download - Institut für Informationswirtschaft und -management (IISM)

Economic Models and Algorithms for
Grid Systems
KUMULATIVE HABILITATIONSSCHRIFT
an der Fakultät für
Wirtschaftswissenschaften
der Universität Karlsruhe (TH)
von
Dr. Dirk Georg Neumann, M.A., Univ. Wisc.
12 October 2007, Karlsruhe

Table of Contents
Table of Contents ___________________________________________________________ i
PART A: OVERVIEW ARTICLE
TOWARDS AN OPEN GRID MARKET _____________________________________________ 1
1 Introduction ________________________________________________________ 2
2 Definitions _________________________________________________________ 3
2.1 Cluster ___________________________________________________________ 3
2.2 Distributed Computing ______________________________________________ 4
2.3 Grid Computing ___________________________________________________ 5
2.4 Utility computing ___________________________________________________ 5
3 Markets as the Business Model for Grid _________________________________ 6
3.1 Pricing strategies ___________________________________________________ 6
3.2 The need for Self-organization ________________________________________ 9
4 Related Work ______________________________________________________ 10
4.1 Market-based Grid systems _________________________________________ 10
4.2 Markets are dead, long live markets __________________________________ 13
5 An Architecture for Open Grid Markets ________________________________ 14
5.1 Architecture Rationale _____________________________________________ 14
5.2 Open Grid Market Architecture _____________________________________ 15
5.2.1 Layer 1 – Core Market Services _____________________________________ 16
5.2.2 Layer 2: Open Grid Market _________________________________________ 16
5.2.3 Layer 3: Intelligent Tools __________________________________________ 17
5.2.4 Layer 4: Grid Application __________________________________________ 18
6 Organization of the book _____________________________________________ 18
6.1 Market Mechanisms for Grid Systems ________________________________ 18
6.1.1 Bridging the Adoption Gap – Developing a Roadmap for Trading in Grids ___ 18
6.1.2 A Truthful Heuristic for Efficient Scheduling in Network-centric Grid OS ____ 19
6.1.3 Economically Enhanced MOSIX for Market-based Scheduling in Grid OS ___ 20
6.1.4 GreedEx – A Scalable Clearing Mechanism for Utility Computing __________ 20
6.1.5 Market-Based Pricing in Grids: On Strategic Manipulation and Computational
Cost ___________________________________________________________ 21
6.1.6 Trading grid services – a multi-attribute combinatorial approach ___________ 21
i

6.1.7 A Discriminatory Pay-as-Bid Mechanism for Efficient Scheduling in the Sun N1
Grid Engine _____________________________________________________ 22
6.1.8 Decentralized Online Resource Allocation for Dynamic Web Service
Applications ____________________________________________________ 22
6.2 Economic Component Design _______________________________________ 22
6.2.1 Self-Organizing ICT Resource Management – Policy-based Automated Bidding _
_______________________________________________________________ 22
6.2.2 A Decentralized Online Grid Market – Best Myopic vs. Rational Response ___ 23
6.2.3 Economically Enhanced Resource Management for Internet Service Utilities _ 23
6.2.4 Situated Decision Support for Managing Service Level Agreement Negotiations _
_______________________________________________________________ 24
6.2.5 Using k-pricing for Penalty Calculation in Grid Market ___________________ 24
6.2.6 Rightsizing of Incentives for collaborative e-Science Grid Applications with TES
_______________________________________________________________ 24
6.3 Design Issues and Extensions ________________________________________ 25
6.3.1 Distributed Ascending Proxy Auction – A Cryptographic Approach _________ 25
6.3.2 Technology Assessment and Comparison: The Case of Auction and E-negotiation
Systems ________________________________________________________ 25
6.3.3 Comparing Ingress and Egress Detection to Secure – Inter-domain Routing: An
Experimental Analysis ____________________________________________ 26
6.3.4 A Market Mechanism for Energy Allocation in Micro-CHP Grids __________ 26
7 Conclusion and Outlook on Future Research ____________________________ 27
7.1 Concluding remarks _______________________________________________ 27
7.2 Outlook __________________________________________________________ 28
Bibliography ______________________________________________________________ 30
Part B: Articles
Article: “Bridging the Adoption Gap – Developing a Roadmap for Trading in Grids” 34
Article: “A Truthful Heuristic for Efficient Scheduling in Network-centric Grid OS” _ 54
Article:“Economically Enhanced MOSIX for Market-based Scheduling in Grid OS” _ 66
Article:“GreedEx – A Scalable Clearing Mechanism for Utility Computing ________ 74
Article:“ Market-Based Pricing in Grids: On Strategic Manipulation and
Computational Cost” _______________________________________________ 95
Article:“Trading grid services – a multi-attribute combinatorial approach ________ 125
Article:“A Discriminatory Pay-as-Bid Mechanism for Efficient Scheduling in the Sun
N1 Grid Engine” __________________________________________________ 154
Article:“Decentralized Online Resource Allocation for Dynamic Web Service
Applications” ____________________________________________________ 164
ii

Article:“Self-Organizing ICT Resource Management – Policy-based Automated
Bidding” ________________________________________________________ 172
Article:“A Decentralized Online Grid Market – Best Myopic vs. Rational Response” 180
Article:“Economically Enhanced Resource Management for Internet Service Utilities” _
_________________________________________________________________ 91
Article:“Situated Decision Support for Managing Service Level Agreement
Negotiations” _____________________________________________________ 205
Article:“Using k-pricing for Penalty Calculation in Grid Market” _______________ 214
Article:“Rightsizing of Incentives for collaborative e-Science Grid Applications with
TES” ___________________________________________________________ 224
Article:“Distributed Ascending Proxy Auction – A Cryptographic Approach” _____ 235
Article:“Technology Assessment and Comparison: The Case of Auction and E-
negotiation Systems” ______________________________________________ 244
Article:“ Comparing Ingress and Egress Detection to Secure – Inter-domain Routing:
An Experimental Analysis _________________________________________ 284
Article:“A Market Mechanism for Energy Allocation in Micro-CHP Grids ________ 307
Lebenslauf ______________________________________________________________ 317
Schriftenverzeichnis ______________________________________________________ 327
iii

Part A
Overview Article
Towards an Open Grid Market
1

TOWARDS AN OPEN GRID MARKET
Dirk Neumann
Institute of Information Systems and
Management (IISM)
Universität Karlsruhe (TH)
neumann@iism.uni-karlsruhe.de
1 Introduction
During the last years, the costs for ICT infrastructures have enormously exploded as a result
of „one-application-one-platform“ style of deployment. This has left most of the ICT infrastructures
with extremely low system utilization and wasted resources (Carr 2005). Examples
can be drawn from virtually all sources: One recent study of six corporate data centers reported
that the bulk of their 1000 servers just utilized 10% to 35% of their available
processing power (Andrzejak, Arlitt et al. 2002). IBM estimated the average capacity utilization
rates of desktop computers to just 5% (Berstis 2002). The Gartner Group indicates that
between 50% and 60% of a typical company’s data storage capacity is wasted (Gomolski
2003). Overcapacity can not only be observed with respect to hardware, but also to software
applications. In essence, highly-scalable applications can serve additional users at almost no
incremental costs – hence redundant installations of the same application create unnecessary
costs (Carr 2003; Carr 2005).
In recent times, ICT is undergoing an inevitable shift from being an asset that companies
posses (e.g. computers, software) to being a service that companies purchase from designated
utility providers. This shift will take years to fully unfold, but the technical building blocks
have already begun to take shape. On the coat tail of this shift, the business model of utility
computing or equivalently e-Business on-demand is more and more emerging.
Utility computing denotes the service provisioning model, in which a service provider makes
computing resources and infrastructure management available to the customer as needed, and
charges them based on usage. This new business model relying on service-oriented architectures
contributes to driving down costs – resulting from low system utilization and waste –
and complexity by plug & play service provisioning in the ICT infrastructure.
Typically, two general pricing models for utility computing are identified. The first pricing
model is the subscription model. Accordingly, users are charged on a periodic basis to subscribe
to a service. The second pricing model is the metered model, where users are charged
on the basis of actual usage. While the first model gives rise to a waste of resources, since
charging is independent of usage, the second, the metered model, is ultimately deemed the
promising to remove the waste of resources (Rappa 2004).
From the user’s point of view, the metering of usage can be compelling, as “one pays for
what one uses”. For the service provider the metered model is appealing, as it offers an opportunity
to sell idle or unused computer capacity. The metered model, however, is not without
problems per-se: It has to be determined, which price is charged from a customer, when
the meter is turned on (Rappa 2004). The price is essential, as it determines the incentive for
resource owners to provide them as well as for consumers to demand them. If the price is too
high, not all the offered resources are demanded contributing to idle resources. If the converse
is true, not all tasks can be realized as the resource provision is too low, which is equally bad.
2

In general, market-based approaches are considered to work well for price determination. By
assigning a value (also called utility) to their service requests, users can reveal their relative
urgency or costs to the service (Buyya, Abramson et al. 2004; Irwin, Grit et al. 2004). Despite
their theoretically excellent properties only few market-based approaches have become operational
systems, let alone commercial. The reason for this lack of market-based operational
systems stems from the fact that many questions arise particularly concerning the applicability
of markets, and their relevance to system design.
This book strives for bridging this gap by providing economic models and algorithms for Grid
systems in general, and for utility computing in particular. It is argued that those economic
models and algorithms supporting technical Grid infrastructures may result in a widespread
usage of utility computing. The results reported in this book have been generated within the
EU funded project “Self-organizing ICT resource management” (SORMA) 1 . While the
project is mainly focused on the technical implementation aspects, this book is devoted to the
underlying economic principles.
This introductory chapter primarily aims at providing a common background for understanding
the selected papers of this book. Moreover, the paper promotes the idea of an open Grid
market, which defines a blueprint for how to construct Grid markets. This blueprint follows a
component-based architecture with well-defined interfaces between the different components.
Most of the components embody economic models and algorithms. This paper introduces the
architecture of an Open Grid Market and describes how the following papers fit into this architecture.
The remainder of this chapter is structured as follows. Firstly, important definitions are introduced
that are commonly used throughout the book (2.1). Secondly, it is motivated why markets
are deemed promising for Grid computing (2.2). Thirdly, related approaches are described;
it is analyzed why those approaches never made it into practice (2.3). Fourthly, the
architecture of an Open Grid Market is proposed, which improves over related approaches
(2.4). Lastly, the structure of the book is illustrated by referring to the overall Open Grid Market
Architecture (2.5).
2 Definitions
The terms cluster, distributed computing, Grid computing and utility computing are often used
as synonyms. Although these terms denote related terms, they impose different requirements
upon both the technical infrastructure as well as on the underlying economic models. In the
following we will define those four concepts, point out their differences and give examples.
2.1 Cluster
A cluster denotes a group of tightly coupled resources that work together on a common problem.
The resources(e.g. computers consisting of memory and computation power) are homogeneous,
which are typically connected via local area networks (Buyya 1999). As all parts of
a cluster are working on a common problem they can be viewed as a single entity. Clusters
can be distinguished into three types:
• Failover clusters
Failover clusters aim at improving the availability of services that the cluster performs.
This implies that they maintain redundant nodes, which can be used once components
of the cluster fail to provide the service (Marcus and Stern 2000). Failover clusters are
frequently used for important databases, business applications, and e-commerce websites.
1 (http://www.sorma-project.eu/)
3

Examples of failover cluster solutions are Linux-HA (High-Availability Linux), Sun-
Cluster, or Microsoft Cluster Server.
• Load-balancing clusters
Load-balancing clusters distribute the workload to a set of computers or servers to
achieve a balanced workload over all nodes of the cluster (Kopparapu 2002). Albeit
those clusters are specialized in load balancing, in many cases they embed failover
properties as well; those clusters are dubbed server farm. Professional load balancers
are the Sun Grid Engine N1, Maui and MOSIX (Barak, Shiloh et al. 2005) and are typically
used as back-ends for high traffic websites such as Wikipedia, Online-shops or
marketplaces.
• High-performance computing clusters
High-performance computing clusters are those clusters that aim at increasing the performance
by exploiting parallelism achieved through splitting a computationally demanding
task into several subtasks. Examples for state-of-the-art high-performance
computing solutions are so-called Beowulf clusters (Bell and Gray 2002). Almost all
supercomputers are clusters. For instance the largest super computer of the world, the
BlueGene/L at Lawrence Livermore National Laboratory, California is devoted to high
performance computing. It consists of 131072 processors attaining performance of
more than 367 TFLOPS 2 at its peak. The investment costs for BlueGene/L were quite
enormous exceeding $ 290 million.
2.2 Distributed Computing
Another computing model is referred to as distributed computing. In essence, it denotes a method
of computer processing, where parts of the program are parallelized on different machines.
Different to clusters, the involved resources are typically heterogeneous. In addition
the resources are only loosely coupled for example over the internet. Distributed computing is
characterized by the fact that there exists a central management component, which divides the
program into smaller pieces and distributes those to the connected machines. The central
management component thereby considers the heterogeneous resources when processing the
distribution job.
The most prominent example of distributed computing is SETI@home developed by the University
of California at Berkeley. It is essentially a model of voluntary computing. End users
can donate their idle computation time of their home PC to the search for extraterrestrial life.
Subsequently, the results of the many PCs are automatically reported back and reassembled
into a larger solution (Werthimer, Cobb et al. 2001). SETI currently combines more than
1,594,396 active users from more than 208 different countries. The impact SETI has is quite
enormous totaling average performance of 261 TFLOPS and up to 500 TFLOPS in its peak
exceeding the performance of the largest supercomputer in the world. SETI has become the
pioneer in voluntary computing – the BOINC framework is an open source software platform
that is used for other distributed computing projects such as Folding@home and Cancer Research
Project
The advantages of distributed computing are impressive, as processor performances of supercomputers
can be reached free of charge, just by utilizing otherwise idle resources. The potential
of voluntary computing is huge, taking into consideration that home PCs are only utilized
by 5 % of the time. Having in mind that investment costs accounts only for 20 % of the total
cost of ownership leaving 80 % for administration, the computation achieved by the
2 TFLOPS stands for Terra FLOPS, where FLOPS (floating point operations per second) denote a performance
metric for processors.
4

SETI@project would incur total cost of ownership of more than $ 1,450 million, which pertains
to the BlueGene/L supercomputer.
2.3 Grid Computing
The term Grid computing was initially introduced by Foster and Kesselman in 1998. Accordingly
a grid denotes “a hardware and software infrastructure that provides dependable,
consistent, pervasive, and inexpensive access to high-end computational capabilities” (Foster
and Kesselman 1998). Later on Foster refined this definition as “direct access to computers,
software, data, and other resources, as is required by a range of collaborative problemsolving
and resource-brokering strategies emerging in industry, science, and engineering”
(Foster and Kesselman 1998).
Obviously, the concept of Grid computing is very related to distributed computing. Both
computing models involve heterogeneous resources that are loosely coupled. Typically organizational
boundaries are crossed. Different to distributed computing, Grid computing relies
on distributed management. A key advantage of Grid systems is their ability to pool together
resources, including both computational resources as well as hard- and software, that can then
be shared among Grid users (Alkadi and Alkadi 2006; Knight 2006). This sharing has major
ramifications for the efficiency of resources, as it
• allows increased throughput because of statistical multiplexing and the fact that users
have a different utilization pattern.
• lowers delay because applications can store data close to potential users
• offers greater reliability because of redundancy in hosts and network connections
All of this results in lower costs than comparable private system (Lai, Huberman et al. 2004).
An example for the potentials of Grid computing offering is Novartis. For their research and
development activities in creating new drugs, they connected their PCs within all branches of
the company. Over night up to 2700 computers were used as number crunching machines.
According to their press releases, the Grid helped reducing projects scheduled on a single supercomputer
for six years to a total run time of 12 hours. Including the productivity gains Novartis
claimed Grid computing to be responsible for savings up to $ 200 million over three
years.
2.4 Utility computing
Utility computing 3 differs from the previous three concepts, as it does not denote an additional
computing model, but a business model on top of Grid computing. More precisely, utility
computing refers to offering a bundle of resources (e.g. computation, and storage) as a metered
service. Comparable with a utility (such as electricity, water or gas) the user is unaware
about where the resources are coming from. Hence, utility computing can be understood as
the on-demand delivery of infrastructure, applications, and business processes in a securityrich,
shared, scalable, and standards-based computer environment over the Internet for a fee.
Customers will tap into IT resources and pay for them (Rappa 2004)]. Utility computing provides
companies with the advantage of low investment costs to acquire hardware, as resources
can be rented. As the terms utility computing and Grid computing is closely related, we will
use them as synonyms throughout the remainder of the book.
The most prominent example of utility computing is Sun’s platform network.com
(http://www.network.com/). Network.com makes use of Grid computing technologies and
offers computational resources for a fix price of $1 per CPU hour. Although utility computing
has the potential to mark “the end of corporate computing” as Nicholas Carr terms it (Carr
3 Synonyms are cloud computing or on-demand computing.
5

2005), neither Sun’s network.com nor any other utility computing initiative has succeed to
exploit the potential. Currently, Sun has had only one single customer, Virtual Computing
Company (VCC) that bought over 1 million CPU hours from Sun. It is a well known fact that
VCC paid a price way less than $1 per CPU hour. Rumors are around that Sun has bought
VCC for marketing purposes. These rumors are to some extent confirmed by the fact that
VCC also heavily invested in the build up of their own infrastructure, which is said to exceed
the 1 million CPU hours by far.
The reasons why utility computing has not yet taken off are manifold. Legal problems are
among the most severe. For instance network.com conducts all computations within the USA.
Requests from other countries could not be accepted due to the absence of legally binding
contracts. Sun has now remedied this shortcoming, allowing users from over 25 countries to
use network.com on-demand. Other shortcomings address the fact that $1 per CPU hour is too
expensive. $1 was merely set by Sun as a marketing message that can be easily communicated.
According to Sun officials, the margin for CPU hours is not yet skimmed off. Apparently
$1 per CPU hour establishes the break-even point for Sun if their infrastructures are utilized
by 30%. Amazon offers their idle CPU hours for 10 cents per CPU hour (Butler 2006).
In their fee structure they also include a fraction for bandwidth usage, which adds to the 10
cents.
3 Markets as the Business Model for Grid
The need for computational resources from industry – especially in the automotive or pharmaceutical
area – is tremendous, but current providers are struggling to accommodate this
need. It seems that the fix prices are still too high. Throughout this book it is argued that viable
business models are missing for Grid computing. This hypothesis is in line with the
“Forrester Research of Grid Computing” The451Group (Fellows, Wallage et al. 2007). Markets
are known to work effectively in allocating resources (who gets which resources and
when?), especially in cases where the resources are under decentralized control, users and
providers are selfish, and where demand and supply are fluctuating. This ability of markets is
gained by establishing dynamic prices that reflect demand and supply situations. Hence markets
seem to be the adequate model to determine accurate prices (which are unknown right
now) for the users and providers to be acceptable. In addition, price signals help to identify
priorities in the requests (Neumann, Holtmann et al. 2006).
3.1 Pricing strategies
In the following, it is explored what kind of pricing strategies are deemed promising for a use
in Grid computing. As in any distributed system, a key problem to sharing resources is find a
suitable allocation (i.e. who gets what?). Resource allocation is not a problem, if sufficient
resources are available. Once, demand exceeds supply this becomes more problematic, as a
scheduling unit in the Grid must decide, which demand will be supplied with the given resource
endowment. Nowadays, resource allocation increasingly becomes an issue, because
resource demands of data mining applications, scientific computing, rendering programs, and
Internet services have kept pace with hardware improvements (Cherkasova, Gupta et al.
2007). This is expected to further increase.
Traditionally, Grid systems address the resource allocation problem with naïve heuristics
(e.g., first-in first out) or with more complex idiosyncratic cost functions (e.g. throughput
maximization) (Buyya, Stockinger et al. 2001; Buyya, Abramson et al. 2004).
Today's schedulers have recognized the need of expressing values by providing a combination
of approaches including user priority, weighted proportional sharing, and service level agreements
that set upper and lower bounds on the resources available to each user or group (Irwin,
Grit et al. 2004; LSF 2005; Sun 2005).
6

One example will show why value-oriented algorithms are needed. Suppose SME Integrated
Systems needs to submit the accounting statement in time in order to avoid a huge fine. Nonetheless,
Integrated Systems resource request cannot be computed right away, as the resources
are blocked by a big rendering job by corporation MovieRender. MovieRender needs the results
of the rendering in two weeks and is indifferent whether this rendering job is computed
now or the next week. The resource allocation should (re)allocate resources to Integrated Systems
which needs them urgently reducing the allocation to MovieRender, as long as the delay
does not change MovieRender value (i.e. the job is conducted within the given timeframe).
Proportional Share is an example of such a scheduling mechanism. In essence, Proportional
Share attaches to any user a weight. The resources allocated to the users depend on their
weighted shares. For example, if there are two users Integrated Systems and MovieRender,
where Integrated Systems has a weight of 1 and MovieRender of 2, Integrated Systems is allocated
1/3 of the resources, leaving 2/3 to MovieRender. This is effective, if MovieRender always
performs more valuable computations than Integrated Systems. In our example, Integrated
Systems clearly has suddenly a higher valuation and MovieRender has a low valuation.
This is rather common, as the arrival process of important work is highly bursty (similar to a
web-, email-, file server) (Lai 2005). As a consequence, the situation would be efficient, if
MovieRender announces that his job is not time-critical and reduces its weight, while Integrated
Systems increases her weight correspondingly. But, if MovieRender’s rendering creates
a small value, then there is no incentive for MovieRender to lower his value, since it would
reduce its own value without compensation.
In other words, maximizing the utility (i.e. the sum of valuations) is possible only if the scheduler
knows the attached valuations, or as in the case of Proportional Share, the exact relative
weights. However, as in the example MovieRender has no incentive to reduce its weight,
when it looses value by not getting its computation done. This is a common phenomenon in
Economics, known as the “Tragedy of the Commons”, which states that the utility maximizing
behavior of the clients results in an overall inefficient solution (Hardin 1968).
Value-oriented approaches are, hence, not sufficient per se, to achieve an efficient solution.
Only if all participants are willing to report their priorities and values honestly, these algorithms
such as Proportional Share will work well. As the example shows, in most cases there
will be no incentive for the participants, who act strategically in order to maximize their utility.
One way to prevent strategic misrepresentation of the weights would be to install a system
administrator that monitors the system and dynamically changes the weights of the users to
maintain high efficiency. Even inside organizational units this is expensive and extremely
error-prone. Once organizational boundaries are crossed, such a system-wide monitoring is
politically almost impossible to achieve. This is where markets enter the discussion.
Markets have the ability to set the right incentives for users to reveal their true valuation as
well as for resource owners to provide those resources that are scarcest in the Grid. By the
dynamic adjustment of the prices, markets can react to the ever changing conditions of resource
demand supply. This flexibility is one major advantage against long-term heuristics
like Proportional Share.
Proportional Share typically determines the proportional value for a longer period, say a week
or a month. Obviously, Proportional Share does neither set the incentive to shift usage from a
high demand resource to a low demand resource, nor does it reward shifting usage from a
high demand period into a low demand period. For instance, the system as a whole may have
high demand for a certain number of CPUs and low demand for memory. Since those resources
are typically allocated separately, there will be no incentive in the system to substitute
CPUs by memory (Sullivan and Seltzer 2000; Lai 2005).
7

With the introduction of prices, incentives will be given to the users to substitute the scarce
resource (e.g. number of CPUs) with less scarce resource (memory). For instance, a fixed
pricing scheme which requests 10 € for one CPU and 1 € per Gigabyte of memory, sets incentives
to reduce the number of used CPUs in favor of the cheaper memory. A fixed pricing
scheme, however, is not enough to achieve an efficient allocation, as the following consideration
will reveal:
Suppose the fixed price is as in Figure 1 (c.f. Lai 2005). Demand changes over time – this will
be depicted by the parabola – without loss of generality the costs of supplying resources are
assumed to be zero. If demand is below the fixed-price (left end of the graph), no resource
will be requested. This is because the value for the resource, represented by the demand curve
is below the price. In our example, neither Integrated Systems nor MovieRender would buy
the resources as they are too expensive. As a consequence, there is a loss of utility, which is
denoted by the area below the demand curve. In the middle part of the Graph there are could
be more buyers willing to pay the fixed price, as their demand exceeds the price. For instance
both Integrated Systems and MovieRender have a valuation higher than the fixed price. If the
seller allocates the resource efficiently to the user who values the resource most (Integrated
Systems in our example), there will be unrealized profit for the seller indicated by the striped
gray area, which refers to the difference between the demand curve and the fixed price.
Figure 1: Fixed Pricing (Lai 2005)
Unfortunately, there is no mechanism for the seller to find the user with the highest resource.
In many cases, the buyer receiving the allocation is someone who is below the highest valuation.
This is shown in the right part of Figure 1, where some user with inferior valuation (MovieRender
in our case) receives the allocation. The difference between the demand curve of
the highest valued user and the one who receives the allocation marks the efficiency loss or in
other words waste. A market mechanism has the ability to set the prices in a way that most, if
not the entire utility under the demand curve can be realized; resulting in an efficient allocation,
where a system designer cannot extract more utility. The gist of a market is to couple
price discovery with the valuations of the users, which are expressed in the form of bids.
For instance, in our example Integrated Systems would submit a bid with a high valuation to
the Grid Market, reflecting the high demand for resources. MovieRender would only mod-
8

erately bid for resources, since it does not matter whether the job is conducted now or at a
later time. For Integrated Systems, it is inadvisable to submit a bid lower than its valuation,
since this would lower the probability of winning. Conversely, bidding higher than the valuation
would create the risk of paying more than the job is worth. For MovieRender, the same is
true. As a result of the market mechanism, Integrated Systems receives the resources and pays
a moderate price.
3.2 The need for Self-organization
In summary, markets in Grids represent a business model for the resource owners, as a way to
get money for their resource provision. On the other hand, markets are adequate way to determine
the most efficient allocation making the Grid attractive for users.
Suppose Integrated Systems needs more than one resource for the completion of the accounting
statement, say CPU cycles and memory. Then Integrated Systems needs to formulate bids
for both resources. If the resource demand is varying over time, Integrated Systems would
have to adjust the bids frequently.
For long-term contracts, manual discovery of the bids, appears to be reasonable, it is ineffective
if low-value bids must be frequently submitted. Especially, if the resources are purchased
on-demand, manual bid formulation is too slow and creates too many transaction costs. To
reduce overhead, it is advisable to automate the bidding process. If the automated bid generator
is coupled with the ICT infrastructure and possesses business knowledge on how to submit
bids, the entire bidding process can be virtualized.
Once the bidding process is virtualized, the ICT system of an organization will self-organize
its resource management. Overcapacity will be provided over the market while undercapacity
initiates a buying process. The idea of the self-organizing resource management in depicted in
Figure 2.
Supply
Demand
Supply
Situation
Information
Policies
Bidding
Agent
Bids
Open
Grid
Market
Bids
Bidding
Agent
Information
Policies
Demand
Situation
Agreements
Agreements
Resource
Owner A
Resource Provision
Payment
User I
Figure 2: Self-organizing Resource Management
Depending on the current demand for resources created by the Grid application, the bidding
agent of the users autonomously defines bids on the basis of user-specific policies. The bids
are submitted to the Open Grid Market. Likewise, on the resource owner side, the bidding
agents publish automatically available resources based on the policies. The Open Grid Market
matches requesting and offering bids and executes them against each other. The allocations
are formulated in agreements (or contracts). The fulfillment of the agreements will be automatically
executed, while the Grid middleware is responsible for the resource provisioning
and the payment system (such as PayPal) for the monetary transfer of funds. Thus, to be effective,
any Grid Market system needs to be composed of three components:
• An Open Grid Market, which determines the allocation in an efficient way and offers in
addition complementary market services. Openness is widely defined and refers to the
9

participants (i.e. all participants are potentially admitted to the market) and to the communication
protocols, so that the market can be accessed by any kind of middleware
system.
• Economic Grid Middleware, which extends common virtualization middleware (e.g.
Globus, GRIA, UNICORE, gLite, VMWare, Xen) in a way that the allocations determined
by the Open Grid Market can be deployed.
• Intelligent Tools, which automate the bidding process.
Any reasonable effort in establishing Grid markets ultimately need to involve all three components,
as these are prerequisite for full automation of the market process (Neumann 2006).
4 Related Work
Since the 60s researchers have motivated the use of markets as a means to cope with those
incentival problems in distributed computing. The first attempt has been made with auctioning
off time slots of the Harvard supercomputer (Sutherland 1968). While this primer was purely
paper-based and restricted to one single computer, subsequent proposals and prototypes offered
automated trading in distributed environments. Despite the fact that the idea of using
markets in distributed computing is not new, no implementation has made it into practice.
4.1 Market-based Grid systems
In the following the most significant efforts will be compiled. Essentially, there are several
ways to classify market-based distributed computing environments. This section will use a
classification along the lines of Buyya et al and will sketch the related work (Buyya, Stockinger
et al. 2001; Yeo and Buyya 2004). The underlying taxonomy is structured around the
market mechanisms, which are incorporated in the distributed computing environments. 4 It
will be differentiated into the following market mechanisms:
• Posted Price
In posted price settings the price is revealed openly to all participants. As we have
shown in our preceding motivation, the posted price is inadequate from an economic
point of view if demand and/or supply are strongly fluctuating.
• Commodity Market
In a commodity market, the resource owners determine their pricing policy and the user
determines accordingly his/her amount of resources to consume. Pricing can depend on
the parameters such as usage time (e.g. peak-load pricing) or usage quantity (e.g. price
discrimination) (Wolski, Plank et al. 2001). In many cases, it is referred to flat rates,
which boils down to fixed price until a certain amount of resources or a certain time is
reached.
• Bargaining
In bargaining markets, resource owners and users negotiate bilaterally for mutually
agreeable prices. 5 By gradually lowering their claims, resource owners and users eventually
reach an agreement. Logrolling for several attributes can also be included in the
bargaining protocol (Rosenschein and Zlotkin 1994; Kersten, Strecker et al. 2004).
4 From an economic point of view, this taxonomy is not detailed enough. Nevertheless, for a rough classification
of running resource management systems this taxonomy is despite its lack of details sufficient.
5 A game-theoretic treatment of applicable bargaining models can be found at (Wolinsky 1988; De Fraja and
Sakovics 2001; Satterthwaite and Shneyerov 2003).
10

• Contract Net Protocol
In the contract net protocol the user advertises its demand and invites resource owners
to submit bids. Resource owners check these advertisements with respect to their requirements.
In case the advertisement is favorable, the resource owners respond with
bids. The user consolidates all bids, compares them and selects the most favorable
bids(s).
• Proportional Share
Proportional Share assigns resources proportionally to the bids submitted by the users.
• Auctions
Auctions are mediated market mechanisms. An auctioneer collects bids from either one
market side (buyers or sellers) or from both. According to the auction rules (which are
known to all bidders) the auctioneer allocates the resources. Typical one-sided auctions
are the English, First-Price Sealed Bid, Dutch and Vickrey auction. Two sided auctions
are the Double Auctions.
A survey using the above mentioned taxonomy over selected market-based resource management
systems are summarized in the following.
Computing
Platform
Marketbased
System
Market
Mechanism
Description
Distributed
Computing
SPAWN Auction The SPAWN system provides a market mechanism for trading
CPU times in a network of workstations (Waldspurger, Hogg et al.
1992). SPAWN treats computer resources as standardized
modities and implements a standard Vickrey auction. It is known
from auction theory that the Vickrey auction attains (1) truthful
preference revelation and (2) an efficient allocation of resources
(Krishna 2002). However, SPAWN does not make use of the
neralized Vickrey auction, which can cope with
ties. 6 Furthermore, the Vickrey auction can traditionally neither
cope with multiple attributes nor with different time slots.
Peer-to-
Peer
Stanford
Peers
Bargaining
The Stanford Peers model is a Peer-to-Peer system which
implements auctions within a cooperative bartering model in a
cooperative sharing environment (Wolski, Brevik et al. 2004). It
simulates storage trading for content replication and archiving. It
demonstrates distributed resource trading policies based on
auctions by simulation.
Internet POPCORN Auction POPCORN provides an infrastructure for global distributed
computation (Nisan, London et al. 1998; Regev and Nisan 1998).
POPCORN mainly consists of three entities: (i) A parallel program
which requires CPU time (buyer), (ii) a CPU seller, and (iii) a
market which serves as meeting place and matchmaker for the
buyers and sellers. Buyers of CPU time can bid for one single
commodity, which can be traded executing a Vickrey auction
repeatedly. POPCORN obviously suffers under the same
shortcomings as SPAWN.
6 Users often demand a bundle of resources. Since they need more than one resource to complete a task, they
value the bundle higher than the sum of its constituents. In an extreme case, the constituents may generate
no value at all – only the bundle creates value. If the resources are sequentially auctioned, the user is facing
the so-called exposure risk, once he received one leg of the bundle, he is in danger of running a loss, if he
does not get the other leg.
11

Computing
Platform
Marketbased
System
Market
Mechanism
Description
Grids Bellagio Auction Bellagio is intended to serve as a resource discovery and resource
allocation system for distributed computing infrastructures. Users
express preferences for resources using a bidding language, which
support XOR bids. The bids are formulated in virtual currency.
The auction employed in Bellagio is periodic. Bids from users are
only accepted as long as enough virtual currency is left (AuYoung,
Chun et al. 2004).
Grids CATNETS Bargaining In CATNETS, trading is divided into two layers, the
application/service layer and the resource layer. In both layers, the
participants have varying objectives which change dynamically
and unpredictably over time.
In the application/service layer, a complex service is a proxy who
negotiates the access to bundles of basic service capabilities for
execution on behalf of the application. Basic services provide an
interface to access computational resources
Agents representing the complex services, basic services and
resources participate in a peer-to-peer trading network, on which
request are disseminated and when an appropriated provider is
found, agents engage in a bilateral bargaining (Eymann, Reinicke
et al. 2003; Eymann, Reinicke et al. 2003; Eymann, Ardaiz et al.
2005).
Grids G-
Commerce
Commodity
market,
auction
G-Commerce provides a framework for trading computer resources
(CPU and hard disk) in commodity markets and Vickrey auctions
(Wolski, Plank et al. 2001; Wolski, Plank et al. 2001; Wolski,
Brevik et al. 2003). While the Vickrey auction has the
abovementioned shortcomings in grid, the commodity market
typically works with standardized products. Additionally, the
commodity market cannot account for the complementarities
among the resources, as only one leg of the bundle is auctioned off,
exposing the bidder to the threshold risk.
Grids Nimrod/G Commodity
market
Grids OCEAN Bargaining/
Contract net
Nimrod/G enables users to define the types of resources needed
and negotiate with the system for the use of a particular set of
resources at a particular price (Buyya, Abramson et al. 2000). This
requires from the system to conduct resource discovery, which can
become quite complex, as the numbers of resources can be large.
Also, does the system need to support price negotiations, which
may be complex as well. Both resource discovery and negotiation
can become very cumbersome, if the users demand bundles instead
of single resources.
OCEAN (Open Computation Exchange and Arbitration Network)
is a market-based infrastructure for high-performance computation,
such as Cluster and Grid computing environments (Acharya,
Chokkaredd et al. 2001; Padala, Harrison et al. 2003). The major
components of the OCEAN’s market infrastructure are user
ponents, computational resources, and the underlying market
chanism (e.g. the OCEAN Auction Component). In the OCEAN
framework, each user (i.e. resource provider or consumer) is
represented by a local OCEAN node. The OCEAN node
ments the core components of the system, for instance a Trader
Component, an Auction Component, or a Security Component.
The implemented OCEAN auctions occur in a distributed Peer-to-
Peer manner. The auction mechanism implemented in the OCEAN
12

Computing
Platform
Marketbased
System
Market
Mechanism
Description
framework can be interpreted as a distributed sealed-bid
nuous double-auction (Acharya, Chokkaredd et al. 2001). A trade
is proposed to the highest bidder and the lowest seller. Afterwards,
the trading partner can renegotiate their service level agreements.
The renegotiation possibility one the one hand allows to cope with
multiple attributes and with the assignment of resources to time
slots. Nonetheless, the negotiation makes the results of the
rent auction obsolete. Neither the auction can enfold its full
tial, nor can the negotiation guarantee to achieve an efficient
cation, as competition is trimmed.
Grids Tycoon Proportional
Share
P2P clusters like the Grid and PlanetLab enable in principle the
same statistical multiplexing efficiency gains for computing as the
Internet provides for networking. Tycoon is a market based
distributed resource allocation system based on an Auction Share
scheduling algorithm (Lai, Rasmusson et al. 2004). Tycoon
distinguishes itself from other systems in that it separates the
allocation mechanism (which provides incentives) from the agent
strategy (which interprets preferences). This simplifies the system
and allows specialization of agent strategies for different
applications while providing incentives for applications to use
resources efficiently and resource providers to provide valuable
resources. Tycoon’s distributed markets allow the system to be
fault-tolerant and to allocate resources with low latency. Auction
Share is the local scheduling component of Tycoon.
Table 1: Summary of market-based systems (adapted from (Yeo and Buyya 2004))
Those market-based resource management systems, however, are purely prototypical implementations.
Apparently, only two prototypes implement all three components being CAT-
NETS and Tycoon. Tycoon developed by Hewlett Packard offers the most comprehensive
realizations of the system. However, Tycoon still fails to meet the challenge, as the supported
economic models and algorithms are too simplistic and the deployment requires rebooting the
resources before provisioning. CATNETS offers very interesting features but lacks comprehensive
support (e.g., monitoring, multi platform deployment).
All other approaches neglect the provision of decision support. In ordinary commodity markets,
bidders have to come up with their bids themselves by rational deliberation. In Grid
markets this can become extremely cumbersome, as demand and supply fluctuates dynamically
and is in most of the cases unknown. In addition, the complex auctions make it difficult
even for humans to devise suitable strategies.
4.2 Markets are dead, long live markets
Despite the shortcomings of providing intelligent tools, economic middleware extensions and
a Grid market, there are several other reasons why previous systems have not (yet) made it
into practice. Those other reasons partially stem from the fact that several prerequisites on the
technical and economic side have not been given. In recent times, advancements in Grid
Computing and Economics have been achieved so that the prerequisites as well as the environment
have changed by now. Those advancements favor the application of market-based
platforms for Grid systems. Among the most important changes are the following three
(Shneidman, Ng et al. 2005):
• Increasing Demand for Resources: In the past, demand for computational resources
was not an issue. As a consequence, there was no scarcity of resources that would have
13

equired value-based scheduling mechanisms. Thus, past market-based systems hardly
(if ever) saw real field tests, and contention was often artificially generated. This has
changed due to the increasing demand for resources that exceeds supply. A deployed
market-based Grid system could solve real resource conflicts. The confrontation of
market-based Grid systems is essential, as the real data will give researchers the insights
to evaluate (and to adjust) their market-based resource schedulers.
• Improved technical infrastructure: Past systems had to deal with limitations in technical
Grid infrastructures. For example, the enforcement of the allocation decision was
not possible. Market mechanisms without proper enforcement are useless, as the agreements
are not binding. Recent advancements in the area of operating systems and Grid
middleware have overcome this and other technical problems.
• Advanced economic models and algorithms: Market-based Grid systems exclusively
use simple market mechanisms. Simple market mechanisms, however, cannot attain
economic efficiency, as they are unable to capture complementarities (i.e. one resource
creates value for a user only in connection with another resource). During the past decade,
advances have been made in the theory and practice of more complex market mechanisms.
Current mechanisms can support combinatorial bidding, which allow the expression
of multiple resource needs. As such, those bidding languages can represent any
logical combination of resources, such as AND, OR, and XOR. On the other hand, these
combinatorial mechanisms come at an additional cost of computation. Solving those resource
allocation problems becomes fairly computationally expensive. Hence, the developments
in market mechanisms need to be complemented by standard solver, which
can determine the allocation within an acceptable time frame. In recent times, off-theshelf
solvers such as CPLEX have been developed that can reach both efficient allocation
and acceptable computation time.
The key to the successful deployment of market-based Grid systems is a holistic and interdisciplinary
approach. On the one hand, it is important to have a strong grip on the technical infrastructure.
On the other hand, it is also important to make use of economic design knowledge.
The concept of an Open Grid Market is expected to fill the current gap paving the way
towards widespread Grid adoption (Neumann 2008).
5 An Architecture for Open Grid Markets
"As the size and complexity of software systems increases, the design problem goes beyond
the algorithms and data structures of the computation: designing and specifying the overall
system structure emerges as a new kind of problem" (Garlan and Shaw 1993). To address this
complexity, an architecture-based system development approach seems promising. The architecture-based
approach focuses on the organization and global control structure of the system,
on protocols for communication, synchronization, and data access (Garlan and Shaw 1993).
Key element of this approach is the architecture whose importance goes far beyond the simple
documentation of technical elements. The architecture rather serves as the blueprint for both
the system itself and the project developing it.
As an Open Grid Market appears to be a very complex system, the architecture will be presented.
Before, the rationale of the Open Grid Market will be given, revealing why “openness”
is important in the Grid context.
5.1 Architecture Rationale
As presented in the related work, there are several ongoing attempts to establish a marketbased
Grid system. The approach pursued here is different, as it does not strive for "yet another"
Grid market system. Instead, a market system is envisioned that is open, with respect to
14

1. the communication protocols being used, and
2. the underlying local resource managers of the traded resources.
This devotion to openness has major ramifications on the scope of the market: Any potential
Grid user can access the Open Grid Market via open communication protocols. Furthermore,
resource providers with different virtualization platforms and resource managers can easily
plug in the Open Grid Market. The idea is set up a flexible market infrastructure, which can
access resources over all kind of virtualization platforms by means of wrappers.
The openness is intended to offer the possibility of loosely integrating emerging Grid markets.
For example the Open Grid Market needs to combine platforms like network.com or Amazon’s
elastic cloud but should also be capable of accessing Beowulf or MOSIX clusters. This
approach should attract already existing resource providers such as Sun to plug in the Open
Grid Market. Obviously, competition on such Open Grid Markets encourages markets to integrate
and weed out too complex or inefficient platforms.
5.2 Open Grid Market Architecture
The architecture of the Open Grid Market describes the functionalities of the entire market
system in terms of its functional entities, their responsibilities and their dependencies (see
Figure 3) (Neumann, Stoesser et al. 2008).
Figure 3: Open Grid Market Architecture
Boxes represent functional entities that can be encapsulated in a corresponding component.
Arrows represent dependencies, where an arrow from an entity A to an entity B means that
entity A either receives information or consumes a service from entity B.
The architecture is structured along four layers. Ultimately, it is the Grid application and the
available resources that create demand and supply for computational resources and thus give
15

ise to the establishment of an Open Grid Market. Accordingly, Grid applications as well as
the resource mark the fourth and most ‘low-level’ layer. Layer 3 consists of the intelligent
tools that translate the resource demand or supply by the applications and machines into bids.
Layer 2 concerns the market functionalities, whereas layer 1 comprises the core market services
comprising the basic communication infrastructure that is common to all potential markets.
The intuition for dividing layer 2 and 1 is very simple. While the core market services
establish the common trading platform every market uses, the Open Grid market components
can be different from market to market.
5.2.1 Layer 1 – Core Market Services
As aforementioned, state-of-the-art Grid middleware does not provide all the infrastructure
services necessary for supporting an Open Grid Market. Layer 1 extends the standard Grid
middleware by additional infrastructure services so that the Open Grid Market can be hooked
up with the Grid middleware:
• Trusted market exchange service: All communication among market participants –
users and their applications as well as and providers and their resources – is mediated
by this service, which assures that information is routed to the designated recipients in
a secure and reliable way. The trusted market exchange service also enforces policies
defined on specific messages for logging, encryption, and signing.
• Logging: All transactions executed on the market are registered in a secure log for auditing
purposes.
• Market directory: The market directory is a market-enabled extension to the commonplace
service registries in (Grid) middleware. The technical information on Grid
markets directories are enriched by economically relevant information like pricing or
quality of service parameters. The enriched information serves as input to the trading
management component on layer 2.
• Market information: The market information service allows market participants to
publish information and to gather information from other participants (e.g., prices, resource
usage levels). Participants can query the service or subscribe to topics. Information
queries consider either instantaneous measures or their history.
5.2.2 Layer 2: Open Grid Market
The Open Grid Market in layer 2 defines the arena, where the published resources are assigned
to the Grid applications, following certain market mechanisms.
• Trading management: The trading management component is the access point for
the users to the Open Grid Market where they can find the offered resources and submit
their according bids. More specifically, the trading management matches the technical
descriptions of the requests obtained from the bids to suitable technical descriptions
of the offered resources collected from the associated Grid market directories.
Subsequently, the trading management manages the bidding process according to given
market mechanisms (e.g. MACE auction (Schnizler, Neumann et al. 2006)). If the
bidding process finishes successfully the corresponding bid and offer are submitted to
the contract management.
• Contract management: The contract management component transforms corresponding
pairs of bids and offers to mutually agreed contracts. One important part of these
contracts are service level agreements (SLAs) which define the agreed terms of usage
of the resources and the pricing. The contract management also initiates the enforce-
16

ment of the contract, especially the allocation of the sold resources (aided by the
EERM) and the payment process.
• SLA enforcement and billing: The SLA enforcement and billing component is responsible
for the surveillance and enforcement of the contracts it receives from the
contract management. The component keeps track of the actual usage of the resources,
makes comparisons to the SLA and (if appropriate) initiates the billing and clearing
according to the results of the comparison.
• Payment: The payment service offers a unified interface to payment, isolating the rest
of the components from the particularities of the payment mechanism. The payment
service also generates appropriate logging/auditing information.
• Security management: The security management component is intended as the entry
point for a single sign-on mechanism and is responsible for a tamper-proof identity
management for the consumers, the suppliers and the constituent components of the
Open Grid Market. Thus, all layers provide security connectors that build the technological
bridges from the respective layers to the security management.
• Economically Enhanced Resource Management (EERM): The EERM component
provides a standardized interface to typical Grid middleware (e.g. Globus Toolkit or
Sun Grid Engine). The EERM can shield clients from resource platform specific issues
and also enhance or complement the management functions provided by job scheduling
and submission systems. The EERM’s main duties include (i) resource management
including the management functionalities to achieve the expected service levels
and notify of deviations. The resource management furthermore coordinates independent
resources to allow co-allocation – if not provided by the fabrics (ii) resource
monitoring (iii) accessing resource fabrics via standardized interfaces to create instances
of resources and later make use of them by the application (Macías, Smith et
al. 2007).
5.2.3 Layer 3: Intelligent Tools
The market participants including the Grid applications and resources are supported by intelligent
tools for an easy access to the Open Grid Market.
• Consumer preference modeling: This component allows for the users to describe
their economic preferences that will determine their bidding strategies on the Open
Grid Market, e.g. they could define if they prefer cheap over reliable resources. One
approach would be to provide the users with a GUI in the form of a simplified ontology
modeling tool to instantiate a given consumer preference modeling language.
• Demand modeling: The user needs a tool to specify the technical requirements of her
Grid application. The technical approach for this component is similar (i.e. ontologybased)
to the preference modeling.
• Business modeling: Analogously to the consumer preference modeling, the providers
have to specify their business models to determine the generation of their offers on the
Open Grid Market. For example one part of such a description could be a pricing
model that specifies if the consumer has to pay for booked time-slots or for the actual
usage. As the example indicates, the models specified by means of this component depend
on the implemented market type.
• Supply modeling: The resource modeling component is the correspondent of the demand
modeling component. Aided by this component the providers can technically
specify their offers.
17

• Bid generation: The bid generation is the intelligent (i.e. agent-based) component that
generates and places the bids of the consumer on the Open Grid Market. For this purpose
it considers the user preferences, the technical requirements and the current state
of the market and derives the bids. The bids are submitted to the trading Management
component of the Open Grid Market. The bid generation component could be implemented
with the help of a rule engine for logical inference over the mentioned inputs.
• Offer generation: The offers are assembled from the technical resource descriptions
and the business model of the respective provider by the offer generation component.
It also publishes the offers at the Grid market directory.
5.2.4 Layer 4: Grid Application
Layer 4 pertains to the Grid applications and Grid resources for trade. At the provider side a
provider IT specialist makes use of the intelligent tools in layer 3 to model the provider's
business strategies and the offered Grid resources.
On the consumer side it has to be distinguished between the Grid application's end user(s)
and the consumer's IT support staff who will use the intelligent tools to model an application's
resource requirements and the consumer's preferences.
6 Organization of the book
This book is devoted to the development of economic models and algorithms for the Open
Grid Market. As such, the book mainly deals with layer 2 and 3 – as those areas comprise the
components, where economic principles can help to improve the entire system. This book
contains eighteen different papers, which are structured along three main sections: “Market
Mechanisms for Grid Systems”, “Economic Component Design”, and “Design Issues and
Extensions”. The papers present joint research conducted with colleagues at the Institute of
Information Management and Systems within the scope of the
• EU-funded project “CATNETS” (http://www.catnets.org/),
• EU-funded project “SORMA”, (http://www.sorma-project.eu/),
• BMBF sponsored project “SESAM”,
(http://www.internetoekonomie.uni-karlsruhe.de/)
• project “Billing the Grid” funded by the Landesstiftung Baden-Württemberg as Landesschwerpunktprogramm
(http://www.billing-the-grid.org/), and
• SSHRC funded project “Electronic Negotiations, Media and Transactions for Socio-
Economic Interactions”
(http://interneg.concordia.ca/views/bodyfiles/enegotiation/index.html.)
6.1 Market Mechanisms for Grid Systems
The first section deals with different market mechanisms for scheduling Grid resources. Apparently,
it is devoted to the economic modeling of the trading management component in
layer 2 of the architecture. All papers focus on the design of market mechanisms for different
market segments of the Grid market. This section consists in total of eight papers, where the
first paper is an overview paper and the subsequent seven papers deal with different market
mechanisms.
6.1.1 Bridging the Adoption Gap – Developing a Roadmap for Trading in Grids
The paper “Bridging the Adoption Gap – Developing a Roadmap for Trading in Grids” argues
that the technology of Grid computing has not yet been adopted in commercial settings
18

due to the lack of viable business models. While in academia Grid technology has already
been taken up, the sharing approach among non for-profit organizations is not suitable for
enterprises. In this paper, the idea of a Grid market is taken up to overcome this Grid adoption
gap. Although this idea is not new, all previous proposals have been made either by computer
scientists being unaware of economic market mechanisms or by economists being unaware of
the technical requirements and possibilities.
This paper derives an economically sound set of market mechanisms based on a solid understanding
of the technical possibilities. More precisely, the paper analyzes the characteristics of
the object traded in Grids. This trading object is closely related to the deployment of software
applications via the EERM. The deployment directly on physical resources or via raw application
services has major ramifications for the trading object and consequently for the requirements
for market mechanisms. Physical resources are essentially commodities, whereas application
services can be both standardized commodities and unique entities.
Based on this analysis, a two-tiered market structure along the distinction between physical
resources and application services is derived, where each tier demands different market mechanisms.
The first tier comprises the markets for physical resources (e.g. CPU, memory) and
raw application services. The second tier comprises the markets for complex application services.
Subsequently, existing Grid market mechanisms are classified according to this market structure.
At the core of this paper, we argue that there is no single market that satisfies all purposes.
Reflecting the distinct requirements of different application models and deployment modes,
a catalogue of co-existing market mechanisms is needed.
This paper provides the outline for the following papers, which present the different mechanisms
of the catalogue and derive their economic properties.
6.1.2 A Truthful Heuristic for Efficient Scheduling in Network-centric Grid OS
The paper “A Truthful Heuristic for Efficient Scheduling in Network-centric Grid OS” argues
that a network-centric Grid OS coupled with market-based scheduling can increase efficiency
in Grid and cluster environments by adequately allocating all available resources. This distinguishes
network-centric Grid OS from state-of-the-art Grid middleware, which rely on batch
processing of idle resources only. While there are highly sophisticated market mechanisms
available for Grid middleware where raw services are being traded (e.g. MACE), there are no
mechanisms available for network-centric Grid OS which rely on interactive application
processing.
The paper strives for filling this need by defining a market mechanism as a multi-attribute
exchange in which resource owners and consumers can publish their demand and supply for
resources deployed as raw services. Exact market-based scheduling mechanisms share one
deficiency: they are quite complex. While this may not be a problem in a cluster setting with a
small number of users, it may prove crucial in an interactive, large-scale Grid OS environment.
Hence a greedy heuristic is designed, which performs a fast scheduling while retaining
truthfulness on the request-side and approximating truthfulness on the provisioning-side of the
market.
In summary, the contributions of this paper are threefold: Firstly, the paper designs a multiattribute
exchange that can be used for Grid OS. Secondly, a greedy heuristic is employed to
solve the scheduling problem. Thirdly, an adequate pricing scheme is developed which assures
that reporting truthfully is a dominant strategy for resource requesters and payments to
resource providers are approximately truthful.
19

6.1.3 Economically Enhanced MOSIX for Market-based Scheduling in Grid OS
The paper “Economically Enhanced MOSIX for Market-based Scheduling in Grid OS” is a
follow-up of the previous paper. Essentially, the paper describes the implementation of the
greedy heuristic in the state-of-the-art network-centric Grid OS MOSIX (i.e. Multi-computer
Operating System for Unix). The paper is a result of the co-operation with Amnon Barak and
Lior Amar (both The Hebrew University of Jerusalem) who co-author this paper.
By introducing market mechanisms to MOSIX, end users can influence the allocation of resources
by reporting valuations for these resources. Current market-based schedulers, however,
are static, assume the availability of complete information about jobs (in particular with
respect to processing times), and do not make use of the flexibility offered by computing systems.
In this paper, the implementation of a novel market mechanism for MOSIX, a state-ofthe-art
management system for computing clusters and organizational grids is described. The
market mechanism is designed so as to be able to work in large-scale settings with selfish
agents. Facing incomplete information about jobs’ characteristics, it dynamically allocates
jobs to computing nodes by leveraging preemption and process migration, two distinct features
offered by the MOSIX system.
The contribution of this paper can be summarized as follows. Firstly, a market mechanism is
proposed, which is specifically tailored towards the needs and technical features of MOSIX so
as to generate highly efficient resource allocations in settings with large numbers of selfish
agents. Secondly, the implementation of the proposed market mechanism in MOSIX is presented
which serves as a proof-of-concept.
6.1.4 GreedEx – A Scalable Clearing Mechanism for Utility Computing
The paper “GreedEx – A Scalable Clearing Mechanism for Utility Computing” extends the
mechanism proposed in the paper “A Truthful Heuristic for Efficient Scheduling in Networkcentric
Grid OS” by a more elaborate pricing scheme. The newly developed market mechanism
is dubbed GreedEx, an exchange for clearing utility computing markets, based on a greedy
heuristic.
This market mechanism can be used to perform this clearing in a scalable manner while at the
same time satisfying basic economic design criteria. Besides its computational speed, the
main characteristic of GreedEx is its ability to account for the inter-organizational nature of
utility computing by generating truthful prices for resource requests and approximately truthful
payments to utility computing providers. In general, however, the computational speed of
heuristic comes at the expense of efficiency. Consequently, a numerical simulation is presented,
which shows that GreedEx is very fast and, surprisingly, highly efficient on average.
These results strengthen our approach and point to several interesting avenues for future research.
Obviously, the contribution of this paper is the design as well as analytical and numerical
evaluation of GreedEx that achieves a distinct trade-off, as it conducts fast near-optimal resource
allocations while generating truthful prices on the demand-side and approximately
truthful prices on the supply-side. GreedEx may be used for two purposes: Firstly, utility
computing providers (such as Sun) may choose to run a proprietary GreedEx-based marketplatform
to allocate their scarce computing resources to resource requests in an efficient manner
and to dynamically price these requests. Moreover, a GreedEx-based platform may be run
as an intermediary market which aggregates demand as well as supply across multiple resource
requesters and utility computing providers, thus increasing liquidity and efficiency.
20

6.1.5 Market-Based Pricing in Grids: On Strategic Manipulation and Computational
Cost
The paper “Market-Based Pricing in Grids: On Strategic Manipulation and Computational
Cost” uses the GreedEx mechanism and shows in several propositions the impossibility of a
perfect heuristic that achieves truthfulness on both market sides. In addition, the paper proofs
propositions on split- and merge-proofness as a measure of fairness. In addition, it numerically
shows that manipulation of individual agents very rarely pays off but is very often punished
by the mechanism.
In summary, the paper make the following contributions: an in-depth analysis of these pricing
schemes with respect to the distribution of welfare among resource requesters and providers,
the incentives of single users to try to manipulate the mechanism, and the computational impact
of the pricing schemes on the mechanism is provided. In addition, basic design options of
how the mechanism can be integrated into current Grid schedulers are outlined.
6.1.6 Trading grid services – a multi-attribute combinatorial approach
The paper “Trading grid services – a multi-attribute combinatorial approach” proposes the
derivation of a multi-attribute combinatorial exchange for allocating and scheduling services
in the Grid. In contrast to GreedEx and other approaches, the proposed mechanism accounts
for the variety of services by incorporating time and quality as well as coupling constraints.
The mechanism provides buyers and sellers with a rich bidding language, allowing for the
formulation of bundles expressing either substitutabilities or complementarities. The winner
determination problem maximizes the social welfare of this combinatorial allocation problem.
The winner determination scheme alone, however, is insufficient to guarantee an efficient
allocation of the services. The pricing scheme must be constructed in a way that motivates
buyers and sellers to reveal their true valuations and reservation prices. This is problematic in
the case of combinatorial exchanges, since the only efficient pricing schedule, the VCG mechanism,
is not budget-balanced and must be subsidized from outside the mechanism.
The main contribution of this paper is the development of a new pricing family for a combinatorial
exchange, namely the k-pricing rule. In essence, the k-pricing rule determines the price
such that the resulting surpluses to the buyers and sellers divide the entire surplus being accrued
by the trade according to the ration k. The k-pricing rule is budget-balanced but cannot
retain the efficiency property of the VCG payments.
As the simulation illustrates, the k-pricing rule does not rigorously punish inaccurate valuation
and reserve price reporting. Buyers and sellers sometimes increase their individual utility
by cheating. This possibility, however, is only limited to mild misreporting and a small number
of strategic buyers and sellers. If the number of misreporting participants increases, the
risk of not being executed in the auction rises dramatically. As a result, the k-pricing schema
is a practical alternative to the VCG mechanism and highly relevant for an application in the
Grid. The runtime analysis shows that the auction schema is computationally very demanding.
However, the use of approximated solutions achieves adequate runtime results and fairly mild
welfare losses for up to 500 participants. Comparing these results with an existing Grid testbed
(PlanetLab) demonstrates the practical applicability of the proposed auction.
21

6.1.7 A Discriminatory Pay-as-Bid Mechanism for Efficient Scheduling in the Sun N1
Grid Engine
The paper “A Discriminatory Pay-as-Bid Mechanism for Efficient Scheduling in the Sun N1
Grid Engine” regards resources as traded object. This work was jointly conducted with Simon
See from Sun Microsystems, who also co-authored this paper.
In essence, this paper proposes an extended model of the Sanghavi-Hajek pay-as-bid mechanism
a promising addition to the N1GE scheduler. Current technical schedulers require an administrator
to specify user weights based on these users relative importance, regardless of the
dynamic demand and supply situation, leading to inefficiencies.
To this end, the contribution of this paper is twofold:
• Mechanism design: A discriminatory pay-as-bid market mechanism by Sanghavi and
Hajek is presented. Conditions are analytically derived under which this mechanism
outperforms market-based proportional share – the currently most prominent grid
market mechanism – with respect to both provider’s surplus and allocative efficiency.
• Integration into Sun N1GE: It is further showed that this mechanism is not a purely
theoretical construct but that it can be integrated into state-of-the-art grid schedulers to
economically enrich the current allocation logics. The basic design considerations are
conducted for the case of the N1 Grid Engine (N1GE) the scheduler of Sun Microsystems’s
grid platform.
6.1.8 Decentralized Online Resource Allocation for Dynamic Web Service Applications
The paper “Decentralized Online Resource Allocation for Dynamic Web Service Applications”
deals with the definition of an online market mechanism that allocates requests to resources
once the requests are submitted to the system.
Economic scheduling mechanisms introduce economic principles to Grids and thus promise to
increase the systems’ overall utility by explicitly taking into account strategic, interorganizational
settings. Current economic scheduling mechanisms, however, do not fully satisfy
the requirements of interactive Grid applications. In this paper, a video surveillance environment
is introduced and it is shown how its functionality can be encapsulated in Grid services
which are traded dynamically using an economic scheduling mechanism. Subsequently,
a mechanism from the general machine scheduling domain is introduced. The applicability of
the market mechanism to the Grid context is showcased by a numerical experiment. However,
the mechanism suffers from some limitations due to the distinct properties of Grids. These
limitations are pinpointed: extensions to the basic mechanism are introduced, which may remedy
these drawbacks.
6.2 Economic Component Design
The second section deals with economic models and algorithms for the component Bid Generation
(6.2.1 and 6.2.2) on layer 3 and for the components EERM (6.2.3), Contract Management
(6.2.4), SLA enforcement (6.2.5), and Payment (6.2.6) on layer 2.
6.2.1 Self-Organizing ICT Resource Management – Policy-based Automated Bidding
The paper “Self-Organizing ICT Resource Management – Policy-based Automated Bidding”
is devoted to the bidding agent. In essence, it is a premise of this paper that markets can undertake
this task better than any other coordination mechanism. This is especially the case if
the bidding process is fully automated, such that the market itself will handle any unforeseen
problem in the resource management process.
22

By bringing autonomic and Grid computing as well as economic principles closer together,
market-based approaches have the potential to achieve an economy-wide adoption of Grid
technology and will thus leverage the exploitation of potential Grid technology offers. To
achieve this potential, many Grid Economics related obstacles must be overcome: markets can
only work well if bids accurately reflect demand and supply. If this is not the case, the market
price cannot reflect the correct situation. Thus, the market price looses its capacity to direct
scarce resources to the bidders who value them most.
This paper is unique by recommending an autonomic computing approach for bidding in Grid
markets. By defining policies the business models of the service consumers and providers can
be represented. The proposed bid generation process is currently quite simple and referring to
observations and straightforward rules.
6.2.2 A Decentralized Online Grid Market – Best Myopic vs. Rational Response
The paper “A Decentralized Online Grid Market – Best Myopic vs. Rational Response” establishes
a bridge between the trading management component and the bidding agent. More precisely,
the paper connects the automated bidding approach with the decentralized online mechanism
proposed in “Decentralized Online Resource Allocation for Dynamic Web Service
Applications”.
To this end, the contributions of this paper are twofold. Firstly, the theoretical analysis of Decentralized
Local Greedy Mechanism is based on the assumption of simple agents behaving
according to a so-called “myopic best response strategy”. By means of a simulation with
learning agents, it is shown that an analysis using myopic best response strategies overestimates
the performance of the mechanism and is thus not an appropriate solution concept for
modeling real-world market-based scheduling mechanisms. As a byproduct of this analysis,
the performance benefits of market-based schedulers as opposed to purely technical schedulers
which are solely based on system-centric measures are demonstrated. Secondly, limitations
of the mechanism for the practical use are pointed out. Furthermore, remedies that may
help to mitigate these drawbacks are suggested.
6.2.3 Economically Enhanced Resource Management for Internet Service Utilities
The paper “Economically Enhanced Resource Management for Internet Service Utilities”
local resource management systems are equipped with economic principles. The results reported
in the paper have been achieved in cooperation with Jordi Torres and Jorid Guitart
(both Barcelona Supercomputing Center) who co-author this paper.
In this work various economical enhancements for resource management are motivated and
explained. In addition a mechanism for assuring Quality of Service and dealing with partial
resource failure without introducing the complexity of risk modeling is presented. It is shown
how flexible dynamic pricing and client classification can benefit service providers.
Various factors and technical parameters for these enhancements are discussed in detail.
Moreover, a preliminary architecture for an Economically Enhanced Resource Manager integrating
these enhancements is introduced. Due to the general architecture and the use of policies
and a policy manager this approach can be adapted to a wide range of situations.
The approach is evaluated considering economic design criteria and using an example scenario.
The evaluation shows that the proposed economic enhancements enable the provider to
increase his benefit. In the standard scenario a 92% of the maximum theoretically attainable
revenue with the enhancements can be achieved in contrast to 77% without enhancements. In
the scenario with partial resource failure the revenue is increased from 57% to 85% of the
theoretical maximum.
23

6.2.4 Situated Decision Support for Managing Service Level Agreement Negotiations
The paper “Situated Decision Support for Managing Service Level Agreement Negotiations”
tackles the interface between contract management and the bidding agent. This paper reports
about the results achieved during the cooperation with Rustam Vahidov (Concordia University,
Montreal) who co-author this paper.
This work proposes the application of the situated decision support approach to managing
automated SLA negotiations. The framework is based on the model for situated decision support
that effectively combines human judgment and autonomous decision making and action
by agent components. The key idea behind the approach lies in the managing the fleet of negotiating
agents by the use of a “manager” agent and human decision maker. It is shown
through simulation experiments how this approach performs under a set of simplifying assumptions.
6.2.5 Using k-pricing for Penalty Calculation in Grid Market
The paper “Using k-pricing for Penalty Calculation in Grid Market” is concerned with the
SLA enforcement component, by proposing penalty schemes for SLA violations. This research
has been conducted in cooperation with Omer Rana and Vikas Deora (both Cardiff
University) who also co-author the paper.
Essentially, the paper considers the important role of the design of service level agreements
(SLAs) to distribute risk in Grids. SLAs are crucial as they are contracts that determine the
price for a service at an agreed quality level as well as the penalties in case of SLA violation.
This paper proposes a price function over the quality of service (QoS) on the basis of the
agreements negotiated upon price and quality objective. This function defines fair prices for
every possible quality of a service, which are in line with the business of the customer and
incentivize the provider to supply welfare-maximizing quality. Therewith, penalties can be
calculated for every possible quality level as the difference between the agreed price and the
output of the price function for the effectively met quality. A price function according to the
k-pricing scheme is presented for a single service scenario and for a scenario with multiple
interdependent services.
6.2.6 Rightsizing of Incentives for collaborative e-Science Grid Applications with TES
The paper “Rightsizing of Incentives for collaborative e-Science Grid Applications with TES”
addresses the payment component. The work has been conducted within the scope of the research
project “Billing the Grid”, which strives for designing and implementing a billing system
for academia.
For business Grids the payment component is comparatively easy, as money can be used as
common denominator. In academia, the use of money is banned. Currently resources are used
purely on availability. This raises incentive problems, as there is no fair allocation that takes
the priority of a job into consideration. Researchers are using Grid resources regardless of
whether others need them more urgently. In addition, researchers often tend to not contribute
their own resources to the Grid, due to missing incentives. Thus, researchers are demanding a
Grid infrastructure, which sets the right incentives to share supporting a fair allocation of resources.
The most common instrument to establish incentives is the use of money. However,
the constraint in academia rules out the use of money.
In this paper, a mechanism called Token Exchange System (TES) is proposed. Considering
the requirements from the particle physics community this mechanism combines the advantages
of reputation and payment mechanism in one coherent system to realize a fair allocation.
It enables to build up a Grid infrastructure with an incentive mechanism which is scalable,
24

incentive-compatible and does not require a face-to-face agreement like the current system.
Transaction between institutes can be accomplished without burdening their financial endowment.
Simulations demonstrate that TES has a better performance than a common payment
system combined with a reputation mechanism.
6.3 Design Issues and Extensions
While the first two sections are focusing on economic models and algorithms for the components
of the Open Grid Market architecture, section 3 broadens the scope by addressing (i)
design issues that helps to introduce the results in practice and (ii) extensions of the results
concerning the applicability of the economic models and algorithms in other domains.
6.3.1 Distributed Ascending Proxy Auction – A Cryptographic Approach
The paper “Distributed Ascending Proxy Auction – A Cryptographic Approach” strives for
decentralizing auctions through the use of cryptography. As Grid computing is characterized
by decentralized management, the use of auctions is questionable as they are centralized in
nature. The results of the paper have been conducted with Michael Conrad and Christoph
Sorge from the Telematics area within the scope of the interdisciplinary research project “SE-
SAM”.
The paper presents a secure mechanism for distributed ascending proxy auctions. It fosters
privacy and overall correct conduct to a degree that enables the trusted application of an asynchronous,
iterative, second-price auction in systems as open as P2P networks.
The approach produces several desirable properties for the auction process: It eliminates the
dependency on one single auctioneer, and the winning bidder can hide his true valuation, respectively
his highest bid. Using an encrypted bid chain for bidding ensures that only a fraction
of information is revealed to each auctioneer. With any new bid chain each bidder can
freely decide, which auctioneer groups to trust. Robustness is achieved by forming groups of
auctioneers, where even one group member suffices to decrypt a bid step.
In contrast to previous distributed second-price auction mechanisms, this approach is suitable
for iterative open-cry auctions. It is never necessary for all participants to be online at the
same time, which is what makes other approaches highly vulnerable, as it allows the blocking
of an auction fairly by attacking any participating node. In the proposed protocol, all a bidder
needs to do is to convey his bid chains, whereas the auctioneer groups must be accessible during
the whole auction process. However, one single obedient auction group member being
online at a time suffices to conduct the auction with a sufficient standard of security.
Obviously, this paper showcases how inherently centralized auctions can be decentralized
through rigorous use of cryptography. The chosen auction is in particular difficult to decentralize.
It is assumed that the mechanisms presented in section 6.1. are much easier to decentralize.
6.3.2 Technology Assessment and Comparison: The Case of Auction and E-negotiation
Systems
The paper “Technology Assessment and Comparison: The Case of Auction and E-negotiation
Systems” investigates whether the use of auctions or negotiations are better suited for a use in
situations, where the trading object is defined by more than one attribute. The paper introduces
an Information Systems framework, which explains the antecedents of system performance
emphasizing the role of mechanisms. A laboratory experiment shows how the framework can
be used. This work helps to argue why the trading management uses auctions rather than negotiations.
The paper has been conducted with Gregory Kersten, Rustam Vahidov and Eva
25

Chen (all from Concordia University, Montreal) within the scope of the Canadian SSHRC
funded project “Electronic Negotiations, Media and Transactions for Socio-Economic Interactions”.
The purpose of this paper has been to propose a model for studying the impacts of electronic
exchange mechanisms on key variables of interests, both objective, as well as subjective ones.
The proposed TIMES model provides a framework that allows studying of types of exchange
mechanisms in their various implementations within different task, environment, and individual
contexts. These mechanisms could range from the simplest catalogue-based models to
advanced auction and negotiation schemas. Thus, the model can accommodate continuity in
the key design principles of the mechanisms, as opposed to considering them as distinct
lasses. Therefore, one of the key contributions of the model is that it enables the comparison
of various exchange structures in terms of the same set of key dependent factors.
6.3.3 Comparing Ingress and Egress Detection to Secure – Inter-domain Routing: An
Experimental Analysis
The paper “Comparing Ingress and Egress Detection to Secure – Inter-domain Routing: An
Experimental Analysis” addresses the issue of protocol security and adoption. The paper concentrates
on the protocol of secure Border Gateway Protocol. It is assumed that the results can
also be transferred to security mechanisms in the Open Grid market. The work has been conducted
together with Christoph Goebel (Humboldt University Berlin) and Ramayya Krishnan
(Carnegie Mellon University, Pittsburgh) who co-author this paper.
Starting point of the paper is that the global economy and society is increasingly dependent on
computer networks linked together by the Internet. The importance of networks reaches far
beyond the telecommunications sector since they have become a critical factor for many other
crucial infrastructures and markets. With threats mounting and security incidents becoming
more frequent, concerns about network security grow.
It is an acknowledged fact that some of the most fundamental network protocols that make the
Internet work are exposed to serious threats. One of them is the Border Gateway Protocol
(BGP) which determines how Internet traffic is routed through the topology of administratively
independent networks the Internet is comprised of. Despite the existence of a steadily
growing number of BGP security proposals, to date neither of them is even close to being
adopted.
The purpose of this work is to contemplate BGP security from a theoretical point of view in
order to take a first step toward understanding the factors that complicate secure BGP adoption.
Using a definition of BGP robustness we experimentally show that the degree of robustness
is distributed unequally across the administrative domains of the Internet, the so-called
Autonomous Systems (ASs). The experiments confirm the intuition that the contribution
ASs are able to make towards securing the correct working of the inter-domain routing infrastructure
by deploying countermeasures against routing attacks differ depending on their position
in the AS topology. It is shown that the degree of this asymmetry can be controlled by the
choice of the security strategy. The strengths and weaknesses of two fundamentally different
approaches in increasing the BGP’s robustness termed ingress and egress detection of false
route advertisements and are compared against each other. Depending on the comparison the
economic implications are discussed.
6.3.4 A Market Mechanism for Energy Allocation in Micro-CHP Grids
The last paper of this book “A Market Mechanism for Energy Allocation in Micro-CHP Grids”
strives for transferring the lessons learnt to other Grid markets. This paper addresses the
26

Energy market and motivates the use of markets for micro Grids. The results of the paper
have been conducted within the scope of the interdisciplinary research project “SESAM”.
Achieving a sustainable level of energy production and consumption is one of the major challenges
in modern society. This paper contributes to the objective of increasing energy efficiency
by introducing a market mechanism that facilitates the efficient matching of energy
(i.e. electricity and heat) demand and supply in Micro Energy Grids. More precisely a combinatorial
double auction mechanism is proposed that performs the allocation and pricing of
energy resources especially taking the specific requirements of energy producers and consumers
into account. The potential role of decentralized micro energy grids and their coupling to
the large scale power grid is outlined. Furthermore an emergency fail over procedure is introduced
that keeps the micro energy grid stable even in cases where the auction mechanism
fails. As the underlying energy allocation problem itself is NP-hard, a fast heuristic for finding
efficient supply and demand allocations is defined. Lastly, the applicability of this approach
is shown through numerical experiments.
7 Conclusion and Outlook on Future Research
In this section the work at hand is recapitulated. Moreover, an outlook on future research in
this domain is outlined.
7.1 Concluding remarks
The vision of a complete virtualization of Information and Communication Technology (ICT)
infrastructures by the provision of ICT resources like computing power or storage over Grid
infrastructures will make the development Open Grid Markets necessary. Over the Open Grid
Market idle or unused resources – computational resources as well as hardware – can be supplied
(e.g. as services) and client demand can be satisfied not only within organization but
also among multiple administrative domains.
In conventional IT environments, the decision “who shares with whom, at what time” is orchestrated
via central server that uses scheduling algorithms in order to maximize resource
utilization. Those central scheduling algorithms, however, have problems, when organizational
boundaries are crossed, and information about demand and supply are manipulated. In particular,
if demand exceeds supply, the scheduling algorithms fail to allocate the resources efficiently.
The further virtualization of ICT infrastructure and services requires technical solutions that
allow for a reasonable – i.e. an economically efficient – allocation of resources. Economic
mechanisms can be the basis for technological solutions that handle these problems by setting
the right incentives to accurately reveal information about demand and supply. Market or
pricing mechanisms foster information exchange and can therefore attain efficient allocations.
Establishing pricing mechanisms raises many questions:
• Can a price be calculated by only considering supply and demand, or will there be the
need for a regulatory body?
• How can the overall supply and demand volume be detected?
• Will participants communicate their supply and demand wishes openly in bids, even if
competitors have access to this information, e.g. to estimate production load?
• How can the bid generation process be automated so as to adequately reflect the underlying
business model?
In this work the design of an Open Grid Market is motivated. In addition, an architecture is
proposed as a blueprint for Grid markets. To establish an Open Grid Market in practice, there
are several obstacles that have to be overcome: The bidding process goes beyond the feasibili-
27

ty of mere manual configuration, so there is a need for intelligent tools that reduce the complexity
of Grid-based systems. In essence, those intelligent tools must support the automation
of the bidding process, which is dependent on the resource supply situation and on business
policies. Additionally, the Open Grid Market needs to be equipped with intelligent monitoring
tools that audit the resources continuously, in order to correct unexpected events such as demand
fluctuations, failure to share resources etc. Most of previous attempts to establish a vivid
market process have hitherto failed, as the underlying economic models were inadequate
or incomplete – most prototypes lacked intelligent tools for automation.
The set-up of an Open Grid Market has even more far-reaching ramifications on the ICT
management. Essentially, the Open Grid Market cannot only be used to allocate idle but all
available resources. In this case, ownership of resources does no longer play a central role in
the resource allocation process, as the cheapest resources that assure the required QoS are
allocated to the processes regardless from which organization they come. It is the price, reflecting
the demand and supply situation, which determines which resources are accessed.
The intelligent tools will help to make the Open Grid Market flexible and transparent enough
such that the doubts concerning market-based resource allocation can be resolved. This paper
tries to address all problems of an Open Grid Market that deal with economic models and
algorithms. It intends to bridge the gap between conceptual strength and real world implementations
of Grid platforms.
7.2 Outlook
Assuming that Grid markets will find their way into practice, more challenges need to be addressed.
Today’s competitive markets require companies to flexibly adapt their products and
services to meet rapidly changing business strategies and business models at low cost. Accommodating
innovative services and products is mainly associated with changes in the business
processes – and thus with IT systems embedding these processes – and in the dimensioning
of the IT infrastructure to deliver the services in proper quality. Service orientation in
software engineering grants companies the necessary flexibility by orchestrating services into
different workflows, thus paving the way for new business processes. The shift towards service
orientation also implies a transition from local optimization of the business processes to
collaborative and (administratively) distributed business processes.
Apparently, Service-Oriented Architectures are the best response to meet the flexibility needs
imposed by competition. In this context, Grid technologies complement service orientation to
manage the underlying IT infrastructure. As will be demonstrated within the articles, Grid
technologies are one way to get a grip on the growing complexity of IT infrastructures which
is going along with innovative enterprise applications requiring different hardware, operation
systems and application platforms. In addition, Grid technologies open up new possibilities in
dimensioning IT infrastructures. While service orientation makes resource demand very difficult
to predict, Grid technologies embedded in market-based business models facilitate ondemand
assignment of resources via access to a pool of resources. Instead of attributing a
fixed number of resources to services supporting specific business processes, Grid technologies
dynamically assign resources depending on the actual demand. This is in particular desirable
for SMEs, as it trims down the costs for IT infrastructure quite considerably. Accordingly,
service orientation and Grid technologies in unison are likely to revolutionize modern enterprises
in the future. It is the premise of this book that Open Grid Markets bolster the widespread
adoption of Grids. There are, however, many open problems associated with Grids and
Grid markets.
Currently, Grid markets almost exclusively deal with application-agnostic deployment
schemes. Trading CPU hours and memory is insufficient to support the deployment of complex
systems such as Enterprise Planning Systems (ERP). If application-independent services
28

are traded, the resource allocation problem will become even more complex due to the dependencies
between services. For instance, a workflow of service requests may require several
services in parallel or in sequential order.
All approaches described within this book are devoted to number-crunching activities taking
CPU as the scarce good. Taking into consideration that hardware and thus CPU gets way
cheaper, it raises the question, whether this is the real problem of Grid computing. The management
of storage or more precisely the distributed data management (e.g. replications) raises
many questions that could be used by referring to economic principles. In all of the presented
models, the impact of bandwidth in combination with the topology was widely neglected.
Assuming that bandwidth will become the critical factor there is a strong need to incorporate
this limiting factor into the resource allocation decision.
At the bottom line, the results of this book suggest several intriguing research avenues:
• Analyze the properties of the proposed mechanisms for the respective application
model classes.
• Implement the proposed economic models and algorithms and conduct field studies in
order to get real data.
• Develop market mechanisms, where theory is mostly silent (e.g. task-oriented applications).
• Develop sustainable business models for companies that provide market platforms for
trading Grid services or resources.
• Identify the size and the potential revenue of the single segments of the Grid market
• Identify limits of the use of market mechanisms in Grid.
29

Bibliography
Acharya, N., C. Chokkaredd, et al. (2001). The open computation exchange & auctioning
network (ocean), Department of Computer & Information Science & Engineering,
University of Florida.
Alkadi, I. and G. Alkadi (2006). "Grid Computing: The past, now, and future." Human Systems
Management 25(3): 161-166.
Andrzejak, A., M. Arlitt, et al. (2002). "Bounding the Resource Savings of Utility Computing
Models." Working Paper HPL-2002-339, Hewlett-Packard Laboratories, Palo Alto,
CA.
AuYoung, A., B. N. Chun, et al. (2004). Resource Allocation in Federated Distributed Computing
Infrastructures. Proceedings of the 1st Workshop on Operating System and
Architectural Support for the On-demand IT InfraStructure.
Barak, A., A. Shiloh, et al. (2005). An Organizational Grid of Federated MOSIX Clusters. CC
Grid, Cardiff, UK.
Bell, G. and J. Gray (2002). "What's Next in High-Performance Computing?" Communications
of the ACM 45(2): 91-95.
Berstis, V. (2002). "Fundamentals of Grid Computing." IBM Redbooks Paper, Austin, TX.
Butler, D. (2006). "Amazon puts network power online." Nature 444: 528.
Buyya, R. (1999). High Performance Cluster Computing: Architectures and Systems, Prentice
Hall.
Buyya, R., D. Abramson, et al. (2000). "Nimrod/G: An architecture of a resource management
and scheduling system in a global computational grid." Proceedings of the 4th International
Conference on High Performance Computing in Asia-Pacific Region: 283–289.
Buyya, R., D. Abramson, et al. (2004). The Grid Economy. IEEE Grid Computing.
Buyya, R., H. Stockinger, et al. (2001). Economic models for management of resources in
peer-to-peer and grid computing. International Conference on Commercial Applications
for High Performance Computing, Denver, CO.
Carr, N. (2003). "IT doesn't matter." Harvard Business Review 81(5): 41-49.
Carr, N. (2005). "The End of Corporate Computing." MIT Sloan Management Review 46(3):
66-73.
Cherkasova, L., D. Gupta, et al. (2007). "When Virtual is Harder than Real: Resource Allocation
Challenges in Virtual Machine Based IT Environments." HPL-2007-25.
De Fraja, G. and J. Sakovics (2001). "Walras Retrouve: Decentralized Trading Mechanisms
and the Competitive Price." Journal of Political Economy 109(4): 842-863.
Eymann, T., O. Ardaiz, et al. (2005). Catallaxy-based Grid Markets. First International Workshop
on Smart Grid Technologies (SGT05), Utrecht, Netherlands.
Eymann, T., M. Reinicke, et al. (2003). Exploring Decentralized Resource Allocation in Application
Layer Networks. Agent Based Simulation.
Eymann, T., M. Reinicke, et al. (2003). Self-Organizing Resource Allocation for Autonomic
Networks. Workshop on Autonomic Computing Systems (ACS/DEXA 2003), Prague,
Cz, IEEE Computer Society.
Fellows, W., S. Wallage, et al. (2007). Grid Computing - The state of the market,
The451Group.
Foster, I. and C. Kesselman (1998). The Grid: Blueprint for a New Computing Infrastructure,
Morgan Kaufmann.
Garlan, D. and M. Shaw (1993). An Introduction to Software Architecture. Advances in
Software Engineering and Knowledge Engineering. V. Ambriola and G.Tortora. New
Jersey, World Scientific Publishing Company. I.
Gomolski, B. (2003). "Gartner 2003 IT Spending and Staffing Survey Results." Gartner Research,
Stamford, Connecticut.
Hardin, G. (1968). "The Tragedy of the Commons." Science 162: 1243–1248.
30

Irwin, D. E., L. E. Grit, et al. (2004). "Balancing Risk and Reward in a Market-based Task
Service." Working Paper.
Kersten, G. E., S. Strecker, et al. (2004). Protocols for Electronic Negotiation Systems: Theoretical
Foundations and Design Issues. Proceedings of the 15th International Workshop
on Database and Expert Systems Applications (DEXA), Zaragosa, Spain.
Knight, W. (2006). "Unlocking the Grid." Engineering & Technology 1(3): 42- 45.
Kopparapu, C. (2002). Load Balancing Servers, Firewalls, and Caches, John Wiley & Sons.
Krishna, V. (2002). Auction Theory. San Diego, CA, Academic Press.
Lai, K. (2005). "Markets are Dead, Long Live Markets." Working Paper.
Lai, K., B. A. Huberman, et al. (2004). "Tycoon: a Distributed Market-based Resource Allocation
System." Working Paper.
Lai, K., L. Rasmusson, et al. (2004). "Tycoon: an Implementation of a Distributed, Marketbased
Resource Allocation System." Working Paper.
LSF. (2005). "LSF."
Macías, M., G. Smith, et al. (2007). Enforcing Service Level Agreements using an Economically
Enhanced Resource Manager. 1st Workshop on Economic Models and Algorithms
for Grid Systems (EMAGS 2007) In conjunction with the 8th IEEE/ACM International
Conference on Grid Computing (Grid 2007), Austin, Texas, USA.
Marcus, E. and H. Stern (2000). Blueprints for High Availability: Designing Resilient Distributed
Systems, John Wiley & Sons.
Neumann, D. (2006). Self-Organizing ICT Resource Management - Policy-based Automated
Bidding. eChallenges e-2006 Conference, Barcelona.
Neumann, D. (2008). Engineering Grid Markets. Negotiation and Market Engineering. H.
Gimpel, N. R. Jennings, G. Kersten, A. Ockenfels and C. Weinhardt, Springer.
Neumann, D., C. Holtmann, et al. (2006). "Grid Economics." Wirtschaftsinformatik 48(3):
206-209.
Neumann, D., J. Stoesser, et al. (2008). "A Framework for Commercial Grids - Economic and
Technical Challenges." Journal of Grid Computing: submitted.
Nisan, N., S. London, et al. (1998). Globally distributed computation over the internet - the
popcorn project. 18th International Conference on Distributed Computing Systems,
Amsterdam, The Netherlands, IEEE Computer Society.
Padala, P., C. Harrison, et al. (2003). Ocean: The open computation exchange and arbitration
network, a market approach to meta computing. International Symposium on Parallel
and Distributed Computing.
Rappa, M. A. (2004). "The Utility Business Model and the Future of Computing Services."
IBM System Journal 43(1): 32-42.
Regev, O. and N. Nisan (1998). The POPCORN market - an online market for computational
resources. First international conference on Information and computation economies,
Charleston, South Carolina, ACM Press.
Rosenschein, J. and G. Zlotkin (1994). Rules of Encounter: Designing Conventions for Automated
Negotiation among Computers. Boston, MIT Press.
Satterthwaite, M. and A. Shneyerov (2003). "Convergence of a Dynamic Matching and Bargaining
Market with Two-sided Incomplete Information to Perfect Competition."
Working Paper.
Schnizler, B., D. Neumann, et al. (2006). "Trading Grid Services - A Multi-attribute Combinatorial
Approach." European Journal of Operation Research: forthcoming.
Shneidman, J., C. Ng, et al. (2005). Why Markets Could (But Don’t Currently) Solve Resource
Allocation Problems in Systems. Proceedings of the 10th conference on Hot
Topics in Operating Systems.
31

Sullivan, D. G. and M. I. Seltzer (2000). "Isolation with Flexibility: a Resource Management
Framework for Central Servers." Proceedings of the USENIX Annual Technical Conference:
337–350.
Sun. (2005). "GridEngine."
Sutherland, I. E. (1968). "A futures market in computer time." Communications of the ACM
11(6): 449-451.
Waldspurger, C. A., T. Hogg, et al. (1992). "Spawn: A distributed computational economy."
IEEE Transactions on Software Engineering 18(2): 103–117.
Werthimer, D., J. Cobb, et al. (2001). "SETI@HOME-massively distributed computing for
SETI." Computing in Science and Engineering 3(1): 78-83.
Wolinsky, A. (1988). "Dynamic Markets with Competitive Bidding." Review of Economic
Studies, 55(1): 71-84.
Wolski, R., J. Brevik, et al. (2003). Grid resource allocation and control using computational
economies. Grid Computing - Making The Global Infrastructure a Reality, John Wiley
& Sons: chapter 32.
Wolski, R., J. Brevik, et al. (2004). Grid resource allocation and control using computational
economies. Grid Computing - Making The Global Infrastructure a Reality, John Wiley
& Sons: chapter 32.
Wolski, R., J. Plank, et al. (2001). "Analyzing market-based resource allocation strategies for
the computational grid." International Journal of High Performance Computing Applications
15(3): 258–281.
Wolski, R., J. Plank, et al. (2001). G-commerce: Market formulations controlling resource
allocation on the computational grid. Proceedings of the International Parallel and Distributed
Processing Symposium (IPDPS).
Yeo, C. S. and R. Buyya (2004). "A taxonomy of market-based resource management systems
for utility-driven cluster computing." Technical Report, GRIDS-TR-Grid Computing
and Distributed Systems Laboratory, University of Melbourne.
32

Part B
Articles
33

Bridging the Adoption Gap – Developing a Roadmap
for Trading in Grids
Dirk Neumann, Jochen Stoesser, Christof Weinhardt
Institute of Information Systems and
Management (IISM)
Universität Karlsruhe (TH)
{Neumann|Stoesser|Weinhardt}@iism.uni-karlsruhe.de
Introduction
Grid computing is increasingly gaining traction in a number of application areas. It describes a computing model
that distributes processing across an administratively and locally dispersed infrastructure. By connecting many
heterogeneous computing resources, virtual computer architectures are created, increasing the utilization of
otherwise idle resources (Foster et al. 2001). By the use of Grid technology, it is possible to set up virtual
supercomputers via the connection of normal servers that cost no more than US$3000 each. Adding and
removing servers is simple, granting extreme flexibility in building up infrastructures. The best example of what
Grid technology can achieve is illustrated by its prominent predecessor SETI@home: Over one million
computers spread across 226 countries were connected to reach a processing power of 418.6 TFLOPS (Cirne et
al. 2006). By comparison, the world’s fastest supercomputer, IBM’s BlueGene/L, has an estimated total
processing power of between 280 and 267 TFLOPS 1 , whereas Google's search engine can muster between 126
and 316 TFLOPS estimated total processing power. The business case for Grids is further underlined by
potential cost savings. It has been projected that Grids may decrease total IT costs by 30 % (Minoli 2004). Thus,
it is not surprising that Insight Research projects an increase in worldwide Grid spending from US$714.9
million in 2005 to approximately US$19.2 billion in 2010 (Anonymous 2006).
1 http://www.top500.org/, accessed on 18 September 2007
34

Currently, Grids are mainly employed within enterprises to connect internal divisions and business units. What is
needed to extend Grid technologies beyond company borders is a set of mechanisms that can enable users to
discover, negotiate, and pay for the use of Grid services on demand. According to The451Group, a leading Grid
research institute, the application of resource trading and allocation models is a crucial success factor for
establishing commercial Grids (Fellows et al. 2007). Recent Grid offerings by Sun Microsystems and
Amazon.com represent a first step in this direction. Sun introduced network.com, which offers computing
services for a fixed price of $1 per CPU hour 2 , while Amazon is currently launching comparable initiatives with
its Amazon Elastic Compute Cloud (Amazon EC2) and the Amazon Simple Storage Service (Amazon S3).
Despite these first approaches, electronic marketplaces for Grid resources have not yet taken off. Very few
customers are using Sun’s network.com. Due to different legal frameworks, network.com was only offered to
customers within the US. The adoption patterns in the US resemble those of the rest of the world – few
customers have adopted Grid markets. 3 But what is the reason for this limited success of Grid markets? Almost
every large computer hardware manufacturer like HP, Sun, or Intel has already worked on or at least considered
the options for Grid markets, but still no Grid market has successfully been launched, creating a Grid adoption
gap.
This paper attempts to explain why Grid market initiatives have failed. The explanation mainly focuses on the
object traded on Grid markets. The problem with current markets for Grids is that they are designed purely as
markets for physical resources. For example, Amazon’s Elastic Compute Cloud only aims to sell CPU hours.
This type of market is by design not relevant for enterprise customers who have deadlines for executing jobs and
who have no idea of how many resources are required to meet this deadline. As a result of the analysis in this
paper, it is concluded that a single Grid market for physical resources such as CPU and memory is insufficient to
ensure successful take-up of Grid computing across organizational boundaries. Instead, a set or catalogue of
different marketplaces is needed to satisfy the diverse needs of different market segments. This paper begins the
process of cataloguing the needed market mechanisms. Thus, this paper provides guidance for potential Grid
2
http://www.network.com/, accessed on 18 September 2007
3 Quite recently, Sun has been facilitating access to customers in 25 countries. It remains to be seen if the
removal of legal concerns will result in broader adoption.
35

market operators (e.g. Telecom companies or hardware vendors such as Sun Microsystems) in the choice of the
market mechanisms needed to increase the impact of Grid markets in commercial settings.
The remainder of this paper is structured as follows. Section 2 discusses ways to define the trading object in Grid
markets and thus forms the background for the discussion of market mechanisms. Section 3 explores whether
one marketplace alone is sufficient for meeting the needs of Grid providers and users. A two-tiered market
structure is proposed as a viable solution for structuring commercial Grids. Section 4 discusses the use of
different market mechanisms for this two-tiered market structure and shows which mechanisms are most
adequate for which kind of application. Section 5 concludes the paper with a summary and points to future work.
Background Discussion
Applications in a Grid environment can be deployed in two ways, either by directly accessing resources that are
distributed over the network or by invoking a Grid service 4 which encapsulates the respective resources behind
standardized interfaces. These alternative ways of deployment give rise to different requirements for potential
markets.
From a technical point of view, resources are simple to describe as there exists only a finite set of well-defined
resources. A resource may be characterized by its operating system (e.g. Windows, Linux), number and type of
CPUs (e.g. 4 * x86), memory (e.g. 128MB RAM), or any finite number of other attributes. The standardization
of resources offers a simple way to describe them semantically. Glue Schema 5 , for instance, provides a
standardized vocabulary for characterizing computing elements and their status. This in turn facilitates resource
discovery, as matchmaking within the finite domain space is straightforward.
Services on the other hand can be extremely difficult to describe. The service space is essentially infinite due to
the myriad of variations in service design. For the description of complex application services (e.g. a virtual web
4 Grid Services are stateful extensions of Web Services (http://www.globus.org/ogsa/, accessed on 18 September
2007).
5 See http://forge.gridforum.org/sf/projects/glue-wg, accessed on 18 September 2007
36

server service “Apache on Linux”), domain-dependent ontologies may become necessary. In the case of
“resource-near” services, henceforth called “raw application services” (for instance, computational services
which use only CPU and memory), standardized languages such as the Job Submission Description Language
(JSDL) 6
exist. Nonetheless, the indefinite search space tremendously exacerbates service description and
likewise service discovery.
Both resources and services are provided on the basis of Quality-of-Service (QoS) assertions, i.e. essentially
guarantees of access to specific resources or services at specific times and under specific conditions. For
resources, the QoS description is simple, as only standardized properties and the duration of resource access
matter. For services, however, QoS is more difficult, as not only time aspects but also precision and accuracy of
the services play a role. The definitions of precision, accuracy and further parameters depend on the individual
service and cannot be standardized. This also has ramifications for monitoring. While monitoring resource
access is relatively simple, the monitoring of complex application services becomes particularly demanding
when services are intertwined.
Figure 1 shows the different aggregation levels of services and resources with respect to the layers of the Grid
Protocol Architecture (Joseph et al. 2004). On top of this stack, the actual Grid application is located, such as a
demand forecasting tool for Supply Chain Management or computer-aided engineering tools for simulations, etc.
While these applications can be run directly on physical resources, they may access several complex application
services, e.g. services for integrating, aggregating and statistically analyzing vast amounts of data, to simplify
development and deployment by shielding parts of the Grid’s complexity from the application. Complex
application services are so diverse that they cannot reasonably be standardized. These complex application
services might in turn access raw application services which provide standardized interfaces for accessing
various data sources and or computational services. These raw application services are resource-near services
(such as storage, memory or computation services). Raw application services also comprise applications and
software libraries which can be standardized. These physical resources may be CPUs, memory, sensors, other
hardware and software or even aggregated resources such as clusters (e.g. a Condor cluster) and designated
computing nodes.
6 See http://www.ogf.org/documents/GFD.56.pdf, accessed on 18 September 2007
37

Figure 1: Grid Protocol Architecture and the different aggregation levels of services and resources
When relying on direct access to physical resources, executables and external libraries need to be transferred
across the Grid. Typically, state-of-the-art Grid middleware only supports limited resource management
functionality. In most cases, the middleware does not enforce policies concerning how many resources a job can
consume. Only the local administrator can specify the degree to which resources can be shared. Trading physical
resources is thus difficult to achieve by means of Grid middleware. Trading physical resources, however, is
possible on the operating system level, which supports effective resource management. So called Grid Operating
Systems, henceforth Grid OS (e.g. MOSIX 7 ), support resource management on the OS Kernel level and are
potentially available for setting up markets (Stößer et al. 2007). With Grid OS, applications need not be altered
to be run on the Grid, as is the case when Grid middleware is being used.
From an economic perspective, Grid resource markets are promising for automation via an organized electronic
market. There are standardized items for sale that potentially attract many buyers and sellers. However, complex
application services have the disadvantage that demand is highly specialized and distributed across niche
markets, such that only few potential buyers are interested in the same or related application services.
One Market Fits All?
In this section, we explore whether it is sufficient to build up and operate a single market for Grids. Based on the
previous background description, the answer to this question is straightforward: Designing one Grid market for
7 See http://www.mosix.org, accessed on 18 September 2007
38

all kinds of resources, from physical resources, such as processing power, memory and storage running on native
platforms, to sophisticated virtual resources or application services that bundle and enrich such physical
resources, seems inappropriate due to both technical and economical factors.
a) Technical factors
From the technical point of view, differences in the monitoring and deployment of services and resources
make it very difficult to devise a generic system capable of supporting all kinds of resource and service
trading. Furthermore, different deployment mechanisms impose different requirements on the market
mechanism, as we will outline below.
b) Economic factors
From the economic point of view, market mechanisms need to achieve the following standard objectives in
mechanism design (Stößer et al. 2007):
• Allocative efficiency: Allocative efficiency is the overall goal of market mechanisms for Grid resource
allocation. A mechanism is allocatively efficient if it maximizes the utility across all participating users
(welfare or overall “happiness”).
• Budget-balance: A mechanism is budget-balanced if it does not need to be subsidized by outside
payments.
• Computational tractability: The market mechanism needs to be computed in polynomial run time in
the number of resource requests and offers.
• Truthfulness: Truthfulness means that it is a (weakly) dominant strategy for users to reveal their true
valuations to the mechanism.
• Individual rationality: A mechanism is individually rational if users cannot suffer a loss in utility from
participating in the mechanism, i.e. if it is individually rational to participate.
As mentioned in the background discussion, trading physical resources and trading application services
impose completely different requirements on the market: while physical resources are more or less
commodities for which auction mechanisms seem to work well, (complex) application services are
inherently non-standardized, thus making auction-like mechanisms inapplicable.
39

From this brief discussion, it can be seen that a one-size-fits-all market for Grids is infeasible from both the
technical as well as from the economic perspective. Due to the heterogeneous properties of Grids, they can be
divided into two different types of markets: resource-near markets for physical resources and raw application
services on the one hand, and markets for complex application services on the other hand, spanning out what
may be called a “two-tiered Grid market structure” 8 as depicted in Figure 2. These two classes of markets are
analyzed below in terms of their requirements for market structure 9 .
Market for Complex
Application Services
Application
Complex
Application
Service
Market for Raw
Application Services
Raw
Application
Service
Application
Complex
Application
Service
Raw
Application
Service
Application
Application
Complex
Application
Service
Complex
Application
Service
Physical
Resource
Physical
Resource
Grid Users
Market for Physical Resources
Integrators
Resource Providers
Tier 2 Tier 1
Tier 1 Markets
Figure 2: Two-tiered market structure
In the resource-near market for physical resources and raw application services, low-level resources such as
processing power, memory, and storage are traded. Demand in this market is generated by complex application
services that need to be executed on these physical resources and raw application services. This setting poses
8 We use the term “market structure” to denote the configuration of marketplaces.
9 It should be noted that there could be n-intermediate markets. We consider only the extreme cases, as they
exhibit different characteristics.
40

special requirements for market mechanisms (Schnizler et al. forthcoming). However, as mentioned earlier, the
requirements also depend on the way resources are deployed.
Deployment as Physical Resource
The technical requirements in markets for physical resources are the following:
• Multi-attributes: Physical resources have quality attributes such as CPU speed, operating platform or
bandwidth. Thus the mechanism needs to cope with multiple attributes simultaneously.
• Bids on bundles: Generally, users require a combination of physical resources to execute an application
(e.g. CPU and memory). If the mechanism does not account for bids on bundles, the user faces the risk
of obtaining only one part of a bundle (the so called “exposure risk”). The market mechanism thus
needs to support requests for bundles of resources.
• Online mechanism: The allocation of the mechanism needs to be made instantaneously, as the market
assumes the role of an operating system scheduler. The mechanism thus needs to be lightweight such
that it requires little computation time. Online mechanisms are crucial in the case where information
such as release time or request processing time is only gradually released to the scheduler.. The
scheduler must be able to mitigate past decisions that prove unfortunate when new information enters
the system. For example, facing a decrease in the performance of an application, the mechanism may be
required to allocate additional physical resources in a timely manner.
• Split-proofness: In some scenarios, one might want the mechanism to treat small and large requests in
a fair manner. The mechanism may need to be split-proof in the sense that users cannot improve their
requests’ priority by splitting them into multiple parts and submitting them under different aliases.
• Merge-proofness: The mechanism must be stable in the face of strategic users, who build coalitions to
improve their priority. Thus the mechanism needs to assure that users do not have an advantage by
merging their requests.
Deployment as Raw Application Service
The technical requirements in markets for raw application services are the following:
41

• Multi-attributes: A virtual machine, for instance, may be characterized by the number of CPUs, share
of memory, cache and bandwidth of the underlying physical resource. A simple computational service
may need to provide a certain speed and accuracy.
• Bids on bundles: A user program may require multiple interdependent raw application services in
parallel.
• Time constraints: When raw application services are traded, the market mechanism needs to take time
attributes into account. The requesters need to specify their demand, so that the market mechanism can
efficiently schedule the requests according to the availability of resources and price. This situation
differs from the trading of resources, where the market mechanism executes requests and the
corresponding applications upon availability.
• Co-allocation: Capacity-demanding Grid applications usually require the simultaneous allocation of
several homogenous service instances from different providers. For example, a large-scale simulation
may require several computation services to be completed at the same time. This is often referred to as
“co-allocation”. A mechanism for raw application services has to enable and control co-allocations.
• Coupling: For some applications, it may be necessary to couple multiple raw application services into a
bundle in order to guarantee that the services are allocated from the same seller and will be executed on
the same machine.
• Resource isolation: Security and performance considerations lead to the requirement of resource
isolation, i.e. that a specific raw application service can only be instantiated once at any given time.
Tier 2 Markets
Along the lines of the two-tiered market structure, complex application services can be decomposed into several
raw application services that can in turn be translated into the physical resources required to execute these
services. For instance, some complex application service might require a basic XML transformer service, which
in turn needs processing power, memory, storage etc. Buyers in such a market request a complex application
service; the provider of this complex application service, the service integrator, is responsible for obtaining the
required raw application services and physical resources in turn, thus hiding parts of the Grid’s complexity from
the buyer. Such a hierarchical masking of complexity seems to be an appropriate approach since service
42

equesters typically have no insight into the resources the complex application service will consume (Eymann et
al. 2006).
The requirements of complex application services for market mechanisms are totally different than for resourcenear
markets. Complex application services are rarely used by two different companies; hence creating
competition via an auction mechanism does not make sense. Instead, the market for complex application services
faces the difficulty of finding a counterpart that offers the exact capabilities needed to execute the application.
As the following requirements suggest, the market mechanism is more “search-oriented”, in terms of the need
for bilateral or multi-lateral negotiation protocols:
• Multi-attributes (see above).
• Workflow support: To support complex services, distributed resources such as computational devices,
data, and applications need to be orchestrated while managing the application workflow within Grid
environments. The market mechanism needs to account for this during design time and run time of the
workflow.
• Scalability: Scalability considers how the properties of a protocol change, as the size of a system (i.e.
the participants in the Grid) increases.
• Co-allocation: (see above).
As pointed out in the introduction, resource-near Grid markets are not a viable solution for enterprises which
typically have to run time-critical applications. However, applications in academia are usually less time
dependent. As such, resource-near markets would be a viable business model for such settings; the users have to
wait until the queued jobs are executed. But clearly, the issue of payment is controversial in academia. Even for
the EGEE Grid 10 , billing and payment will soon become an issue as demand exceeds supply. It seems that
resource-near markets will soon become an adequate model for academic Grids such as EGEE or D-Grid, the
German Grid initiative.
Grid markets that will be widely accessed by enterprises need to be of the form of markets for complex
application services. Consider a manufacturer interested in executing a computer-aided engineering application
10 Enabling Grids for eScience, http://www.eu-egee.org/, accessed on 18 September 2007
43

and deploying it on a computationally intense platform. This complex application service showcases a very
specific service which is likely to be demanded by a single requester only. To accommodate this complex
application service, the service must be decomposed into its constituent raw application services and the required
physical resources. Integrators – companies such as EDO2 11, GigaSpaces 12 , etc. that specialize in aggregating
and disaggregating services into physical resources – are needed to facilitate this decomposition process. The
specialization stems from experience, allowing the identification of service needs by comparing each service
request with similar service requests in the past, where similarity is established in terms of algorithms, data
structures and sizes, etc. Telecommunication companies and hardware producers seeking to virtualize IT
infrastructures naturally have the interest and competency to become Grid services integrators. Table 1
summarizes this discussion.
Market level Tier 1-market (resource-near) Tier 2-market
Deployment Physical Resource Raw Application Service Complex Application Service
Description
languages
GLUE JSDL/RSL Domain-dependent ontologies
Time limits No Yes Yes
Application areas
Small enterprises,
Integrators and resource
All enterprises
academia
providers
Requirements
Multi-attributes, Bids on
Multi-attributes, Bids on
Multi-attributes,
bundles, Online
bundles, Time attributes, Co-
Workflow support,
mechanism, Split-proof,
Allocation, Coupling, Resource
Scalability,
Merge-proof
Isolation
Co-Allocation
A Roadmap for Grids
Table 1: Types of grid markets
In previous sections, we argued that Grids do not require something like a single global market where all Grid
requests and supplies are aggregated, but a more complex two-tiered market structure. Figure 2 above
summarizes this (meta-)market structure. Applications demand the execution of several complex application
services in Tier 2 markets. In resource-near markets (Tier 1), complex services request physical resources either
11 http://www.cdo2.com/, accessed on 18 September 2007
12 http://www.gigaspaces.com/, accessed on 18 September 2007
44

plainly deployed or accessed via service interfaces. Integrators assume the responsibility for mediating between
requesters unaware of their resource needs and the needed resources.
Based on the preceding discussion, we set up a taxonomy below of known market mechanisms that support
different types of Grid applications. This taxonomy is conceived as a roadmap for further Grid market
developments to help bridge the Grid adoption gap.
Market Mechanisms for Trading Physical Resources
For Grids where physical resources (but not services) are traded, no time restrictions apply. Usually, the
resources themselves are not traded, but rather shares of computing units (e.g. nodes) are. The idea is that
bidding determines the share a user receives from the pool of available nodes. The more resources a user obtains,
the faster the application is completed. The following mechanisms have been proposed for this setting:
• Fair Share (Kay and Lauder 1988): In the Fair Share mechanism, all users get the same share of the
respective resource. This share is dynamically adjusted as new users enter the system or existing users
leave.
• Proportional Share (Lai et al. 2004): In contrast to the Fair Share mechanism, with Proportional Share
the shares can differ across users to reflect each user’s priority. This priority may be determined
dynamically based on the users’ bids for resources: If there are n users in the system and user i has
submitted a valuation of v i , then i will receive a share amounting to v i
/ ∑ j =
v
1 j
.
n
• Pay-as-bid (Sanghavi and Hajek 2005; Bodenbenner et al. 2007): The pay-as-bid mechanism operates
in the same settings as Fair Share and Proportional Share but is specifically designed so as to induce
users to truthful reporting of valuations. It has been shown that pay-as-bid improves on the prominent
Proportional Share mechanism as regards efficiency and provider’s revenue (Bodenbenner et al. 2007).
All three mechanisms, however, share a common drawback: they can only be used in scenarios where one
resource provider serves several consumers, that is, there is no competition among providers. In cases where all
resources are under fully centralized control, this condition is unproblematic. But the idea in Grids is to cross
administrative boundaries. Consequently, there is a need for market mechanisms that support multiple resource
providers.
Market Mechanisms for Trading Application Services
45

Dividing the service market into two parts – raw and complex application services –is too simplistic, as the
timing of demand for services has not yet been considered. This timing is determined by the application itself
and depends on the task the application is performing. We use the term “application model” as a characterization
of the processing mode of the application. This encompasses in particular the workload of the application as well
as the interaction model between applications and the Grid middleware virtualizing the execution platform.
Depending on the application model, different requirements upon the market mechanisms emerge.
• Batch applications (e.g. data mining) are characterized by a planned execution and expected
termination time. Execution is serial and resource demands depend on the parameters, such as the size
of input data.
• Interactive applications (e.g. online data analysis) are applications that require services on demand,
depending on the interactions with users. In contrast with batch applications, it is not possible to plan
the execution and expected termination time of interactive applications far in advance, so unpredictable
peaks of requests can occur within a short time.
• Task-oriented applications are dynamically composed from (sub-)tasks to build more complex tasks.
Service demand depends on the (work-) flow of requests from multiple users (e.g. transaction system of
a bank constitutes).
Most of the market-based approaches relate to batch applications. Batch applications are comparably easy for
two main reasons. First, there is no need to consider a whole workflow with different resource demands on each
echelon of the workflow. Second, the time to determine the allocation can be relatively long; immediacy is not
essential. Thus, complex resource allocation computations can be performed without hampering the whole
application due to latency times devoted to the calculation of the optimal allocation. Furthermore, most of the
practical market-based Grid prototypes consider only one single resource type (e.g. CPU only) and thus make
use of standard auctions (e.g. English auction). In applications other than pure number crunching, those auction
types are inadequate as more than one object (e.g. memory) is required at the same time.
Market Mechanisms for Raw Application Services
As mentioned above, the market mechanisms for raw services depend on the application model. Mechanisms
that we identify as being adequate for batch applications are:
46

• Multi-attribute Combinatorial Auction (Bapna et al. 2006): In the model of Bapna et al., multiple
requesters and providers can trade both computing power and memory for a sequence of time slots.
First, an exact mechanism is introduced. By imposing fairness constraints and a common valuation
across all resource providers, the search space in the underlying combinatorial allocation problem is
structured so as to establish one common, truthful price per time slot for all accepted requests.
Additionally, this introduces a linear ordering across all jobs and time which reduces the complexity of
the allocation problem, which however still remains NP-hard 13 . To mitigate this computational
complexity, a fast, greedy heuristic is proposed at the expense of both truthfulness and efficiency.
• Multi-attribute Exchange (Stößer et al. 2007): The model of Stößer et al. extends the model of Bapna
et al. (2006) in that it accounts for strategic behavior not only on the demand side of the market but also
among resource providers. Stoesser et al. design a truthful scheduling heuristic for Grid operating
systems that achieves an approximately efficient allocation fast. A greedy heuristic is employed to solve
the scheduling problem. It is complemented with an adequate pricing scheme which assures that
reporting truthfully is a dominant strategy for resource requesters and payments to resource providers
are approximately truthful.
• MACE-mechanism (Schnizler et al. forthcoming): Schnizler et al. elaborate a Multi-Attribute
Combinatorial Exchange (MACE). Users are allowed to request and offer arbitrary bundles of computer
resources and can specify quality attributes of these resources. The scheduling problem in this
combinatorial setting is NP-hard, and the pricing scheme of MACE yields approximately truthful prices
on both sides of the market. Due to the NP-hardness of the problem, the mechanism is not applicable to
large scale settings and interactive applications that require the immediate allocation of resources.
• Combinatorial Scheduling Exchange (AuYoung et al. 2004): The Bellagio system also implements an
exact combinatorial exchange for computing resources (AuYoung et al. 2004). Its pricing is based on
the approximation of the truthful Vickrey-Clarke-Groves prices proposed by Parkes et al. (2001). As an
exact mechanism it shares the computational drawbacks of MACE.
• Augmented Proportional Share (Stoica et al. 1996): One major drawback of Proportional Share as
proposed by Lai et al. (2004) is that users do not get any QoS guarantee. Generally, users will receive a
13 Informally, for a complex / large problem instance, the allocation problem is not solvable with justifiable
47

larger or smaller share than required and constantly need to monitor their resource needs and the actual
allocation. To mitigate this deficiency, Stoica et al. propose an extension of the Proportional Share
mechanism so that users can request and receive a fixed share of the resource. Thus, this approach
inherits the benefits of the simplicity of Proportional Share and resource reservation if required.
For interactive applications it is impossible to predict demand for raw application services. Thus, market
mechanisms need to allocate these services continuously. This can be realized by frequent call mechanisms,
where bids are collected for a very short time span and right away cleared. This requires that the mechanism is
solvable in few milliseconds. The greedy approaches of the Multi-attribute Combinatorial Auction Heuristic
(Bapna et al. 2006) and the Multi-attribute Exchange (Stößer et al. 2007) are capable of approximately solving
the combinatorial allocation problem in the required timely manner. Alternatively, the mechanism could be a
true online mechanism, which allows the real-time job submission to available resources (e.g. nodes or clusters).
A third way is to introduce a derivative market to hedge against the risk of supernormal resource demand.
Adequate market-based schedulers for interactive applications comprise:
• Decentralized Local Greedy Mechanism (Heydenreich et al. 2006; Stößer et al. 2007): This
mechanism is designed to provide a scalable and stable auctioning protocol. It operates in an online
setting where jobs arrive over time. Its aim is to schedule these jobs so as to minimize the overall
weighted completion times. This scheduling mechanism is complemented by a pricing rule which
induces the users to truthful behavior.
• Augmented Proportional Share (Stoica et al. 1996): The shares which are allocated to the users and
the resulting prices can be adjusted very efficiently as new requests are submitted or requests are
finished. Moreover, users can choose their risk by either fixing the price – and receiving a
corresponding share which may or may not correspond to the true demand of their interactive
application – or by fixing their share and having to pay the corresponding price, which may or may not
match their valuation.
• Derivative Markets (Kenyon and Cheliotis 2002; Rasmusson 2002): Derivative markets perform two
basic functions, hedging and speculation, which become essential in the context of non-storable goods
such as Grid resources. This generates early price signals which support capacity planning for both
buyers and sellers. Rasmusson (2002) proposes the use of options to price network resources. Kenyon
effort.
48

and Cheliotis (2002) follow the same idea by proposing option contracts for network commodities such
as bandwidth. They elaborate a pricing scheme for options which specifically takes into account aspects
of the underlying network topology that may affect prices, e.g. the existence of alternative paths.
Options cannot only be used for hedging purposes but also for speculation (arbitrage) which contributes
liquidity to the market.
The requirements for market mechanisms that support task-oriented applications are very demanding, as all
constituents of the workflow need to be allocated simultaneously, since otherwise the application cannot be fully
executed and is of no value to the user. Currently, there are only bargaining protocols available that guide the
users in their search for all components of the workflow, e.g.:
• SNAP (Czajkowski et al. 2002): The Service Negotiation and Acquisition Protocol (SNAP) supports
the reliable management of Service Level Agreements (SLAs). In particular, SNAP supports the
process of negotiating simultaneously across multiple resource providers, a key requirement for
complex and resource-demanding services.
Market Mechanisms for Complex Application Services
Trading complex application services is very demanding as there are not many providers and requesters.
Currently, there is not much research available that aims to develop market mechanisms for trading complex
application services. Hence, the market mechanisms for batch, interactive and task-oriented applications do not
differ substantially. In particular, the MACE-mechanism (Schnizler et al. forthcoming), the Bargaining Protocol
(Czajkowski et al. 2002), and the Decentralized Local Greedy Mechanism (Heydenreich et al. 2006; Stößer et al.
2007) might be appropriate for trading such services. Furthermore, for all applications, the licensing model of
software-as-a-service 14 is applicable. It refers to take-it-or-leave-it pricing, where the vendor sets the price and
the users decide whether or not to purchase.
Table 2 summarizes the existing market mechanisms for each application model.
Application model Physical Resource Raw Application Service Complex Application Service
14 Software-as-a-service refers more to model of software delivery where a service provider (e.g. SAP) offers to
requesters applications that are specifically implemented for one-to-many hosting.
49

Batch
MACE
Multi-attribute Combinatorial
Auction
Augmented Proportional Share
MACE
SNAP
Software-as-a-service
Interactive
Fair Share
Proportional Share
Sanghavi-Hajek
Decentralized Local Greedy
Mechanism
Multi-attribute Auction
Augmented Proportional Share
Derivative Markets
Decentralized Local Greedy
Mechanism
SNAP
Software-as-a-service
Task-oriented
SNAP
SNAP
Software-as-a-service
Table 2: Market Mechanisms Canon for Grids
Conclusion
This paper argues that, despite its promising features, the technology of Grid computing has not yet been adopted
by enterprises due to an absence of adequate business models. Grid markets may help to overcome this obvious
“adoption gap”. This paper attempts to derive an economically sound set of market mechanisms based on a solid
understanding of the technical possibilities.
In the background section, we analyzed the characteristics of the object traded in Grids. This trading object is
closely associated with the deployment of software applications. The deployment directly on physical resources
or via raw application services has major ramifications for the trading object and consequently for the
requirements for market mechanisms. Physical resources are essentially commodities, whereas application
services can be both standardized commodities (raw application services) and unique entities (complex
application services).
Based on this analysis, we derived a two-tiered market structure along the distinction between physical resources
and application services, where each tier demands different market mechanisms. The first tier comprises the
markets for physical resources (e.g. CPU, memory) and raw application services. The second tier comprises the
markets for complex application services.
50

We then presented and classified existing Grid market mechanisms according to this market structure. At the
core of this paper, we argue that there is no single market that satisfies all purposes. Reflecting the distinct
requirements of different application models (physical resource vs. raw application service vs. complex
application service) and deployment modes (batch vs. interactive vs. task-oriented), a catalogue of co-existing
market mechanisms is needed. Thus, in order to overcome the adoption gap, Grid market operators such as Sun
Microsystems and Amazon must deploy multiple market mechanisms via a Grid market platform to satisfy the
needs of their enterprise customers, or else integrators must step in to bridge these niche markets.
In essence, this paper suggests several intriguing research avenues. Further research is needed to specify the
properties of both the various application scenarios and the needed market mechanisms. Since existing market
mechanisms represent rather theoretical constructs, they need to be deployed in real-world settings to verify the
underlying models and analytic results. However, existing mechanisms only represent a first step towards filling
the identified market structure and satisfying the various requirements. This basic catalogue needs to be extended
by means of improved and new mechanisms in order to further enhance efficiency and to account for the
strategic dimension inherent in Grid markets. And, of course, the commercial adoption and adaption of Grid
markets requires an analysis of the supply side as well as of the demand side, which was our focus in this paper.
An in-depth analysis of the supply side would address questions such as the following: What are sustainable
business models for companies that provide market platforms for trading Grid services or resources? And how
large is each of the individual markets within the two-tiered market structure?
Chargeable Grid services and sustainable business models are not the final or only answers to the Grid adoption
problem. Other factors, such as trust in the security and reliability of Grid technology, are also relevant. But
better technology alone will not ensure widespread adoption of the Grid: sound market mechanisms are also
necessary.
51

References List
Anonymous (2006). Grid Computing: A Vertical Market Perspective 2006-2011. Boonton, NJ, USA, The Insight
Research Corporation.
AuYoung, A., Chun, B. N., Snoeren, A. C. and Vahdat, A. (2004) 'Resource Allocation in Federated Distributed
Computing Infrastructures', Proceedings of the 1st Workshop on Operating System and Architectural
Support for the On-demand IT InfraStructure.
Bapna, R., Das, S., Garfinkel, R. and Stallaert, J. (2006) 'A market design for grid computing', INFORMS
Journal of Computing forthcoming.
Bodenbenner, P., Stößer, J. and Neumann, D. (2007) 'A Pay-as-Bid Mechanism for Pricing Utility Computing ',
20th Bled eConference, Merging and Emerging Technologies, Processes, and Institutions, Bled,
Slovenia.
Cirne, W., Brasileiro, F., Andrade, N., Costa, L. B., Andrade, A., Novaes, R. and Mowbray, M. (2006) 'Labs of
the World, Unite!!!' Journal of Grid Computing 4(3): 225-246.
Czajkowski, K., Foster, I., Kesselman, C., Sander, V. and Tuecke, S. (2002) 'SNAP: A Protocol for Negotiating
Service Level Agreements and Coordinating Resource Management in Distributed Systems', 8th
Workshop on Job Scheduling Strategies for Parallel Processing, Edinburgh, Scotland.
Eymann, T., Neumann, D., Reinicke, M., Schnizler, B., Streitberger, W. and Veit, D. (2006) 'On the Design of a
Two-Tiered Grid Market Structure ', Business Applications of P2P and Grid Computing, MKWI 2006.
Fellows, W., Wallage, S. and Davis, J. (2007). Grid Computing - The state of the market, The451Group.
Foster, I., Kesselman, C. and Tuecke, S. (2001) 'The Anatomy of the Grid - Enabling Scalable Virtual
Organizations', Euro-Par 2001 Parallel Processing, Springer.
Heydenreich, B., Müller, R. and Uetz., M. (2006) 'Decentralization and mechanism design for online machine
scheduling.' in L. Arge and R. Freivalds (eds.) SWAT, Springer, 136-147.
Kay, J. and Lauder, P. (1988) 'A Fair Share Scheduler', Communications of the ACM 31(1): 44-55.
Kenyon, C. and Cheliotis, G. (2002) 'Forward Price Dynamics and Option Design for Network Commodities',
Bachelier Finance Society 2nd World Conference, Crete.
52

Lai, K., Rasmusson, L., Adar, E., Sorkin, S., Zhang, L. and Huberman, B. A. (2004) 'Tycoon: an Implementation
of a Distributed, Market-based Resource Allocation System', Working Paper.
Minoli, D. (2004) A Networking Approach to Grid Computing, New York, John Wiley & Sons, Inc.
Rasmusson, L. (2002) Network capacity sharing with QoS as a financial derivative pricing problem: algorithms
and network design, Ph.D. thesis, Royal Institute of Technology, Stockholm.
Sanghavi, S. and Hajek, B. (2005) 'A new mechanism for the free-rider problem', ACM SIGCOMM Workshop
Peer-to-Peer Economics, Philadelphia, PA.
Schnizler, B., Neumann, D., Veit, D. and Weinhardt, C. (forthcoming) 'Trading Grid Services – A Multiattribute
Combinatorial Approach', European Journal of Operational Research.
Stoica, I., Abdel-Wahab, H., Jeffay, K., Baruah, S., G., J. and Plaxton, G. C. (1996) 'A Proportional Share
Resource Allocation Algorithm for Real-Time, Time-Shared Systems', IEEE Real-Time Systems
Symposium.
Stößer, J., Neumann, D. and Anandasivam, A. (2007) 'A Truthful Heuristic for Efficient Scheduling in Network-
Centric Grid OS', European Conference on Information Systems (ECIS 2007), St. Gallen.
Stößer, J., Roessle, C. and Neumann, D. (2007) 'Decentralized Online Resource Allocation for Dynamic Web
Service Applications ', IEEE Joint Conference on E-Commerce Technology (CEC'07) and Enterprise
Computing, E-Commerce and E-Services (EEE '07), Tokyo, Japan.
53

A TRUTHFUL HEURISTIC FOR EFFICIENT SCHEDULING IN
NETWORK-CENTRIC GRID OS
Jochen Stoesser, Institute of Information Systems and Management (IISM), Universität
Karlsruhe (TH), Englerstr. 14, 76131 Karlsruhe, Germany, stoesser@iism.uni-karlsruhe.de
Dirk Neumann, Institute of Information Systems and Management (IISM), Universität
Karlsruhe (TH), Englerstr. 14, 76131 Karlsruhe, Germany, neumann@iism.unikarlsruhe.de
Arun Anandasivam, Institute of Information Systems and Management (IISM), Universität
Karlsruhe (TH), Englerstr. 14, 76131 Karlsruhe, Germany, anandasivam@iism.unikarlsruhe.de
Abstract
The Grid is a promising concept to solve the dilemma of increasingly complex and demanding
applications being confronted with the need for a more efficient and flexible use of existing resources.
Network-centric Grid Operating Systems (OS) aim at providing users and applications with
transparent and seamless access to heterogeneous Grid resources across different administrative
domains. Scheduling in these heterogeneous environments becomes the key issue since scarce
resources must be distributed between strategic and self-interested users. Market-based algorithms
are deemed to provide a good fit to the Grid environment by accounting for its distinct properties and
the needs of the users. Current market mechanisms, however, leave room for inefficiency and are
computationally intractable. The speed of heuristics becomes a desirable feature in interactive, largescale
Grid OS settings. Heuristics achieve a fast, approximately efficient allocation of resources but
generally fail in preserving truthfulness, i.e. users can benefit from cheating the mechanism. The
contribution of this paper is the proposal of a greedy, market-based scheduling heuristic for networkcentric
Grid OS which does achieve this distinct trade-off: it is designed so as to obtain an
approximately efficient allocation schedule at low computational cost while accounting for strategic,
self-interested users in a heterogeneous environment.
Keywords: network-centric Grid OS, market-based scheduling, heuristic, truthful
1052
54

1 INTRODUCTION
Grids allow the aggregation and sharing of a wide variety of geographically dispersed computer
resources owned by different administrative units (Foster and Kesselman 2004). Grids are typically
used for scientific applications that require massive amounts of computational power and storage such
as protein folding, weather forecasts, or gene encoding. These applications potentially process and
store terabytes of data and the sharing of resources is necessary to reduce the enormous processing
time (Berlich et al. 2005; Cirne et al. 2006). The most famous predecessor of modern Grids is
SETI@home which connected in its peak time over one million computers spread across 226 countries
to achieve processing power of 418.6 TFLOPS (Cirne et al. 2006). For comparison, the world’s fastest
supercomputer – IBM’s BlueGene/L system – has an estimated total processing power of 280
TFLOPS, while Google's search engine system can muster between 126 and 316 TFLOPS. Grids offer
enormous potential by aggregating idle resources that are geographically distributed. For businesses, it
is projected that Grids may reduce total IT costs by 30 % (Minoli 2004).
The use of market mechanisms has often been suggested in situations that establish a price for
resources. The price being paid to resource owners that share idle resources reflects the scarcity of the
provided resource. Since consuming resources is not for free, demand will be restricted and shifted to
those time slots where the resources are really needed. The employment of market mechanisms in
distributed computing is not new. The first attempt was reported at Harvard University where auctions
were used to allocate computing time to users of the PDP-1 computer (Sutherland 1968). Other
examples are Tycoon, Bellagio, and Nimrod/G (Buyya et al. 2000; Lai et al. 2004).
From an economic point of view, the use of market mechanisms in Grid computing is promising, but
most of the potential in state-of-the-art Grid middleware such as Globus, UNICORE, and gLite is
being wasted. Current Grid middleware is based on the premise to share idle resources only and
computational jobs are submitted to these idle resources where they are being processed in batch
mode. It would be more efficient to allocate all available resources over markets. Ideally, resource
allocation should be on-demand such that interactive processing of jobs is permitted as well (Gorlatch
and Müller 2005).
Recently, a new stream of research on network-centric Grid operating systems (OS) has emerged in
Grid computing that especially addresses interactive applications. Network-centric Grid OS provide
the user with seamless and transparent access to Grid resources such as memory and computing
power. While current OS provide only limited support for Grid computing, network-centric Grid OS
adjust their view on the system at run-time contingent upon the particular resources. By means of
network-centric Grid OS, the Grid can be used in an interactive manner. This interactivity certainly
raises problems concerning bandwidth and latency – network-centric-Grid OS are an active research
topic (Padala and Wilson 2003; Gorlatch and Müller 2005). Nonetheless, there are currently some
implementations available that come close to this notion of a network-centric Grid OS. MOSIX, for
example, is realized as an OS virtualization layer that provides a single system image with the Linux
run time environment to users and applications. As such, it allows applications to run on remote nodes
as if they ran locally. Users can run their applications while MOSIX is automatically and transparently
seeking resources and migrating processes among nodes in order to improve overall performance
(Barak et al. 2005).
The concept of network-centric Grid OS coupled with markets is deemed promising to revolutionize
resource allocation in Grids. From their conception, network-centric Grid OS do not only allocate
resources being idle but all available resources. The employment of markets for allocating and
scheduling jobs to resources appears promising to achieve an efficient resource allocation in the
economic sense – allocating those jobs to the resources that value it most.
The problem of scheduling resources is, however, computationally demanding. To be harnessed for
network-centric Grid OS, market mechanisms needs to be solvable quickly. Currently, there is no
market mechanism available that is specifically tailored towards the needs of a network-centric Grid
1053
55

OS. This paper seeks to remedy this gap by designing a truthful scheduling heuristic for MOSIX that
achieves an approximately efficient allocation fast. The contributions of this paper are threefold:
Firstly, the paper designs a multi-attribute exchange that can be used for Grid OS. Secondly, a greedy
heuristic is employed to solve the scheduling problem. Thirdly, an adequate pricing scheme is
developed which assures that reporting truthfully is a dominant strategy for resource requesters and
payments to resource providers are approximately truthful.
The remainder of this paper is structured as follows. In Section 2, the market-based scheduling
problem of network-centric Grid OS systems is introduced. Section 3 suggests (i) a heuristic to solve
this problem and (ii) a pricing scheme that provides the market mechanism with desirable properties.
Section 4 discusses related work in the light of the presented market mechanisms. Section 5 concludes
the paper with a summary and points to future work.
2 MARKET-BASED SCHEDULING IN GRID OS
There are two classes of system users in network-centric Grid OS: Resource requesters who want to
use the computing resources offered in the system for solving a computational problem and resource
providers who want to sell idle resources. We introduce the term “job” to refer to a computational
problem and we will call an atomic bundle of resources on which the job or parts of it can be
computed a “node”, e.g. one server within a cluster. If a job gets allocated to a node, we will call this
job (node) a “winning” job (node). We intend to design a direct, sealed-bid mechanism which
allocates periodically: the mechanism collects resource requests and offers from the users for a period
of time and then allocates jobs to nodes on the basis of these submitted requests and offers. Users do
not get to know the requests and offers submitted by other users before the allocation is determined by
the mechanism.
2.1 Economic Requirements
The mechanism is intended to perform job scheduling in a distributed computing environment with
heterogeneous and selfish users. There are a number of economic requirements which a market-based
scheduling mechanism should ideally satisfy in this environment (Schnizler et al. forthcoming):
• Allocative efficiency: A mechanism is said to be allocatively efficient if it maximizes the
utility across all participating users (welfare), i.e. the sum over the valuations of all winning
resource requesters less the sum over the reservation prices of all winning resource providers.
• Budget-balance: the mechanism does not need to be subsidized by outside payments. The
payments coming from the resource requesters cover the payments made to the resource
providers.
• Computational tractability: the mechanism can be computed in polynomial run time in the
size of its input, i.e. the number of resource requests and offers.
• Truthfulness: it is a (weakly) dominant strategy for users to reveal their true valuations to the
mechanism.
• Individual rationality: users cannot suffer any loss in utility from participating in the
mechanism, i.e. it is individually rational to participate.
Budget-balance and individual rationality are hard constraints which the mechanism must suffice. If
these two requirements are not met, the mechanism will not be sustainable. Truthfulness is a desirable
feature since it tremendously simplifies the strategy space of the users; there is no need to reason about
the strategies of other users. Due to the celebrated impossibility result of Myerson and Satterthwaite
(1983), this inherently implies that allocative efficiency can only be approximated: There is no
exchange which is at the same time budget-balanced, individually rational, truthful and allocatively
efficient in equilibrium.
1054
56

There are three basic components to be designed: the bidding language which defines how requests
and offers are specified, the allocation algorithm which decides which job is to be executed on which
node at what time, and the pricing scheme which translates the resulting allocation schedule in
corresponding monetary transfers. In this section, the bidding language and the winner determination
problem will be formalized. The complexity of this scheduling problem forms the rationale for a
market-based heuristic. The pricing scheme will only be specified for the heuristic due to the focus of
this paper.
2.2 Bidding Language
There is on-going research on applying machine learning, regression and filtering techniques for
predicting run times of future jobs (Ali et al. 2004). In many cases a certain resource requester will
submit jobs to a distributed computing environment which are similar in terms of algorithms, data
structures, job size etc. Thus the run time of a new job is likely to follow the run times of “similar”
jobs in the past. Building on this research we will assume that the resources needed for the
computation of a job are known to the requester a priori. This comprises the amount of computing
power, memory, and the run time of the job. A resource requester wanting to compute a job j submits a
request (b j , c j , m j , s j , e j ) to the market mechanism where b j ∈ R + denotes the requester’s willingness to
pay per unit of computing power and time slot, c j ∈ℵand m j ∈ ℵ the minimum amount of computing
power and memory, s j ∈ ℵ the time slot at which the job needs to be started, and e j ∈ ℵ the time slot
until which the job needs to be run. A job can only be executed in its entireness (atomicity). It cannot
be parallelized on multiple nodes but migrate across different nodes over time. Note that job migration
is one of the main features which distinguish Grid OS in general and MOSIX in particular from
traditional machine scheduling domains. A resource offer containing node n consists of a tuple (r n , c¯n,
m¯ n, ε n , λ n ) where r n ∈ R + denotes the reservation price per unit of computing power and time slot, c¯n ∈
ℵ and m¯ n ∈ ℵ the maximum amount of computing power and memory, and ε n ∈ ℵ and λ n ∈ ℵ the
earliest and the latest time slot at which the resources are available. Contrary to atomic resource
requests, resources offered by some node are divisible and moreover freely disposable.
Example: The following resource requests and offers have been submitted to the system:
Job j b j c j m j s j e j Node n r n c¯n m¯ n ε n λ n
J1 11 60 50 2 3 N1 5 150 100 1 5
J2 10 80 50 1 3 N2 7 100 120 2 6
J3 9 90 100 2 5 N3 8 140 100 1 5
J4 8 130 100 4 5
J5 7 80 90 1 3
J6 6 70 50 2 4
J7 5 30 40 1 3
Table 1. Sample resource requests and offers.
Job J1 requests to be run in time slots 2 and 3 and requires at least 60 units of computing power and 50
units of memory in each time slot. The resource requester for job J1 is willing to pay up to $11 per unit
of computing power and time slot, i.e. $11 * 60 * 2 = $1,320 in total. The provider of node N1 offers
at most 150 units of computing power and 100 units of memory in time slots 1 to 5 and requires a
minimum payment of at least $5 per unit of computing power per time slot.
To simplify notation, in the following we will assume that each request and each offer has been
submitted by a separate user and we will use the terms request and job (offer and node)
interchangeably. In contrast to Schnizler et al. (forthcoming), this bidding language does not support a
combinatorial exchange where users can request and offer arbitrary sets of goods and possibly link
multiple requests and offers by means of logical operators. In the scenario at hand, each request and
offer specifies one and the same bundle of computing resources and memory and thus rather
implements a multi-attribute exchange (Bichler et al. 1999) by allowing the specification of multi-unit
bundles, e.g. 100 units of computing power and 50 units of memory. Note, however, that while the
bidding language is not combinatorial, the scheduling problem of allocating jobs to nodes remains a
combinatorial assignment problem as will be seen below.
1055
57

2.3 Exact Scheduling
Exact mechanisms such as the well-known Vickrey-Clarke-Groves (VCG) mechanism are guaranteed
to find an optimal solution to the scheduling problem. Let N be the set of nodes contained in resource
offerings, J the set of jobs contained in resource requests, T S (n) := {t ∈ ℵ | ε n ≤ t ≤ λ n } the set of time
slots in which node n ∈ N is available according to the respective resource offering, T b (j) := {t ∈ ℵ | s j
≤ t ≤ e j } the set of time slots in which job j ∈ J requests to get executed, and N(j) := {n ∈ N | T b (j) ⊆
T S (n), c j ≤ c¯n, m j ≤ m¯ n} the set of nodes which can potentially execute job j. We introduce the binary
decision variable x jnt ∈ {0,1} with x jnt = 1 if job j will be executed on node n in time slot t and x jnt = 0
otherwise. The scheduling problem can be formalized as the following mathematical integer
programme:
[SP]
max V : = c x ( b −r
)
x jnt
n∈N( j)
b
j∈J j t∈T ( j) n∈N( j)
jnt j n
b
s.. t x ∈{0,1}, j∈J, t∈T ( j), n∈N( j) ( C1)
jnt
∑
∑
∑
∑
∑ ∑ ∑
j∈J
j∈J
b
x ≤1, j∈J, t∈T ( j) ( C2)
jnt
S
x m ≤ m , n∈N, t∈T ( n) ( C3)
jnt j n
S
x c ≤ c , n∈N, t∈T ( n) ( C4)
jnt j n
∑
∑
b
x = ( e − s + 1) x , j∈J,
t∈T
( j) ( C5)
b
u∈T ( j) n∈N( j)
jnu j j n∈N( j)
jnt
b
≥ ∈ ∈ ∈
j jnt n
b x r , j J, n N( j), t T ( j) ( C6)
SP assigns jobs to nodes as to maximize welfare V. The economic scheduling scenario is encoded in
the constraints of this combinatorial assignment problem as follows: (C1) introduces x jnt as binary
decision variable. Jobs can only be assigned to nodes which provide the necessary resources during the
right time slots. (C2) ensures that a job is allocated to at most one node at a time. (C3) and (C4)
specify that for any given time slot and node, the total resource consumption in terms of memory and
computing power cannot exceed the resources provided by this node. (C5) defines atomicity, i.e. every
job is either executed as a whole or it is not executed at all. (C6) ensures that for each allocation of a
job to a node, the requester’s willingness to pay is equal to or greater than the provider’s reservation
price. The optimal allocation schedule X* := (x*) j∈J,n∈N,t∈T b(j) yielding welfare V* can be derived
directly from the solution to SP.
Example: For the sample resource requests and offers listed in Table 1, building and solving SP yields
the optimal allocation schedule X* shown in Figure 1. In time slot 1, only job J2 runs on node N1. The
requester of job J2 is willing to pay up to $10 [per computing cycle] * 80 [computing cycles] = $800
per time slot. The provider of node N1 requires a minimum payment of $5 per computing cycle and
time slot and receives – depending on the pricing which will be elaborated below – at least $5 * 80 =
$400. Thus the allocation of job J2 to node N1 in time slot 1 yields welfare of $800 - $400 = $400.
Across all time slots, X* generates welfare of V* = $3,420.
J1
J2
J3
J4
J5
J6
J7
Exact
scheduling
Slot 1 Slot 2 Slot 3 Slot 4 Slot 5 Slot 6
J2 J2
J2
J4 J4
J1 J1
Slot 1 Slot 2 Slot 3 Slot 4 Slot 5 Slot 6
Node N1
J3 J3 J3 J3 Node N2
Slot 1 Slot 2 Slot 3 Slot 4 Slot 5 Slot 6
Node N3
Figure 1.
Optimal Allocation Schedule.
1056
58

While this exact mechanism generates an optimal allocation schedule in terms of economic efficiency,
this optimality only comes at the expense of computational efficiency: SP is an instance of a multiple
knapsack problem (Ferrari 1994) and thus NP-hard. Computational costs of determining the allocation
and the pricing may become a problem in dynamic network-centric Grid OS with a large number of
resource requests and offers (Waldspurger et al. 1992). In these settings, a trade-off between efficiency
and computational complexity may be desirable to support interactive applications.
3 TRUTHFUL SCHEDULING HEURISTICS
Heuristics aim at obtaining a good but generally suboptimal (in terms of welfare) allocation fast.
Relaxing allocative efficiency not only impacts the outcome determination but also the pricing
scheme: compromising a mechanism in terms of allocative efficiency generally implies compromising
truthfulness as well (Mu’alem and Nisan 2002). Lehmann et al. (2002) and Mu’alem and Nisan
(2002), however, elaborate necessary and sufficient conditions that do yield truthful, approximating
mechanisms for a restricted class of users: known single-minded bidders. Informally, a bidder (user) is
known single-minded if she only desires one specific set of goods and this set is known to the
mechanism. Applying these conditions to the scenario above in order to design a truthful mechanism
suggests itself: in a non-combinatorial multi-attribute exchange such as a market-based scheduler for
Grid OS, resource requesters and providers are only allowed to request and offer one single bundle of
resources corresponding to the bundle of attributes – in the scenario at hand this is computing power
and memory.
3.1 A Greedy Scheduling Heuristic
Lehmann et al. (2002) and Mu’alem and Nisan (2002) propose greedy heuristics for constructing
truthful heuristics for single-unit combinatorial auctions with one seller and multiple buyers. We build
on this proposal and apply this greedy principle to construct a heuristic for the multi-attribute Grid
exchange at hand with multiple sellers and multiple buyers. This heuristic consists of two basic
phases:
1. Sort the requests j ∈ J and offers n ∈ N in non-increasing and non-decreasing order with
respect to some norm η;
2. Sequentially run through the resulting rankings and allocate the requests with the highest
ranking to the offers with the highest ranking (cf. pseudo code in Figure 2).
(1) Sort jobs j ∈ J in non-increasing order and nodes n ∈ N in non-decreasing order of η.
(2) Run sequentially through the job ranking, starting with the highest ranked job. For each job j:
(3) Run sequentially through the node ranking, beginning with the highest ranked node, and
check if j can be executed, i.e. whether conditions (b j ≥ r m ) ∧ (c j ≤ c¯m) ∧ (m j ≤ m¯ m) are
satisfied for some m ∈ {n 1 ,...,n j } ⊆ N, i.e. the nodes with the highest ranking which can
together accomodate job j, in time slots t ∈ T b (j).
(4) If so, allocate j to {n 1 ,...,n j }; Update the residual capacities of m ∈ {n 1 ,...,n j }, i.e. subtract
m j from m¯ m and c j from c¯m in time slots t ∈ T b (j).
(5) Continue at (2).
Figure 2. Heuristic for obtaining the initial allocation.
The allocative efficiency of this greedy heuristic essentially hinges on norm η used in the sorting phase
(Lehmann et al. 2002). This norm must lead to “efficient” rankings in the sense that the heuristic can
create an approximately efficient allocation based on these rankings. A straightforward choice is to use
b i and r i respectively as these set the valuation for a job (node) in relation to the amount of resources
requested (offered). In conjunction with the sequential allocation rule in phase two, the resulting
heuristic truly implements a greedy allocation algorithm: in each allocation step, it intends to
maximize b j – r n from the objective function of SP above.
1057
59

Example: Assume η = b j , j ∈ J, and η = r n , n ∈ N.
Job j b j c j m j s j e j Node n r n c¯n m¯ n ε n λ n
J1 11 60 50 2 3 N1 5 150 100 1 5
J2 10 80 50 1 3 N2 7 100 120 2 6
J3 9 90 100 2 5 N3 8 140 100 1 5
J4 8 130 100 4 5
J5 7 80 90 1 3
J6 6 70 50 2 4
J7 5 30 40 1 3
Figure 3. Example of the greedy heuristic.
Jobs J1 and J2 can be fully executed on node N1. In time slots 2 and 3, job J3 does not fit on N1 but
only on N2 which is next in the ranking. In time slots 4 and 5, J3 does fit on N1. Since J3 is allocated
to N1 in time slots 4 and 5, J4 does not fit on N1 anymore but only on N3, in contrast to the optimal
allocation X*. In total, X greedy yields welfare V greedy = $3,000 and an approximation ratio of V greedy / V* =
$3,000 / $3,420 ≈ 88 %. The greedy heuristic generates the following allocation schedule X greedy :
J1
J2
J3
J4
J5
J6
Heuristic
scheduling
Slot 1 Slot 2 Slot 3 Slot 4 Slot 5 Slot 6
J2 J2
J2
J3 J3
J1 J1
Slot 1 Slot 2 Slot 3 Slot 4 Slot 5 Slot 6
J3
J3
Slot 1 Slot 2 Slot 3 Slot 4 Slot 5 Slot 6
J7 J4 J4
Node N1
Node N2
Node N3
Figure 4. Greedy allocation schedule.
The question is: How can these allocations of requests to resources and vice versa be translated into
corresponding monetary transfers (prices and payments) according to the desired design criteria
introduced above? Mu’alem and Nisan (2002) and Lehmann et al. (2002) propose a pricing scheme
which generates truthful prices in restricted one-sided, combinatorial auctions with one provider and
multiple requesters. In the remainder, these theoretical results will be applied to the Grid OS context to
generate truthful prices for resource requesters. In contrast to the setting of Mu’alem and Nisan (2002)
and Lehmann et al. (2002), the Grid OS market consists of multiple requesters and multiple providers.
An algorithm will hence be proposed which yields approximately truthful payments to resource
providers.
3.2 Critical Value-based Pricing of Resource Requests
Following the spirit of the truthful Vickrey auction, the payment of job j corresponds to its critical
value given sets J and N of jobs and nodes (Lehmann et al. 2002, Mu’alem and Nisan 2002). The
critical value of φ j is the minimal valuation that j has to report in order to remain in the allocation
schedule, keeping all other requests and offers unchanged. Note that it is not possible to simply take
the valuation of the job with the highest ranking which does not get executed since this job may not be
executable in general due to capacity constraints. Moreover, removing j from the winning allocation
might change the allocation of other jobs within the winning allocation as well. To determine the
critical value for each winning job j, we need to determine the allocation without j: we successively
allocate all other jobs from the ranking and, after having allocated a job, check whether j can still be
accommodated (cf. pseudo code in Figure 5).
1058
60

(1) Sort jobs k ∈ J, k ≠ j, in non-increasing order and nodes n ∈ N in non-decreasing order of η.
(2) Is any job k ≠ j left in the ranking?
If not, set φ j = max r
m∈ { n1
,..., nj
) m
where {n 1 ,...,n j } are the nodes with the highest rankings
which can together accommodate j, and finish.
If so, select job k with the highest ranking.
(3) Can k be accommodated? If not, continue at (2). If so, allocate k to the nodes {n 1 ,...,n k } ⊆ N
with the highest ranking that can accommodate k, and update the residual capacities of m ∈
{n 1 ,...,n k }.
(4) Can j still be accommodated on some nodes {n 1 ,...,n j } ⊆ N? If not, set φ j = b k and finish.
(5) If so, set r: = maxm ∈{ n1
,..., nj
)
rm. Either r < b k , then continue at (2); or r ≥ b k , then it is cheaper
for j to push away k than to take the next available nodes {n 1 ,...,n j }; set φ j = b k and finish.
Figure 5.
Determining the critical value of job j and refining the initial allocation.
The critical value must be computed once for each winning job j. However, compared to the
computational intractability of the VCG mechanism, the computation now runs in polynomial time.
Every winning job j has to pay p greedy,j = (e j – s j + 1)c j φ j whereas every rejected job pays nothing.
Example: Applying the algorithm to the example above, we get p greedy,J1 = (e J1 – s J1 + 1)c J1 φ J1 =
120·b J7 = 120·r N1 = $600, i.e. the “cheapest” option for job J1 is to push away J7 by bidding $5 (plus
some ε) which happens to equal the minimum reservation price of node N1. In total, the mechanism
collects a revenue R of $6,400 from the resource requesters.
Job j J1 J2 J3 J4
φ j 5 5 7 8
p greedy,j 600 1,200 2,520 2,080
Table 2. Payments of resource requesters.
The greedy allocation scheme combined with critical-value based pricing creates a truthful scheduling
heuristic with respect to resource requesters: It is a weakly dominant strategy for resource requesters
to report their true valuation for a job (the required resources). The payment depends on the critical
value which is independent of the reported valuation. Assume a risk-neutral requester j has reported
her true valuation v j , i.e. b j = v j . Then there are four cases to be considered.
State\Action Decrease b j Increase b j
j was accepted Case 1 Case 2
j was rejected Case 3 Case 4
Table 3. Proof sketch.
Suppose j has won. Reporting a lower valuation (Case 1) might have resulted in j not being accepted
while it would not have reduced j’s payment. Reporting a higher valuation (Case 2) would not have
generated more utility either since j has been accepted anyways.
Now suppose j has been rejected (φ j > b j = v j ). Reporting a lower valuation (Case 3) would still have
left j outside the winning allocation. Reporting a higher valuation (Case 4) might even have resulted in
a loss: if j had been accepted, b j > φ j > v j .
3.3 Budget-balanced, Approximately Truthful Payments to Resource Providers
The concept of critical value-based pricing cannot be applied when determining the payments for
divisible resource offers. The concept requires a binary decision in the sense that an offer is either
fully executed or it is not executed at all. We allow divisibility of offers, i.e. not all offered resources
need to be used but the offer can be executed partially.
1059
61

The algorithm for determining the critical values ensures that each winning job pays at least the
reservation price of the node it is allocated to. If there is competition from another job for the same
slot, the critical value will exceed the reservation price. The question is: How can revenue R be
distributed to the resource providers so as to ensure budget-balance, individual rationality and to
approximate truthfulness? The VCG mechanism is generally not budget-balanced (Schnizler et al.
forthcoming) and, combined with the heuristic above, not truthful (Mu’alem and Nisan 2002). A
straightforward approach is to adopt the algorithm of Parkes et al. (2002) to our greedy heuristic. The
basic idea of their algorithm is to approximate the VCG mechanism’s truthfulness by computing
discounts Δ parkes that minimize the distance to the VCG-discounts Δ vick while ensuring budget-balance.
In our algorithm, we first compute Δ temp,n := V greedy – (V –n ) greedy for each winning node n. The
underlying assumption is that the greedy Δ temp,n (which may in total exceed the surplus
R – ∑ n∈N
v^n(X greedy ) from the request-side) approximate the truthful VCG-discounts Δ vick,n . We then
solve the mathematical programme
[BB greedy ]
( Δ Δ
temp greedy )
∑ Δ = −
n∈N
greedy,
n ∑
min L ,
Δ greedy
greedy, n temp,
n
greedy
( )
s.. t R vˆ
X ( C7)
n∈N
0 ≤Δ ≤Δ , n∈N ( C8)
The aim of BB greedy is to compute discounts Δ greedy,n as to approximate Δ temp,n in turn while ensuring
(strong) budget-balance by distributing the surplus R – ∑ n∈N
v^n(X greedy )) to resource providers (C7).
Parkes et al. (2002) suggest and examine various distance functions analytically and numerically and
recommend the “threshold function” L 2 (Δ temp, Δ greedy ) := ∑ n∈N
(Δ temp,n – Δ greedys,n ) 2 as it minimizes the
“residual degree of manipulation freedom” (Parkes et al. 2002), that is the maximum amount of utility
a user can gain from reporting untruthfully. Analogously to the algorithm of Parkes et al. (2002), the
optimal Δ greedy,n can be computed without explicitly having to solve the non-linear programme BB greedy .
Let ∞ = Δ temp,0 ≥ Δ temp,1 ≥ Δ temp,2 ≥ … ≥ Δ temp,|N| ≥ Δ temp,|N|+1 = 0 be the partial ordering over the
temporary greedy discounts for the winning nodes. Then, for the threshold function L 2 , Parkes et al.
(2002) show that there is an interval K and a unique point C t * within this interval such that
∑
greedy
( R−∑
vˆ
n ( X ))
K
Δ −
* i=
1 temp,
i
n∈N
C
temp K + t temp K
Δ ≤ = ≤ Δ ( C9)
, 1 ,
K
and Δ greedy,n = max{0, Δ temp,n - C t * } solves BB greedy . Node n receives a positive discount if n’s temporary
greedy discount is greater than the threshold C t * . Otherwise n does not get any discount. C t * can be
computed by running sequentially through the partially ordered temporary discounts and checking if
condition (C9) is satisfied.
Example: In the example, R = $6,400 > $6,040 = ∑ n∈N
v^n(X greedy ), i.e. there are $360 above the
reservation prices which are to be distributed as to approximate truthful payments. Solving BB greedy
with distance function L 2 leads to C t * = $1,680 and the greedy discounts listed in Table 4. The total
surplus of $360 is allocated to node N1, N2 and N3 only receive payments amounting to their
reservation prices.
Node n v^n(X greedy ) (V –n ) greedy p temp,n Δ temp,n p greedy,n Δ greedy,n
N1 -2,700 960 -4,740 2,040 -3,060 360
N2 -1,260 2,820 -1,440 180 -1,260 0
N3 -2,080 3,000 -2,080 0 -2,080 0
Σ -6,040 -8,260 2,220 -6,400 360
Table 4. Approximately truthful greedy payments.
n
1060
62

3.4 Evaluation of the Heuristic
The greedy heuristic is designed so as to obtain approximately efficient allocations fast. It runs in
polynomial time in the size of resource requests and offers: In the first phase, requests and offers get
sorted in O(|J|·log |J|) and O(|N|·log |N|). In the second phase, the greedy allocation scheme
sequentially runs through the |J| sorted requests and for each job j tries to allocate j to one of the |N|
sorted offers. Hence the allocation phase runs in O(|J|·|N|). Note that the feasibility of allocating a
specific job to a specific node can be tested in O(1) since the number of attributes which need to be
checked is constant. Inherently, this speed is only achieved at the expense of allocative efficiency
compared to exact mechanisms. In a theoretical worst case analysis, the approximation of the efficient
allocation can be made arbitrarily bad with respect to norm b j , r n suggested above by means of simple
examples. The pricing scheme of the greedy mechanism consists of two parts: a truthful pricing
scheme for resource requesters and an approximately truthful payment scheme for resource owners.
BB greedy links these two sides as to ensure budget-balance. The resulting prices are individual rational;
a user reporting her true valuation will not have to pay more than her reported valuation if her request
gets accepted and she will not have pay anything if she does not obtain the resources, and accordingly
for resource owners.
4 RELATED WORK
The study of market mechanisms for Grid and the implementation of running Grid market prototypes
have received significant attention in the past. SPAWN (Waldspurger et al. 1992) implements a market
for computing resources in which each workstation auctions off idle computing time to multiple
applications by means of a Vickrey auction. All resources are allocated to at most one application at a
time regardless of this application’s actual demand which yields low resource utilization. Chun and
Culler (1999) realized a prototpyical market for computing time in which one resource provider sells
computing time to multiple requesters. Resource requesters get allotted computing time proportional
to their share in the total reported valuation across all requesters. The POPCORN market (Regev and
Nisan 1998) is an online market for computing power which implements a Vickrey auction as well as
two double auctions. All of these approaches share two major drawbacks: First, they allow the
specification and trading of computing power only. But requesters require a bundle of resources such
as computing power, memory, and bandwidth. On the one hand, the approaches at hand thus lead to
inefficient allocations since requests with the same demand for computing power but different
memory requirements, for instance, are treated the same. On the other hand, requesters are exposed to
the risk of only being able to obtain one leg of the bundle of required resources without the other
(“exposure risk”, Schnizler et al. forthcoming). A second limitation of these approaches is that they do
not support reservations of resources in advance which are essential for Quality of Service assertions.
Schnizler et al. (forthcoming) propose a comprehensive model that targets these deficiencies. They
suggest the use of a multi-attribute combinatorial exchange (MACE). Users are allowed to request and
offer arbitrary bundles of grid resources and can specify quality attributes on these resources. MACE
implements an exact mechanism. The scheduling problem in this combinatorial setting is NP-hard, the
pricing scheme of MACE yields approximately truthful prices. With truthful prices, strategic users do
not have an incentive to report any valuation other than their true valuation. Due to the NP-hardness of
the problem, the mechanism is adequate for batch applications – the use for interactive applications is
rather limited.
The work of Bapna et al. (forthcoming) is most relevant to the work presented in this paper. In their
model, multiple requesters and providers can trade both computing power and memory for a sequence
of time slots. First, Bapna et al. (forthcoming) introduce an exact mechanism. By introducing fairness
constraints and imposing one common valuation across all resource providers, they structure the
search space in the underlying combinatorial allocation problem as to establish one common, truthful
price per time slot for all accepted requests. Additionally, this introduces a linear ordering across all
jobs and time which reduces the complexity of the allocation problem, which however still remains
1061
63

NP-hard. To mitigate this computational complexity, they thus propose a fast, greedy heuristic at the
expense of both truthfulness and efficiency.
Archer et al. (2003) also design a heuristic on the basis of Mu’alem and Nisan (2002) and Lehman et
al. (2002) which, however, cannot be applied to the Grid OS context since multiple items of resources
are traded, such as x computing cycles. The truthfulness of their mechanism, however, only holds for a
small number of items.
5 CONCLUSIONS AND FUTURE RESEARCH DIRECTIONS
In Section 1, we argued that a network-centric Grid OS coupled with market-based scheduling can
increase efficiency in Grid and cluster environments by adequately allocating all available resources.
This distinguishes network-centric Grid OS from state-of-the-art Grid middleware such as Globus,
gLite, and UNICORE which rely on batch processing of idle resources only. While there are highly
sophisticated market mechanisms available for Grid middleware (e.g. MACE), there are no
mechanisms available for network-centric Grid OS which rely on interactive application processing. In
Section 2 we formalized the market mechanism as a multi-attribute exchange in which resource
owners and consumers can publish their demand and supply. Exact market-based scheduling
mechanisms share one deficiency: they are NP-hard. While this may not be a problem in a cluster
setting with a small number of users, it may prove crucial in an interactive, large-scale Grid OS
environment. We thus proposed a greedy heuristic in Section 3 which performs a fast scheduling while
preserving truthfulness on the request-side and approximating truthfulness on the provisioning-side of
the market. In Section 4, we presented related work on market-based resource allocation in Grids.
In essence, our market mechanism suggests several intriguing research avenues. We intend to
implement a prototype of the heuristic for dynamic scheduling in MOSIX, a state-of-the-art Grid OS
presented in Section 1. Numerical analyses need to be performed to further analyze the heuristic’s
properties with respect to the economic and technical requirements in Grid OS. The approximation of
efficiency needs to be compared to the optimal allocation achieved by exact mechanisms. This ratio
essentially depends on the norm used in the ranking phase of the heuristic. More sophisticated norms
will be developed and analyzed. The run time of the presented scheduling mechanisms needs to be
evaluated to determine the critical number of requests and offers at which an exact mechanism might
become infeasible and to determine the speedup achieved by the heuristic. And finally, the
approximation of truthfulness on the provisioning-side of the market needs to be analyzed. It is
desirable to allow users to specify dependencies between resources, such as substitutability of
computing power for memory. The model in this paper hence needs to be expanded to allow for such
extended bidding logics.
Acknowledgement
This work has been partially funded by EU IST programme under grant 034286 “SORMA – Self-
Organizing ICT Resource Management”.
References
Ali, A., A. Anjum, J. Bunn, R. Cavanaugh, F. van Lingen, R. McClatchey, M. Atif Mehmood, H.
Newman, C. Steenberg, M. Thomas, I. Willers (2004). Predicting the Resource Requirements of a
Job Submission. Proceedings of Computing for High Energy Physics, Interlaken, Switzerland.
Archer, A., C. Papadimitriou, K. Talwar and E. Tardos (2003). An Approximate Truthful Mechanism
for Combinatorial Auctions with Single Parameter Agents. Proceedings of the 14 th annual ACM-
SIAM symposium on Discrete algorithms, Baltimore, Maryland, pp. 205-214.
Bapna, R., D. Sanjunkta, R. Garfinkel and J. Stallaert (forthcoming). A market design for grid
computing. INFORMS Journal of Computing.
1062
64

Barak, A., A. Shiloh and L. Amar (2005). An Organizational Grid of Federated MOSIX Clusters.
Proceedings of the 5 th IEEE International Symposium on Cluster Computing and the Grid
(CCGrid), Cardiff, Wales.
Berlich, R., M. Kunze and K. Schwarz (2005). Grid Computing in Europe: from research to
deployment. Proceedings of the 2005 Australasian Workshop on Grid computing and e-research,
Newcastle, Australia, 44, pp. 21-27.
Bichler, M, M. Kaukal and A. Segev (1999). Multi-attribute auctions for electronic procurement.
Proceedings of the 1 st IBM IAC Workshop on Internet Based Negotiation, Yorktown Heights, NY.
Buyya, R., D. Abramson and J. Giddy (2000). Nimrod/G: An Architecture for a Resource
Management and Scheduling System in a Global Computational Grid. Proceedings of the HPC
ASIA’2000, the 4 th International Conference on High Performance Computing in Asia-Pacific
Region, Beijing, China.
Chun, B. and D.E. Culler (1999). Market-based proportional resource sharing for clusters. Millenium
Project Research Report. Available at http://www.cs.berkeley.edu/~bnc/papers/market.pdf
(November 16, 2006).
Cirne, W., F. Brasileiro, N. Andrade, R. Santos, A. Andrade, R. Novaes and M. Mowbray (2006).
Labs of the World, Unite! Journal of Grid Computing, 4 (3), pp. 225-246.
Ferrari, C.E. (1994). On Combinatorial Optimization Problems Arising in Computer Systems Design.
Ph.D. Thesis, Technische Universität Berlin.
Foster, I., and C. Kesselman (2004). The Grid: Blueprint for a New Computing Infrastructure. Morgan
Kaufmann, San Francisco, 2 nd edition.
Gorlatch, S., and J. Müller (2005). From Interactive Applications Toward Network-Centric Operation
Systems. Workshop on Network-Centric OS, Brussels.
Lai, K., B.A. Huberman and L. Fine (2004). Tycoon: a Distributed Market-based Resource Allocation
System. Technical Report, Hewlett Packard.
Lehmann, D., L.I. O’Callaghan and Y. Shoham (2002). Truth Revelation in Approximately Efficient
Combinatorial Auctions. Journal of the ACM, 49 (5), pp. 577-602.
Minoli, D. (2004). A Networking Approach to Grid Computing. John Wiley & Sons, Hoboken, New
Jersey.
Mu’alem, A. and N. Nisan (2002). Truthful Approximation Mechanisms for Restricted Combinatorial
Auctions. In AAAI (poster). Also presented at Dagstuhl workshop on Electronic Market Design.
Myerson, R.B. and M.A. Satterthwaite (1983). Efficient mechanisms for bilateral trading. Journal of
Economic Theory, 28, pp. 265-281.
Padala, P. and J.N. Wilson (2003). GridOS: Operating System Services for Grid Architectures. LNCS
2913/2003, pp. 353-362. Springer Verlag, Berlin.
Parkes, D.C., J. Kalagnanam and M. Eso (2002). Achieving budget-balance with vickrey-based
payment schemes in exchanges. IBM Research Report, draft, March 13, 2002.
Regev, O. and N. Nisan (1998). The POPCORN market – an online market for computational
resources. Proceedings of the 1 st international conference on Information and computation
economies, New York, NY, pp. 148-157.
Schnizler, B., D. Neumann, D. Veit and C. Weinhardt (forthcoming). Trading Grid Services – A
Multi-attribute Combinatorial Approach. European Journal of Operational Research.
Sutherland, I.E. (1968). A futures market in computer time. Communications of the ACM, 11 (6), pp.
449-451.
Waldspurger, C., T. Hogg, B.A. Huberman, J.O. Kephart and W.S. Stornetta (1992). Spawn: A
Distributed Computational Economy. IEEE Transactions on Software Engineering, 18 (2), pp. 103-
117.
1063
65

Economically Enhanced MOSIX for Market-based Scheduling in Grid OS
Lior Amar ∗ , Jochen Stößer ∗∗ , Amnon Barak ∗ , Dirk Neumann ∗∗
∗ Institute of Computer Science
The Hebrew University of Jerusalem
Jerusalem, 91904 Israel
{lior,amnon}@cs.huji.ac.il
∗∗ Institute of Information Systems and Management
Universität Karlsruhe (TH)
Englerstr. 14, 76131 Karlsruhe, Germany
{stoesser,neumann}@iism.uni-karlsruhe.de
Abstract
Applying economic principles to grids is deemed promising
to improve the overall value provided by such systems.
By introducing market mechanisms, end users can influence
the allocation of resources by reporting valuations for
these resources. Current market-based schedulers, however,
are static, assume the availability of complete information
about jobs (in particular with respect to processing
times), and do not make use of the flexibility offered by computing
systems. In this paper, we present our progress in implementing
a novel market mechanism for MOSIX, a stateof-the-art
management system for computing clusters and
organizational grids. The market mechanism is designed
so as to be able to work in large-scale settings with selfish
agents. Facing incomplete information about jobs’ characteristics,
it dynamically allocates jobs to computing nodes
by leveraging preemption and process migration, two distinct
features offered by the MOSIX system.
1. Introduction
Grid computing denotes a computing paradigm where
computer resources are shared across administrative boundaries
in order to efficiently serve compute-intense applications
at low cost and to accommodate peak loads [6]. So
called “Grid Operating Systems” (Grid-OS) are a revolutionary
approach to grid computing [13]. Grid OS encapsulate
the access to and management of grids on the operating
system layer already. Applications can be left unchanged,
leading to ease of use and hence increased acceptance of
grid technologies. Moreover, instead of sharing only idle
resources, grid OSs can be leveraged to efficiently share all
resources.
MOSIX 1 is a state-of-the-art grid OS [1] that makes an
1 MOSIX R○ is a registered trademark of A. Barak and A. Shiloh.
x86 based Linux cluster and an organizational grid perform
like a single computer with multiple processors. A production
organizational grid with 15 MOSIX clusters (∼ 650
CPUs) in different departments is operational at Hebrew
University (see http://www.MOSIX.org/webmon). The features
of MOSIX allow better utilization of grid resources
by users who need to run demanding applications but cannot
afford such a large cluster.
A key issue in inter-organizational grids is the decision
which resource is allocated to whom. Current technical
schedulers are designed to maximize the resource utilization
or to balance the overall system load. In case of excess
demand, such schedulers do not maximize the overall value
of the system since users do not have means to report their
valuations for resources.
Market mechanisms are thus deemed promising to improve
resource allocation in grids – and thus the systems’
value – by explicitly involving users in the allocation process.
Users report their valuations for resources to the system
and are allocated resources and computational tasks
based on these reports. Corresponding monetary transfers
between resource requesters and providers serve so as to induce
users to report truthfully about resource demand and
supply, and to provide incentives to contribute resources to
the system.
Applying market mechanisms to resource allocation in
computing systems is not a new idea [16, 3, 12]. However,
this previous work is not tailored towards the use in
grid OS for three reasons: (1) The Bellagio system [3] and
the OCEAN system [12] are both located above the OS
layer and interoperate with existing grid middleware such
as Globus [5]. (2) From a market mechanism perspective,
the employed matching algorithms are not suited for grid
OS. The combinatorial allocation scheme of Bellagio and
its computationally hard pricing scheme make this mechanism
infeasible for settings with potentially large numbers
of users and where immediacy of the resource allocation is
crucial. The same holds for the negotiation protocol em-
66

ployed in OCEAN which places significant communicational
load on the system. (3) These existing market mechanisms
assume complete information about resource requests
and offers, in particular concerning processing times, and
determine static allocations which do not dynamically adapt
as the status of the system and its entities changes over time.
The contribution of this paper is twofold:
• We propose a market mechanism which is specifically
tailored towards the needs and technical features
of MOSIX so as to generate highly efficient resource
allocations in settings with large numbers of selfish
agents; and
• We present a first implementation of the proposed
market mechanism in MOSIX which serves as a proof
of our concept.
The paper is organized as follows. The next section
presents an overview of MOSIX and its main relevant features.
Section 3 presents the proposed market mechanism
we use to enhance the resource management of MOSIX. In
Section 4, we present the design and the pilot implementation
of the MOSIX Economic Infrastructure (MEI). Section
5 concludes the paper and points to future work.
2. MOSIX Background
MOSIX is a management system for clusters and organizational
grids [1]. Its main feature is to make all the participating
nodes perform like a single computer with multiple
processors, almost like an SMP. In a MOSIX system, users
can run applications by creating multiple processes, then let
MOSIX seek resources and automatically migrate processes
among nodes to improve the overall performance, without
changing the run-time environment of migrated processes.
The goal of MOSIX is to allow owners of nodes to share
their computational resources from time to time, while still
preserving the autonomy of the owners, allowing them to
disconnect their nodes from the grid at any time, without
sacrificing guest processes from other nodes.
MOSIX is implemented as a software layer that provides
applications with an unmodified Linux run-time environment.
Therefore, there is no need to change or even link applications
with any special library. Moreover, MOSIX supports
most additional Linux features that are relevant to ordinary,
non-threaded Linux applications, so that most Linux
programs can run unchanged (and be migrated). To submit
processes via MOSIX, users use the program mosrun. For
example, the command “mosrun -m450 myprog myprogargs”,
starts and later may migrate the program my-prog
only to nodes with at least 450 MB of free memory.
The relevant features of MOSIX to the work presented
in this paper are the automatic resource discovery, the process
migration and the freezing mechanisms. The following
subsections provide some detailes about these features.
Other features of MOSIX include a secure run-time environment
(sandbox), that prevents guest processes from accessing
local resources in hosting nodes; “live-queuing”
that preserves the full generic Linux environment of queued
jobs; gradual release of queued jobs, to prevent flooding of
the grid or any cluster; checkpoint and recovery and support
of batch jobs.
2.1. Automatic Resource Discovery
Resource discovery is performed by an on-line information
dissemination algorithm [2], providing each node with
the latest information about the availability and state of grid
resources. The dissemination is based on a randomized gossip
algorithm, in which each node regularly monitors the
state of its resources, including the CPU speed, current load,
free and used memory, etc. This information, along with
similar information that has been recently received by that
node is routinely sent to a randomly chosen node, where a
higher probability is given to choosing target nodes in the
local cluster. The outcome of this scheme is that each node
maintains a local
information-vector with information about all active
nodes in the local cluster and the grid. Any client requiring
information about cluster or grid nodes can simply access
the local node’s information-vector and use the stored
information. Information about newly available resources,
e.g., nodes that have just joined the grid, is gradually disseminated
across the active nodes, while information about
nodes in disconnected clusters is quickly phased out.
2.2. Process Migration and Scheduling
MOSIX supports cluster and grid-wide (preemptive)
process migration [4]. Process migration can be done either
automatically or manually. The migration itself amounts
to copying the memory image of the process and setting
its run-time environment. In MOSIX, the node where the
process was initially started is referred to as the process’s
home-node. After a process is migrated out of its homenode
all the system-calls of that process are forwarded to
and processed in the home-node. The resulting effect is that
there is no need to copy file or libraries to the remote nodes,
and a migrated process can be controlled from its homenode
as if it was running there.
Automatic migration decisions are based on (run-time)
process profiling and the latest information on availability
of grid resources, as provided by the information dissemination
algorithm. Process profiling is performed by continuously
collecting information about its characteristics, e.g.,
size, rates of system-calls, volume of IPC and I/O. This in-
67

formation is then used by competitive on-line algorithms [7]
to determine the best location for the process. These algorithms
take into account the respective speed and current
load of the nodes, the size of the migrated process vs. the
free memory available in different nodes, and the characteristics
of the processes. This way, when the profile of a
process changes or when new resources become available,
the algorithm automatically responds by considering reassignment
of processes to better locations.
2.3. Freezing Support
In a dynamic grid environment the availability of resources
may change over time, e.g., when clusters are connected
or disconnected from the grid. In MOSIX, guest processes
running on nodes that are about to be disconnected
are migrated to other nodes or back to their home-node. In
the last case, the available memory at the home-node may
not be sufficient to receive all returning processes.
To prevent the loss of running processes, MOSIX has a
freezing mechanism that can take any running MOSIX process,
suspend it, and store its memory image in a regular file
(on any accessible filesystem). This mechanism can be activated
automatically when a high load is detected, or manually
upon specific request. The freezing mechanism ensures
that a large number of processes can be handled without exhausting
CPU and memory.
Automatically frozen processes are reactivated in a circular
fashion in order to allow some work to be done without
overloading the owner’s nodes. Later, when more resources
become available, the load-balancing algorithm migrates
running processes away, thus allowing reactivation of
additional frozen processes.
3. A Market Mechanism for MOSIX
In order to enhance the economic value provided by the
system and to provide incentives to contribute resources
to the grid, we implemented a market mechanism in the
MOSIX system. In this section, we will specify our first
approach towards designing such a market mechanism for
MOSIX which accounts for its requirements – in particular
online scheduling without knowing processing times in
advance – while at the same time leveraging its distinct features
such as migration and freezing.
3.1. Requirements for Market-based
Scheduling
Scheduling in MOSIX gives rise to a number of requirements
towards the market mechanism (e.g. [14]). The most
prominent requirements are:
Immediacy: The timing of the allocation is crucial. The
allocation scheme must be scalable in order to be able to allocate
large numbers of resource requests and offers (thousands
of jobs and CPUs)
Allocative efficiency: There shall be no “waste” of resources;
the system is supposed to make optimal use of its
resources and thus maximize the system’s overall which is
returned to its users.
Truthfulness: Participants cannot benefit from cheating
the mechanism, i.e. reporting any valuation for a resource
other than their true valuation.
A market mechanism consists of three elementary components:
a bidding language which specifies what information
is exchanged between the mechanism and its users
(Subsection 3.2), an allocation scheme which determines
the assignment of offered resources to requests (Subsection
3.3), and a pricing scheme which determines corresponding
monetary transfers between the market mechanism
and its participating users (Subsection 3.4).
3.2. The Model
Let J be the set of resource requests. We assume that
request (or “job”) j ∈ J is fully described by the tuple
(v j , c j , m j , s j , l j ). Computing power is the central scarce
resource in the system. Consequently, v j ∈ R + denotes
j’s valuation (or “maximum willingness to pay”) per unit
of computing power and time (e.g. per standard processor
and hour of processing time), c j ∈ N and m j ∈ N are j’s
minimum required amount of computing power (e.g. number
of processors in case of a parallelizable job) and memory
(e.g. in MB) respectively, s j ∈ R + denotes j’s release
time (the point in time at which is reported to the mechanism),
and l j ∈ R + is j’s processing time (or length, e.g. in
hours). However, with processes being located on the operating
system level, users typically do not know the length
of their jobs before these jobs are actually completed. Consequently,
we assume that resource requesters submit the
tuple (v j , c j , m j ) to the mechanism at time s j .
Moreover, let N be the set of resource offers. We assume
that offer (or “node”) n ∈ N is fully described by the
tuple (r n , c n , m n , ɛ n , λ n ) where r n ∈ R + denotes n’s valuation
(or “reservation price”) per unit of computing power
and time, c n ∈ N and m n ∈ N is the maximum available
amount of computing power and memory respectively, and
ɛ n ∈ R + and λ n ∈ R + specify the timeframe during which
these resources are available. For now, we further assume
that nodes can potentially execute multiple jobs in parallel
but that one job can only be executed by one node at a time. 2
Again, we assume that resource providers do not know how
long they can contribute their resources, e.g. because they
2 In future extensions of this basic model we plan to also support parallelism.
68

do not know when these resources will be required by a local
process, and thus submit the tuple (r n , c n , m n ) to the
mechanism at time ɛ n .
This setting is oftentimes referred to as being an online
setting where information is gradually being released to the
mechanism as opposed to the offline setting where all information
is known to the mechanism at the time it makes its
allocation and pricing decisions. Furthermore, we are facing
a so called non-clairvoyant scheduling problem where
processing lengths are unknown [10].
Example: Assume at time t when the market mechanism
is to make its allocation and pricing decisions (“clearing”),
the sample resource requests and offers listed in Table 1
have been collected. Job J1 requires one processor and 200
Job j v j c j m j Node n r n c n m n
J1 1.8 1 200 N1 1.4 1 240
J2 2.2 2 100 N2 1.0 2 200
J3 2.0 1 80
J4 1.0 1 120
Table 1. Sample resource requests and offers
MB of memory. It is willing to pay up to $1.8 per processor
(and thus in total) per time unit. Node N2 is offering two
processors and 240 MB of memory while requiring at least
$1.0 per processor (and thus $2.0 in total) per time unit.
3.3. The Allocation Scheme
The mechanism clears at discrete points in time t ∈ R + .
At time t, the requests J(t) = {j ∈ J | s j ≤ t ≤ s j + l j }
and the offers N(t) = {n ∈ N | ɛ n ≤ t ≤ ɛ n + λ n } have
been collected by the market mechanism. As introduced
in Subsection 3.1, we intend to design and implement an
allocation scheme with the aim of maximizing allocative
efficiency (i.e. welfare), where allocative efficiency is
the sum over all users’ utility resulting from a specific
market outcome (allocations and prices) and may also be
interpreted as overall “happiness”. Let V (t) be the resulting
welfare at time t and V = ∑ t∈R +
V (t) the overall welfare
generated by the mechanism. Since we are facing an online
setting, in order to maximize V , we must maximize V (t)
for all t. This scheduling problem can be formalized as the
following integer program:
max X(t) V (t) = ∑ j∈J(t) c ∑
j n∈N(t) x jnt(v j − r n )
s.t. x jnt ∈ {0, 1}, v j ≥ r n , j ∈ J(t), n ∈ N(t) (1)
∑
n∈N(t) x jnt ≤ 1, j ∈ J(t) (2)
∑
j∈J(t) x jntc j ≤ c n , n ∈ N(t) (3)
∑
j∈J(t) x jntm j ≤ m n , n ∈ N(t) (4)
X(t) gives the (optimal) allocation at time t. Constraint
(1) introduces the binary decision variable x jnt with
x jnt = 1 if job j ∈ J(t) is allocated to node n ∈ N(t) at
time t and x jnt = 0 else. A job can only be allocated to
a node whose reservation price (its minimal required payment)
does not exceed this job’s maximum willingness to
pay. Furthermore, a job can only be allocated to one node at
a time (Constraint (2)). Constraints (3) and (4) enforce that
the resource requirements by all jobs allocated to one specific
node at a time do not exceed the amount of resources
available on this node.
This program is an instance of a Generalized Assignment
Problem and thus NP-hard [9]. But in settings with
large numbers of resource requests and offers, we must still
be able to determine allocations quickly. Heuristics are a
promising approach towards tackling this problem. They
are very speedy while at the same time being simple to understand,
implement, and extend.
We present the pseudocode for a greedy heuristic in Algorithm
1. We successively construct a feasible alloca-
Algorithm 1 Greedy heuristic
Require: A set J(t) of resource requests and a set N(t) of
resource offers.
Ensure: A feasible allocation X(t) = (x jnt ) of requests
j ∈ J(t) to offers n ∈ N(t).
1: init ∀j ∈ J(t), n ∈ N(t) : x jnt = 0
2: while J(t) ≠ ∅ do
3: Select and remove j ∈ argmax k∈J(t) v k
4: N ′ (t) = ∅
5: while N(t) ≠ ∅ do
6: Select and remove n ∈ argmin k∈N(t) r k
7: if c j ≤ c n and m j ≤ m n and r n ≤ v j then
8: x jnt = 1
9: c n = c n − c j and m n = m n − m j
10: N ′ (t) ← n and break
11: else
12: N ′ (t) ← n
13: end if
14: end while
15: N(t) ← N(t) ∪ N ′ (t)
16: end while
17: return X(t)
tion request-by-request by employing a single-pass priorityrule-based
scheduling heuristic, where the priority of requests
and offers is determined by their reported valuation.
3 By running through the list J(t) of requests in non-
3 Ties may be broken randomly or by considering the memory requirements,
for instance.
69

ascending order of v j and trying to allocate this request to
the cheapest possible offer n ∈ N(t), we greedily maximize
the term v j −r n in the objective function of the integer
program introduced above.
Example: Job J2 has the highest ranking and can only
be allocated to node N2. Consequently, there is only capacity
left on node N1, which is then assigned to job J3 which
is considered next in the job ranking. This allocation generates
welfare amounting to v J2 c J2 + v J3 c J3 − r N2 c N2 −
r N1 c N1 = $2.2 × 2 + $2.0 − $1.0 × 2 − $1.4 = $3 per time
unit.
This heuristic generates a feasible and approximately
efficient allocation at low computational cost. The lists
J(t) and N(t) can be sorted in O(|J(t)| log |J(t)|) and
O(|N(t)| log |N(t)|) respectively. The actual allocation
phase then runs in O(|J(t)||N(t)|).
In our example, the generated welfare happens to correspond
to the optimal welfare produced by solving the
scheduling problem exactly. However, in general the computational
speed of the heuristic will only come at the expense
of allocative efficiency [15]. This is exacerbated by
the fact that we are facing a non-clairvoyant online setting
in which earlier allocation decisions may prove unfortunate
(in terms of allocative efficiency) when new requests and offers
(and hence new information) arrive in the system. We
thus make use of the unique features provided by MOSIX to
improve upon such unfortunate decisions: If the allocation
scheme determines that a specific job j ∈ J(t) is allocated
to some resource n ∈ N(t) at time t (x jnt = 1), in the
next run at time t ′ the allocation scheme may decide to preempt
this currently running job in case a job with a higher
willingness to pay enters the system. This preempted job
may then either be migrated to another node in case the allocation
scheme decides to keep this job in the allocation
(x jmt ′ = 1, m ∈ N(t ′ ), m ≠ n), or it may suspend the
job and send it back to its home-node in case it decides to
not keep the job in the allocation ( ∑ n∈N(t ′ ) x jnt ′) = 0).
The implementation of this logic will be explained in detail
in Section 4.
3.4. The Pricing Scheme
In inter-organizational grid settings, we assume users to
be selfish and rational, meaning both resource requesters
and providers try to maximize their individual benefit from
participating in the system. Resource requesters, for instance,
may benefit from understating their valuation, thus
potentially lowering their price. We assume that resource
requesters can only misreport about their valuations (ṽ j ≠
v j ) and their release time (˜s j ≥ s j ). Overstating resource
requirements can only increase the risk of not being allocated
or having to pay a higher price. Understating resource
requirements can easily be detected and punished by the
system, e.g. by killing the job. We further assume that resource
providers do not have an incentive to misreport their
available resources. Again, overstating can easily be detected
and punished by the system whereas providers obviously
do not have an incentive to understate, as this cannot
increase their resulting allocation and thus their payment.
However, they can misreport about their reservation prices
(˜r n ≠ r n ).
Consequently, the users’ actions will generally not work
towards the social optimum and we need to complement the
allocation scheme with an “adequate” pricing scheme. With
truthful prices (in dominant strategies), users cannot benefit
from misreporting about their valuations, independently
of the other resource requests and offers in the system.
Building on the concept of “critical-value-based” prices as
elaborated in [8] and [11], we will first construct such a
scheme for determining truthful prices of resource requests
and will then complement this scheme so as to distribute
the resulting payments to the supply side of the market.
Critical-value-based Pricing of Resource Requests
Losing requests j ∈ J(t) (i.e. requests which have not been
accepted) are not required to pay anything, so we set the
price of these requests at time t to p j (t) = 0. For accepted
requests, we set p j (t) = φ j c j , where φ j (t) ∈ R + denotes
j’s critical value at time t. This critical value equals the
minimum valuation j would have had to report to the system
in order to remain in the allocation X(t).
The strongest plus of critical-value-based pricing is that
it generates truthful prices of resource requests [8, 11, 14].
Truthfulness is a powerful concept. With truthful prices, resource
requesters do not need to reason about other users’
bidding strategies, instead they can “simply” report their
true valuation to the system. They cannot benefit from
trying to cheat the mechanism. The drawback of criticalvalue-based
pricing is the additional computational burden
it adds to the mechanism. The allocation is not independent
of a specific request, meaning the allocation of the other requests
might change in case a winning request is removed
from the allocation. Consequently, in order to determine the
critical value φ j (t) of a specific request j at time t, we need
to determine the allocation without j.
Critical-value-based prices guarantee payments from
resource requesters which at least cover the reservation
prices of resource providers and potentially generate a surplus
exceeding these reservation prices. Intuitively, we will
want to distribute this surplus so as to also induce resource
providers to truthful reporting of their valuations. This,
however, is impossible, as analytically shown in [8] and
[15], and we can thus only aim to produce approximately
truthful payments.
70

Proportional Payments to Resource Providers
Proportional payments distribute the surplus generated by
critical-value-based pricing to resource providers according
to these providers’ contribution of computing power as
compared to the overall contribution. First numerical experiments
indicate that this approach exhibits desirable strategic
properties while only adding marginal computational
burden on the mechanism [15]. More formally, let
S(t) = ∑
p j (t) −
∑ ∑
j∈J(t)
˜r x jntc j
n
c n
j∈J(t)
n∈N(t)
be the surplus at time t, i.e. overall payments by resource
requesters less overall reservation prices. Then resource
provider n receives a payment of
∑
j∈J(t)
p n (t) = ˜r x ∑
jntc j
j∈J(t)
n +S(t)·
x jntc j
∑ ∑
c n
m∈N(t) j∈J(t) x jmtc j
per time unit the current allocation is valid.
Example: Assume that ties between J2 and J1 and between
J3 and J1 are broken in favor of J2 and J3. Then
the winning jobs J2 and J3 both need to outbid J1 in order
to remain in the allocation (φ J2 = φ J3 = $1.8). Consequently,
J2 has to pay 2 · $1.8 = $3.6, J3 has to pay
$1.8, and J1 and J4 do not have to pay anything. Node
N2 receives 2 · $1.0 + 2 3 · $2 ≈ $3.33 while N1 receives
$1.4 + 1 3 · $2 ≈ $2.07.
4. Economically Enhanced MOSIX
In this section we describe the enhancement of MOSIX
to support economy aware scheduling using the market
mechanism presented in the previous section. The MOSIX
features used were the automatic resource discovery, manual
process migration and manual freezing (the automatic
process migration and freezing were disabled). A layer
called the “MOSIX Economy Infrastructure” (MEI) was
built on top of the existing MOSIX system. The MEI is
composed of 3 main components: the provider component,
the client component and the market manager component.
Figure 1 shows a schematic view of those components.
4.1. The Provider Component
In our implementation each provider runs a daemon
called providerd which is responsible for managing the economic
properties of that provider. The economic properties
currently includes: the provider market status, its reservation
price and its current payment, see Figure 1. The market
status may be ON or OFF, where ON indicates the provider
is currently participating in the market. If the market status
is ON, the amount of time the provider is expected to stay
Figure 1. MEI components
ON is also reported. This amount of time is only used as
a hint and the provider may join or leave the market at any
given moment. For example, student farm nodes can join
the market from 22:00 until 8:00 the morning after (during
this time the farm is closed for students). The reservation
price specifies the minimal price the provider is willing to
accept in order to run processes, and the current payment
indicates how much the provider is currently getting payed
(if it is used by a job).
Those new economic properties were added to the information
system of MOSIX and are constantly circulated
among the cluster nodes. This way the market manager (see
below) can easily obtain information about all the available
providers by querying one node in the cluster and obtaining
its information vector. An example of the relevant information
which each provider supply is presented in Figure 2.
A command line tool called provider-ctl enables the
owners to online modify the economic properties of the
provider. For example an owner may set different reservation
prices for different nodes, or can set different time
windows for participating in the market for each node.
4.2. The Client Component
The program erun (see Figure 4) allow the user to specify
economic parameters in addition to the MOSIX standard
parameters. The economic parameters includes the
71

Figure 2. Sample provider XML information
job’s maximal price and its budget. The standard MOSIX
(non-economic) parameters are forwarded to the mosrun
program. For example, the command “erun -p40 -m180
myprog myprog-args“ submits a job with a maximal price
of 40 and a memory requirement of 180 MB.
Once an erun program is launched, it connects to a daemon
called assignd, on the local host, which is responsible
for managing all the economy-aware processes in each
node. The assignd keeps track of all the erun programs and
reporting to the market manager about newborn erun processes
and also about erun processes which finished to run.
Figure 3 shows an example of a message the assignd send
to the market manager when a new job is submitted by the
user.
Figure 3. Sample job XML information
In the figure it can be seen that each job has a unique jobid
which is used to identify the job in the cluster. The user
may specify the maximal amount of memory the job may
need. This property helps the system to avoid assigning
the job to nodes with not enough free memory. The cpunum
property specifies how many CPUs this job requires,
enabling the user to submit parallel jobs which need more
than one CPU to run. The max-pay property stands for the
maximal price (valuation) the user is willing to pay for running
the jobs on a standard CPU for one hour. And finally
the max-budget property let the user assign a limited budget
to this job. Once the job consumes this budget the job will
be suspended until the budget is increased again.
The assignd is also responsible for receiving instructions
regarding the allocation of erun processes from the market
manager. Once such instructions arrive, the assignd uses the
underlying MOSIX system to enforce the market decision.
For example, if the central market instruct the assignd to
move a given erun processes from one provider to another
provider, the assignd uses the manual migration capabilities
of MOSIX to send the erun to the new location. If on
the other hand the assignd is ordered to suspend a process,
then the process is frozen to the local disk using the MOSIX
freezing mechanism.
4.3. The Market Manager Component
The market manager component is responsible for performing
the economic scheduling decisions. It consists of a
daemon called marketd and a market solver program (which
we refer to as the solver). The marketd receives job requests
from client nodes (via the assignd’s) and collects information
about the available providers using the MOSIX automatic
resource discovery. Once there are jobs in the system
and also available providers, the marketd passes this
information to the solver and wait for the solver to output a
scheduling decision. The solver is implemented as shown in
Section 3. Once the schedule is computed, the solver sends
to the marketd a scheduling decision for each job in the system.
The marketd then forwards the scheduling decision of
each job to the assignd responsible for that job (which carry
on the decision). There are two possibilities for running the
solver, the first is periodically and the second is upon arrival
of a new process. Currently, periodic approach is used.
Beside changing the status of the job, the market mechanism
also specifies the payment of the job (in case it in
a running state), this payment is used to update the totalpayment
property of the job. So for each job we always
know how much money it spent until now. Figure 4 shows
two examples of scheduling decisions the solver may output.
The first is for job 12345 which schedules the job to run
on provider 132.65.174.10 with price of 10. In this case, assuming
the job is already running on another provider, the
assingd responsible for the job will migrate the job to the
new provider. The second schedule is for job 1233, but in
this case the job is suspended. Here the assignd responsible
for the job will use the freezing mechanism of MOSIX and
will freeze (suspend) the job in it home-node.
5. Conclusion & Future Work
We have presented an enhancement to the MOSIX system
which performs market-based scheduling instead of the
existing load-balancing-based scheduling. The process migration
and freezing capabilities allow our system to dy-
72

Figure 4. Sample scheduling decisions
namically allocate and reallocate jobs to any provider node,
not only to idle nodes. The market mechanism we propose
is specifically tailored towards the needs and technical features
of grid OS so as to generate highly efficient resource
allocations in large-scale settings. We presented a pilot implementation
of the economic enhancement of MOSIX.
Our work suggests several intriguing avenues for future
research. First, we plan to perform real runs of the system
with thousands of jobs and hundreds of providers. We
will research how to keep the scheduling period as short as
possible (several seconds) in order to keep the system responsive.
An interesting issue is supporting short jobs with
run times smaller than the clearing period of the market.
To do that we need to assign the jobs as soon as they arrive.
Here we plane to run a fast version of the market solver each
time a new jobs arrives, while running the full version periodically.
Another important goal is to introduce technical
limitations to the market in order to take into account both
performance and economic issues. For example, in the current
market, the migration cost of a job is not yet taken into
account. This can lead to a scenario where each new schedule
causes a large number of concurrent migrations, which
in turn can overload the network. We plan to enhance the
market by taking the cost of migration into account as well
as limiting the number of concurrent migrations. This will
make the system more stable and usable in practice.
Acknowledgements
We wish to thank Amnon Shiloh for his contribution in
the development of the erun and assignd programs, and for
his valuable ideas. This research was supported in part by
grants from the EU IST programme under grant no. 034286
SORMA – “Self-Organizing ICT Resource Management”.
[3] A. AuYoung, B. Chun, A. Snoeren, and A. Vahdat. Resource
allocation in federated distributed computing infrastructures.
Proceedings of the 1st Workshop on Operating
System and Architectural Support for the On-demand IT InfraStructure,
2004.
[4] A. Barak, A. Shiloh, and L. Amar. An organizational grid of
federated mosix clusters. In 5-th IEEE International Symposium
on Cluster Computing and the Grid (CCGrid05), May
2005.
[5] I. Foster and C. Kesselman. Globus: a Metacomputing Infrastructure
Toolkit. International Journal of High Performance
Computing Applications, 11(2):115, 1997.
[6] I. Foster, C. Kesselman, and S. Tuecke. The Anatomy of
the Grid: Enabling Scalable Virtual Organizations. International
Journal of High Performance Computing Applications,
15(3):200, 2001.
[7] A. Keren and A. Barak. Opportunity cost algorithms for
reduction of i/o and interprocess communication overhead in
a computing cluster. IEEE Tran. Parallel and Dist. Systems,
14(1):39–50, 2003.
[8] D. Lehmann, L. O’Callaghan, and Y. Shoham. Truth revelation
in approximately efficient combinatorial auctions. Journal
of the ACM (JACM), 49(5):577–602, 2002.
[9] S. Martello and P. Toth. Knapsack problems: algorithms
and computer implementations. John Wiley & Sons, Inc.
New York, NY, USA, 1990.
[10] R. Motwani, S. Phillips, and E. Torng. Nonclairvoyant
scheduling. Theoretical Computer Science, 130(1):17–47,
1994.
[11] A. Mualem and N. Nisan. Truthful approximation mechanisms
for restricted combinatorial auctions. Proc. 18th
National Conference on Artificial Intelligence (AAAI-02),
2002.
[12] P. Padala, C. Harrison, N. Pelfort, E. Jansen, M. Frank, and
C. Chokkareddy. OCEAN: the open computation exchange
and arbitration network, a market approach to meta computing.
Parallel and Distributed Computing, 2003. Proceedings.
Second International Symposium on, pages 185–192,
2003.
[13] P. Padala and J. Wilson. GridOS: Operating System Services
for Grid Architectures. Proceedings of International
Conference On HP computing, 2002.
[14] J. Stößer, D. Neumann, and A. Anandasivam. A Truthful
Heuristic for Efficient Scheduling in Network-Centric Grid
OS. Proceedings of the European Conference on Information
Systems (ECIS), 2007, to appear.
[15] J. Stößer, D. Neumann, and C. Weinhardt. Market-based
Pricing in Grids: On Strategic Manipulation and Computational
Cost. Submitted for publication.
[16] I. Sutherland. A futures market in computer time. Communications
of the ACM, 11(6):449–451, 1968.
References
[1] Mosix. http://www.MOSIX.org, 2007.
[2] L. Amar, A. Barak, Z. Drezner, and I. Peer. Gossip algorithms
for maintaining a distributed bulletin board with guaranteed
age properties. Submitted for publication.
73

GreedEx – A Scalable Clearing Mechanism for Utility Computing
Authors:
Jochen Stoesser, Dirk Neumann ∗
Institute of Information Systems and Management (IISM)
Universität Karlsruhe (TH)
{stoesser,neumann}@iism.uni-karlsruhe.de
Englerstr. 14, 76131 Karlsruhe, Germany
Corresponding author:
Jochen Stoesser
Institute of Information Systems and Management (IISM)
Universität Karlsruhe (TH)
Englerstr. 14, 76131 Karlsruhe, Germany
Email stoesser@iism.uni-karlsruhe.de
Phone +49–721–6088372
Fax +49–721–6088399
∗ The authors have been partially funded by the EU IST programme under grant 034286 “SORMA –
Self-Organizing ICT Resource Management”.
74

Abstract
Scheduling becomes key in dynamic and heterogeneous utility computing settings.
Market-based scheduling offers to increase efficiency of the resource allocation
and provides incentives to offer computer resources and services. Current
market mechanisms, however, are inefficient and computationally intractable in
large-scale settings.
The contribution of this paper is the proposal as well as analytical and numerical
evaluation of GreedEx, an exchange for clearing utility computing markets,
based on a greedy heuristic, that does achieve a distinct trade-off: GreedEx
obtains fast and near-optimal resource allocations while generating prices that
are truthful on the demand-side and approximately truthful on the supply-side.
Keywords: Market-based Scheduling, Scalable Heuristic, Truthful
75

In current business domains, information technology (IT) departments are forced to cut
down on IT expenses but are at the same time required to provide the business side with
systems that can flexibly be adjusted to new business requirements and processes. This
is reflected in the vision of utility computing where computer resources can be accessed
dynamically in analogy to electricity and water: “Utility computing is the on-demand delivery
of infrastructure, applications, and business processes in a security-rich, shared, scalable, and
standards-based computer environment over the Internet for a fee. Customers will tap into
IT resources – and pay for them – as easily as they now get their electricity or water”
[Rappa, 1]. According to Rappa, utilities are characterized by necessity, reliability, ease of
use, fluctuating utilization patterns, and economies of scale [Rappa, 1].
Computer resources match this profile very well. Rather than being a strategic asset, IT
has become a basic necessity for almost all businesses [Carr, 2]. If IT systems fail, this can
result in unsatisfied customers and severe contractual penalties. Moreover, IT systems must
be easy to use for customers and employees and must adapt and connect to heterogeneous
systems. However, especially small- and medium-size companies will not constantly require
massive amounts of computer resources. Instead, these resources will be required in the
design phase of a new product only or to create daily/monthly/yearly reports, leading to
fairly dynamic and unpredictable (from the provider’s point of view) demand. The validity
of the business model “utility computing” mainly stems from its immense economies of scale.
Instead of having each company and research center to maintain its own costly computing
and data centers, utility computing providers benefit from economies of scale and offer these
resources on-demand. Setting up computing and data centers for the first customer takes
tremendous fixed costs, but serving an incremental customer or request with these existing
resources only requires (comparably) minor efforts. Sun’s One-Dollar-Per-CPU-Hour and
Amazon’s Elastic Compute Cloud are prominent examples for the industry take-up of utility
computing offerings [Sun, 3; Amazon, 4].
Rappa suggests to base pricing in utility computing on metering usage (also coined “paywhat-you-use”
or “pay-as-you-go”), as is the case with classic utilities such as water, telephone
and Internet access [Rappa, 1]. This approach has also been adopted by Sun ($1/CPU-hour)
and Amazon (e.g. $0.20 per GB of data transfer). However, even in this metered model,
prices are temporarily static and do not fully reflect the dynamicity of demand and supply.
Moreover, setting appropriate prices is a complex task for utility computing providers.
The use of market mechanisms has often been suggested in situations in which allocations
of and prices for scarce resources need to be established. The prices being paid reflect the
resources’ scarcity. Consequently, dynamic market prices support on the one hand resource
requesters in coordinating and distributing their demand over time, thus balancing the system
load. On the other hand, utility computing providers have an incentive to offer resources
in return for the payment of the market price. In dynamic, large-scale utility computing settings,
however, the allocation problem of the market is computationally demanding. Market
mechanisms must be solvable quickly. Currently, there is no market mechanism which can
fully satisfy the economic requirements of utility computing markets.
The contribution of this paper is the proposal as well as analytical and numerical evaluation
of GreedEx, an exchange for clearing utility computing markets, based on a greedy
heuristic, that does achieve a distinct trade-off: it obtains fast and near-optimal resource
allocations while generating prices that are truthful on the demand-side and approximately
76

truthful on the supply-side. GreedEx may be used for two purposes:
• Utility computing providers may choose to run a proprietary GreedEx-based market
platform to allocate their scarce computing resources to resource requests in an efficient
manner and to dynamically price these requests.
• A GreedEx-based platform may be run as an intermediary market which aggregates
demand as well as supply across multiple resource requesters and utility computing
providers, thus increasing liquidity and efficiency.
This paper is structured as follows. Section 1 presents related work on market-based
scheduling. In Section 2, we elaborate the heuristic with its allocation and pricing scheme.
The heuristic’s speedup and its efficiency are evaluated numerically in Section 3 before Section
4 concludes the paper and points to promising future research directions.
1 Related Work
There are two main streams of research in market-based allocation of utility resources: mechanisms
which consider the trading of one type of resource only and mechanisms which account
for dependencies between multiple utility resources.
Spawn implements a market for computing resources in which each workstation auctions
off idle computing time to multiple applications by means of a Vickrey auction [Waldspurger
et al., 5]. All resources are allocated to at most one application at a time regardless of
this application’s actual demand which yields low resource utilization. Chun and Culler
present a market for computing time in which one resource provider sells computing time
to multiple requesters [Chun and Culler, 6]. Resource requesters get allotted computing
time proportional to their share in the total reported valuation across all requesters. The
Popcorn market is an online market for computing power which implements a Vickrey
auction as well as two double auctions [Regev and Nisan, 7]. All of these approaches share
two major drawbacks: First, they allow the specification and trading of computing power
only. But requesters require a bundle of resources such as computing power, memory, and
bandwidth. On the one hand, the approaches at hand thus lead to inefficient allocations since
requests with the same demand for computing power but different memory requirements, for
instance, are treated the same. On the other hand, requesters are exposed to the risk of only
being able to obtain one leg of the bundle of required resources without the other (“exposure
risk”). A second limitation of these approaches is that they do not support reservations of
resources in advance which are essential for Quality of Service assertions [Foster, 8].
Schnizler et al. [Schnizler et al., 9] elaborate a multi-attribute combinatorial exchange
(MACE) which targets these deficiencies. Users are allowed to request and offer arbitrary
bundles of computer resources and can specify quality attributes on these resources. MACE
implements an exact mechanism. The scheduling problem in this combinatorial setting is
NP-hard, the pricing scheme of MACE yields approximately truthful prices. With truthful
prices, strategic users do not have an incentive to report any valuation other than their true
valuation. Due to the NP-hardness of the problem, the mechanism is adequate for batch
applications – the use for interactive applications is rather limited. The Bellagio system
77

also implements an exact combinatorial exchange for computing resources [AuYoung et al.,
10]. Its pricing is based on the approximation of the truthful Vickrey-Clarke-Groves prices
proposed by Parkes et al. [Parkes et al., 11]. Being an exact mechanism it shares the
computational drawbacks of MACE. The work of Bapna et al. [Bapna et al., 12] is most
relevant to the work presented in this paper. Multiple requesters and providers can trade
both computing power and memory for a sequence of time slots. First, an exact mechanism is
introduced. By introducing fairness constraints and imposing one common valuation across
all resource providers, the search space in the underlying combinatorial allocation problem
is structured so as to establish one common, truthful price per time slot for all accepted
requests. Additionally, this introduces a linear ordering across all jobs and time which reduces
the complexity of the allocation problem, which however still remains NP-hard. To mitigate
this computational complexity, a fast, greedy heuristic is proposed at the expense of both
truthfulness and efficiency.
However, as shown in Lehmann et al. [Lehmann et al., 13] and Mu’alem and Nisan
[Mu’alem and Nisan, 14], heuristics do not necessarily imply a loss of both truthfulness and
efficiency. In their work, which is essential to this paper, Mu’alem and Nisan establish a
set of necessary and sufficient properties which can be leveraged to indeed design a truthful
heuristic.
2 A Truthful Clearing Heuristic
In this section we elaborate GreedEx, a market-based heuristic which obtains fast and
near-optimal resource allocations while generating prices that are truthful on the demandside
and approximately truthful on the supply-side of the utility computing market. To this
end, we will first introduce the bidding language which defines the content of the messages
being exchanged between the market mechanism and its participants and then we present
the formal model of the underlying allocation problem. Secondly, we will elaborate an allocation
algorithm which achieves a fast approximation of the optimal allocation schedule.
Thirdly, pricing schemes will be introduced which translate these resource allocations into
corresponding monetary transfers.
2.1 The Model
There are two classes of participants in utility computing markets: requesters who would like
to access computer resources such as processing power and memory, and utility computing
providers who offer these resources in return for the market price. We introduce the term
“job” to refer to a computational problem and we will call an atomic bundle of resources
on which the job or parts of it can be computed a “node”, e.g. one server within a cluster.
If a job gets allocated to a node, we will call this job and node a “winning” job and node
respectively. The market mechanism is a centralized entity which clears periodically, meaning
it collects resource requests and offers for a period of time before making its allocation and
pricing decisions. Larger clearing periods generally allow for more efficient allocation decisions
due to additional degrees of freedom when optimizing over a large number of requests and
offers [Parkes et al., 11]. Note, however, that the clearing period can be made as small as
78

a couple of seconds in order to support interactive applications [Stoesser et al., 15]. The
market mechanism is a sealed-bid mechanism in the sense that both resource requesters and
providers do not get to know the other users’ requests and offers before market clearing
[Parkes et al., 11].
A resource requester j who would like to submit a job to the utility computing system
reports the job’s characteristics (v j , c j , m j , s j , e j ) to the market mechanism where v j ∈ R + 0
denotes j ’s maximum willingness to pay (i.e. j ’s valuation) per unit of computing power
and time slot, c j ∈ N and m j ∈ N the minimum required amount of computing power and
memory respectively, and s j ∈ N and e j ∈ N specify the job’s estimated runtime. In the
following we will use the terms “resource requester” and “job” interchangeably. There is ongoing
research about using historical data and information about jobs’ size and application
in order to predict resource requirements and runtimes [Degermark et al., 16; Smith et al.,
17]. Based on this research we assume users to be able to report estimates about these
characteristics to the system. We require the market mechanism to make atomic allocations
in the sense that each job can only be executed if there are sufficient resources available in all
requested time slots. Furthermore, jobs can potentially be migrated between several nodes
over time but each job can only be executed on one node at a time.
A utility computing provider n who would like to contribute a node to the utility computing
system reports the node’s characteristics (r n , c n , m n , ɛ n , λ n ) to the market mechanism
where r n ∈ R + 0 specifies this node’s reservation price per unit of computing power and time
slot, c j ∈ N and m j ∈ N the maximum amount of computing power and memory available
on this node, and ɛ n ∈ N and λ n ∈ N the time frame during which the node can be accessed.
Given sufficient resources, we assume that each node is able to virtually execute multiple
jobs in parallel, for instance by using virtualization middleware.
Example: Assume the resource requests and offers listed in Table 1 have been submitted
to the system. Job J1 requests to be run in time slots 1 to 7 and requires a minimum of 54
Table 1: Sample resource requests and offers
Job j v j c j m j s j e j Node n r n c n m n ɛ n λ n
J1 12 54 126 1 7 N1 4 84 71 2 10
J2 4 85 32 3 7 N2 7 100 101 1 9
J3 12 43 43 1 8
J4 16 35 43 2 7
J5 11 47 37 3 7
J6 6 31 19 2 7
units of computing power and 126 units of memory in each time slot. J1 is willing to pay
up to $12 per unit of computing power and time slot, that is $12 * 54 * 7 = $4,536 in total.
Node N1 offers 84 units of computing power and 71 units of memory in time slots 2 to 10
and requires a reservation price of $4 per unit of computing power and time slot.
In real settings and with respect to designing usable systems, human users cannot be
expected to frequently submit requests and offers manually. Instead, software agents may
serve so as to hide the system’s complexity by automatically trading resources based on
79

the current resource consumption of applications and configurable bidding rules which automatically
derive corresponding valuations [MacKie-Mason and Wellman, 18; Neumann et al.
19].
2.2 Greedy Allocation Scheme
Let J be the set of resource requests, N the set of resource offers, and T = {t ∈ N | s j ≤ t ≤
e j , j ∈ J} ∪ {t ∈ N | ɛ n ≤ t ≤ λ n , n ∈ N} the set of time slots across all requests and offers,
i.e. the allocation problem’s time horizon. Then the winner determination problem which
solves the allocation problem exactly can be formalized as the following integer program:
max V = ∑ ∑ ∑
c j x jnt (v j − r n )
X j∈J t∈T n∈N
s.t. x
∑ jnt ∈ {0, 1}, s j ≤ t ≤ e j , ɛ n ≤ t ≤ λ n , v j ≥ r n , j ∈ J, n ∈ N (C1)
x jnt ≤ 1, s j ≤ t ≤ e j , j ∈ J
(C2)
n∈N ∑
x jnt c j ≤ c n , ɛ n ≤ t ≤ λ n , n ∈ N
(C3)
j∈J
∑
x jnt m j ≤ m n , ɛ n ≤ t ≤ λ n , n ∈ N
(C4)
j∈J
e j
∑
∑
u=s j n∈N
x jnu = (e j − s j + 1) ∑
n∈N
x jnt , s j ≤ t ≤ e j , j ∈ J
The objective of this integer program is to maximize welfare V, the total difference between
the requesters’ valuations and the providers’ reservation prices across all time slots. Constraint
(C1) introduces the binary decision variable x and ensures that a job can only be
allocated to a node which is accessible during the right time slots and whose reservation
price does not exceed the job’s willingness to pay. Furthermore, a job can only be allocated
to at most one node at a time (C2). Constraints (C3) and (C4) specify that the jobs allocated
to one node at a time are not allowed to consume more resources than are available on
this node. Constraint (C5) enforces atomicity, i.e. a job is either fully executed or it is not
executed at all.
Example: Figure 1 shows the schedule which results from optimally allocating the sample
requests and offers above.
The allocation problem at hand is an instance of a multi-dimensional Generalized Assignment
Problem [Martello and Toth, 20] and as such clearly NP-hard. Heuristics have the
desirable property of generating suboptimal allocations fast. We define the following greedy
heuristic:
(C5)
1. Sort jobs j ∈ J in non-ascending order of their reported willingness to pay v j . Sort
nodes n ∈ N in non-descending order of their reported reservation prices r n .
2. Starting with job j with the highest ranking (i.e. the highest reported willingness
to pay), allocate j to the nodes n 1 , . . . , n k with the highest ranking (i.e. the lowest
reported reservation prices) which can together accommodate j.
3. Repeat the allocation procedure of Step 2 with the next job in the ranking until there
are no more jobs which can be allocated to the available nodes.
80

Figure 1: Sample allocation schedule
This heuristic truly implements a greedy allocation scheme: it tries to greedily maximize
the term v j − r n in the objective function of the exact allocation problem above.
Example: For the sample requests and offers at hand, the greedy heuristic is illustrated
in Figure 2.
Figure 2: The greedy heuristic applied to the sample requests and offers
The winning jobs and nodes are highlighted. Job J4 can be allocated to node N1 since
N1 offers sufficient resources over all required time slots and its reported reservation price is
less than J1’s reported willingness to pay. In time slot 1, J3 can be allocated to node N1.
However, in time slots 2 to 8, there is not sufficient residual capacity left due to the execution
of J4. So J3 is subsequently allocated to the next available node N2. J1 cannot be executed
at all due to its excessive memory requirements. The heuristic proceeds until the ranked
list of jobs ends and finally happens to generate the same allocation schedule as the exact
mechanism for the example at hand.
2.3 Pricing Schemes
The allocation algorithm of market mechanism intends to achieve some global or social aim, in
this case maximize social welfare. In achieving this goal, it depends on the resource requesters
and providers to report their true valuations and resource characteristics. These participants,
however, are assumed to be rational and self-interested agents trying to maximize their
81

individual benefit which may not per se be aligned with the social aim. Pricing schemes
are introduced and closely linked to the allocation algorithm so as to induce participants to
indeed report truthfully about their valuations and characteristics while maintaining other
economic criteria [Stoesser et al., 15; Schnizler et al., 9]. In this section, we will first introduce
GreedEx’s pricing scheme. The prices of resource requests are determined by critical-valuebased
pricing while the resulting revenue is distributed to the utility computing providers
by proportional pricing. Subsequently, we will briefly outline K-Pricing [Schnizler et al., 9]
as an alternative pricing scheme which we will use in section 4 to numerically analyze and
benchmark both GreedEx as well as its pricing scheme.
2.3.1 Critical-value-based Pricing
The rationale behind the pricing of resource requests is similar to the well-known Vickrey-
Principle: the payment of each winning job amounts to the lowest (total) willingness to pay
it would have needed to report in order to still remain in the allocation, keeping all other
resource requests and offers fixed. This is also called the critical value φ j of job j [Lehmann
et al., 13; Mu’alem and Nisan, 14]. This is illustrated best by looking at the example at
hand.
Example: The critical values and resulting payments for the winning jobs J3, J4, J5 and
J6 are given in Table 2.
Table 2: Critical-value-based prices of the sample resource requests
Job j J3 J4 J5 J6
v j 12 16 11 6
φ j 7 7 7 4
p cv,j $2,408 $1,470 $1,645 $744
Job J3 would have needed to report a willingness to pay of at least $7 – the reservation
price of node N2 – in order to still remain in the allocation schedule. Consequently, φ j = $7
and the resulting critical-value-based price of J3 is p cv,j = (e j − s j + 1)c j φ j = 8 ∗ 43 ∗ $7 =
$2, 408. The prices of J4, J5 and J6 are determined accordingly. Jobs J1 and J2 have not been
allocated and consequently no payment is required. Overall, the pricing scheme generates
revenue of $6,267.
Truthfulness with respect to a job’s resource requirements is straightforward. If a job’s
resource requirements are understated, the job will not be finished; it has to pay for the used
resources but these are of no value. Overstating a job’s requirements either increases the
job’s payment or the job is not scheduled at all. Regarding the reported willingness to pay,
the main feature of the proposed pricing scheme is its ability to generate truthful prices of
resource requests: It is a (weakly) dominant strategy for resource requesters to report their
true valuations [Stoesser et al., 15]. We assume quasli-linear utility functions. For utility
P
j∈J x jntc j
c n
computing provider n ∈ N, u n (X) = p n (X) − ∑ t∈T r n where X denotes a specific
allocation schedule and p n (X) is provider n’s resulting payment. For resource requester j ∈ J,
u j (X) = v j c j (e j − s j + 1) − p j (X) if j is among the winning jobs and u j (X) = 0 else, where
82

p j (X) denotes j ’s price under allocation schedule X. Then, with truthful critical-value-based
prices p cv,j , j ∈ J,
∀ ṽ j ∈ R + 0
: u j (X) = v j c j (e j − s j + 1) − p cv,j (X) ≥ ṽ j c j (e j − s j + 1) − p cv,j ( ˜X) = u j ( ˜X)
Truthfulness is a strong and desirable feature as it tremendously simplifies the resource
requesters’ strategy space. For the individual participant, there is no need to reason about
the other participants’ strategies. Instead, she can “simply” report her true valuation. Note
that we assume participants to know their true valuations. As suggested above, a policybased
approach may assist users in deriving adequate valuations from the current resource
consumption of applications [MacKie-Mason and Wellman, 18; Neumann et al. 19].
The critical value of a job essentially hinges on the competition of other jobs for the
same resources. If there is no competition, the job is only required to pay the reservation
price. Otherwise, the job at least needs to outbid these competing jobs. Consequently,
critical-value-based pricing implements a desirable dynamicity from the providers’ point of
view. Providers do not need to constantly monitor the demand-side of the market to set
appropriate prices. Instead, they can simply delegate this to GreedEx. If there is sufficient
competition such that critical values are above reservation prices, critical-value-based pricing
of resource requests potentially generates a surplus, i.e. overall payments exceeding the utility
computing providers’ reservation prices.
2.3.2 Proportional Payments to Utility Computing Providers
Like the resource requesters, these providers are assumed to be selfish agents trying to report
to the mechanism so as to maximize their individual benefit. Utility computing providers
clearly do neither have an incentive to understate nor to overstate their availability of resources;
understating may reduce the generated revenue while overstating can be monitored
and punished. Unfortunately, the critical-value-based pricing scheme is not applicable to
resource offers; it requires a binary decision in the sense that either all resources of an offer
are allocated or none at all. However, in our model, we allow the partial use of nodes. In
order to at least approximate truthful payments, we thus fall back to payments proportional
to each provider’s contribution of processing power to the allocation schedule. Let
∑
S(X) = ∑ p cv,j (X) − ∑ x
∑ jnt c j
j∈J
r n
c n
j∈J
n∈N t∈T
be this surplus for allocation schedule X. Then provider n will receive a proportional payment
amounting to
∑ ∑ ∑
p prop,n (X) = ∑ c j
x jnt c j
j∈J
t∈T j∈J
r n + S · ∑ ∑ ∑
c n x jmt c
t∈T
j
t∈T j∈J m∈N
i.e. n will receive a share of the surplus according to its share in the total contribution of
processing power across all providers.
Example: In the example at hand, nodes N1 and N2 have reservation prices of $1,756
83

and $3,752 respectively for the allocation schedule. Thus there is a surplus of $6,267 - $1,756
- $3,752 = $759. In total, 975 units of processing power are provided, 439 by N1 and 536 by
N2. Consequently, N1 receives a total payment of $1,756 + 0.45 * $759 = $2,097.55 and N2
receives $3,752 + 0.55 * $759 = $4,169.45.
As stated above, there are further economic design criteria besides truthfulness. The two
most prominent and desirable are budget-balance and individual rationality.
A mechanism is said to be budget-balanced if its pricing scheme does not need to be
subsidized by outside payments, i.e. the payments from
∑
the resource providers cover the
payments made to utility computing providers: ∀X : p cv,j (X) − ∑ p prop,n (X) ≥ 0. The
pricing scheme of GreedEx is specifically designed so as to obtain strongly budget-balanced
payments. Any surplus from the demand-side is distributed to the supply-side of the market.
After settling the market, there are neither payments left, nor does the market need to be
subsidized with outside payments.
Individual rationality is satisfied if no participant can suffer a loss from participating in
the mechanism: ∀X : u j (X) ≥ 0, j ∈ J, and u n (X) ≥ 0, n ∈ N. GreedEx (respectively
participation) is individually rational. No resource requester needs to pay anything in case
her request is not accepted, nor does she have to pay more than her reported willingness to
pay in case her request is served. Each utility computing provider is guaranteed to receive
at least her reported reservation price for her contributed resources.
Budget-balance and individual rationality are hard feasibility constraints and must be
satisfied in order for the market to be sustainable.
2.3.3 K-Pricing
K-Pricing is introduced in [Schnizler et al., 9]. The basic idea is to distribute the welfare
generated by the allocation algorithm between resource requesters and utility computing
providers according to a factor k ∈ [0, 1]. For instance, assume an allocation of resources
from a specific provider to a specific requester. The buyer values these resources at $10 while
the provider has a reservation price of $5. Then the (local) welfare is $10 - $5 = $5 and k *
$5 of the surplus is allotted to the requester – thus having to pay $10 - k * $5 – and (1 - k)
* $5 is allotted to the provider – thus receiving $5 + (1 - k) * $5.
Besides allowing fairness considerations, the main advantage of K-Pricing is that it can be
determined in polynomial runtime. On the downside, however, it only yields approximately
truthful prices and payments on both sides of the market. In contrast to critical-value-based
pricing, K-Pricing can be applied to both GreedEx and the exact mechanism and will thus
be used for benchmark purposes in the following numerical experiment.
j∈J
n∈N
3 Numerical Experiment
Heuristics inherently imply a trade-off between computational speed and allocative efficiency.
A theoretical analysis of the heuristic only allows examining its general behavior. Thus,
computational speed and efficiency can only be evaluated properly by means of numerical
simulations. In this section, we will shortly introduce the simulation setting which we chose
84

to benchmark GreedEx against the exact mechanism before we will subsequently present
and discuss the results of this simulation.
3.1 Data Generation
With numerical analysis, there are two general possibilities to come up with data sets upon
which the heuristic can be tested [Böttcher et al., 21; Feitelson, 22]. Practical cases, e.g.
derived from real workload traces of distributed systems, are of high practical relevance.
However, it is hard to obtain complete traces about all job and node characteristics and to
generalize from one workload trace to others. More importantly, to the best of our knowledge,
there is currently no trace available containing the reported valuations of users. We
consequently chose the second option, which is the creation of artificial instances. Artificial
data sets yield the benefit that they can be constructed so as to specifically investigate certain
properties of the heuristic.
With our heuristic, reflecting the bidding language introduced in Subsection 2.1, there is
a set of parameters upon which data sets need to be generated:
Parameter
Table 3: Simulation settings
Number of orders per side 10, 20, . . . , 170
Domain
Processing power and memory requested Lognormal LogN(4, 0.28)
Processing power and memory offered Lognormal LogN(4.5, 0.16)
Start and end time of requests Normal N(2, 0.5) and N(6, 0.5)
Start and end time of offers Normal N(2, 0.5) and N(8, 0.5)
Willingness to pay Uniform [0, 20]
Reservation price Uniform [0, 10]
We call a specific combination of such parameter values a “setting”. We chose to fix the
number of requests and offers within one setting but to increase this number across settings
from 10 orders per side to 20, 30 and so forth up to 170 requests and 170 offers. As will be
seen in our results, we found that the exact mechanism cannot be run within acceptable time
with more than 170 orders per side upwards. Consequently, we defined 17 settings in total.
According to Feitelson [Feitelson, 22], the lognormal distribution generates a good fit to
characteristics in real workloads, such as file size. It represents a normal distribution in log
space, is modal and positively skewed. The distributions specified in Table 3 have been kept
fixed across all 17 settings. The same applies to time constraints and valuations, for which
we chose normal and uniform distributions respectively.
For each setting, we generated 30 order books with the corresponding number of requests
and offers, meaning we ran each setting 30 times in order to mitigate stochastic outliers. We
then fed both the heuristic as well as the exact mechanism with identical order books. The
simulations were run using jCase on an Intel Pentium Xeon computer with 3.2 GHz and 2
GB memory. jCase is a Java-based simulation environment for multi-attribute combinatorial
85

environments. The exact formulations of the allocation problem were solved using CPLEX
9.1, a state-of-the art commercial solver for mathematical optimization problems. The results
presented below represent the averages across all 30 runs of each setting.
In order to be able to run the simulation within acceptable time, we decided to stop the
CPLEX solver after 300 seconds (dark red line) and after 1800 seconds (blue line). CPLEX
then returned the best solution found so far. Since the results for the large settings do not
always match with the optimum for large settings, we call these runs “Anytime[300s]” and
“Anytime[1800s]” in the spirit of Sandholm et al. [Sandholm et al., 23] and Schnizler et al.
[Schnizler et al., 9]. We chose K-Pricing as pricing scheme for the exact mechanism. Recall
that critical-value-based pricing cannot be applied to the exact mechanism while the heuristic
can be combined with both pricing schemes.
3.2 Data Analysis
In this subsection we will present the results of the experiment with respect to computational
tractability and allocative efficiency.
3.2.1 Computational Tractability
GreedEx is designed so as to perform fast allocations. It runs in polynomial time of the
number of resource requests J and offers N : In the first phase, requests and nodes are sorted
in O(|J|log|J|) and O(|N|log|N|) respectively. In the second phase, jobs are allocated to
nodes in O(|J||N|). While the allocation algorithm can thus be run very fast, critical-valuebased
pricing adds computational complexity to GreedEx. The allocation schedule is not
independent of a specific job j in the sense that the allocation of other jobs might change if
j is not being considered. Thus, for every winning job j, in order to determine j ’s critical
value, the allocation without j needs to be determined. This is similar to the drawback of
the Vickrey-Clarke-Groves mechanism [Parkes et al., 11; Schnizler et al., 9]: The allocation
needs to be computed n+1 times if n is the number of winning jobs. However, in contrast to
the Vickrey-Clarke-Groves mechanism, this allocation can now be computed in polynomial
time.
The average runtime µ CT and the coefficient of variation CV CT = σ CT /µ CT (where σ CT is
the standard deviation) for each setting and combination of allocation algorithm and pricing
scheme is listed in Table 5 in the Appendix.
In Figure 3, a linear scale was used. Note that the runtime of Anytime[300s] and Anytime[1800s]
grows quickly in the number of orders per side. Especially the runtime of Anytime[1800s]
shows an exponential growth and takes almost 22.3 minutes on average already
with 170 orders per side. The curves of the anytime versions begin to flatten for large order
books due to the predefined time limit, i.e. Anytime[1800s] and Anytime[300s] will converge
asymptotically to an (average) runtime of 1800s and 300s respectively. In contrast,
GreedEx clears very fast. It takes less than 800 milliseconds for 170 orders per side.
Figure 4 shows the same data in milliseconds on a logarithmic scale in order to illustrate
the runtime behavior of GreedEx in more detail. Moreover, this graph shows the computational
cost of the critical-value-based pricing scheme compared to K-Pricing. Even though
the allocation needs to be determined multiple times in order to compute the critical values of
86

Figure 3: Runtime behavior in seconds on linear scale
the winning jobs, GreedEx with critical-value-based pricing still only takes 763 milliseconds
on average to clear order books with 170 orders per side and thus exhibits a good-natured
linear growth in runtime. The speed of GreedEx allocation algorithm is showcased by
GreedEx with K-Pricing. While Anytime[1800s] with K-Pricing takes about 22.3 minutes
on average to clear order books with 170 orders per side, GreedEx with K-Pricing only
takes 39 milliseconds.
3.2.2 Allocative Efficiency
The typical approach when analyzing the allocative efficiency of a heuristic is to benchmark
it against the optimal welfare generated by an exact mechanism. Heuristics generally only
produce suboptimal allocations. Furthermore, for most heuristics the ratio of the generated
suboptimal allocation and the optimal allocation can be made arbitrarily small for special
cases. This also holds for GreedEx. Such special cases, however, are mostly only of theoretical
value since they do not reflect realistic workloads. We would consequently like to
compare the efficiency achieved by the exact mechanism with the efficiency generated by
GreedEx by means of numerical experiments. Unfortunately, as the analysis above illustrates,
the runtime of the exact mechanism is prohibitively high. Anytime[1800s], however, is
an accurate proxy for the optimal solution when benchmarking GreedEx. Table 4 indicates
how many times the CPLEX solver returns a suboptimal solution (i.e. it hits its predefined
time limit) dependant on the order book size:
The results of Anytime[1800s] with order books with less than 130 orders per side cor-
87

Figure 4: Runtime behavior in milliseconds on logarithmic scale
Table 4: Suboptimal solutions
Number of orders per
side
Number of runs with
suboptimal solutions
10 . . . 120 130 140 150 160 170
0 . . . 0 1 1 7 7 14
88

espond to the optimum. For larger order books, Anytime[1800s] always finds a feasible
solution. Even though it returns an increasing number of suboptimal solutions, it still finds
16 optimal solutions for the largest order book size while the suboptimal solutions will be
fairly close to the optimum.
Comparing the welfare generated by the anytime versions of the exact mechanism with the
welfare generated by GreedEx yields the results shown in Figure 5. The average welfare µ W V
and the coefficient of variation CV W V = σ W V /µ W V (where σ W V is the standard deviation)
for each setting and combination of allocation algorithm is listed in Table 6 in the Appendix.
GreedEx achieves a surprisingly high efficiency ratio. While it shows some minor spikes
with small order books, efficiency successively levels off at around 95%. We conjecture that
the spikes are settled due to the additional degrees of freedom the heuristic obtains with
larger order books when making its allocation decisions. The efficiency of Anytime[300s]
drastically deteriorates with order books of 100-110 orders per side upwards. The explanation
is straightforward looking at Figure 4: With these order book sizes, Anytime[300s] begins to
hit its predefined time limit. Consequently, the returned solutions are increasingly distant
to the optimum or even worse, Anytime[300s] does not even find feasible allocations within
5 minutes.
Figure 5: Approximation of efficiency compared to the anytime versions of the exact mechanism
3.3 Implications for Utility Computing
While the results of this experiment must not be generalized overhastily, they are promising
and clearly strengthen the case for heuristics such as GreedEx and related approaches.
In real utility computing settings, the markets will be much larger than the sizes used
in this experiment. Thus the results illustrate that exact mechanisms and even anytime
89

approximations based on exact mechanisms are not appropriate for real utility computing
markets but can merely be used as a baseline and reference point. The mere computational
infeasibility of exact mechanisms even becomes exacerbated in interactive settings where
immediacy of the allocation is the key requirement [Stoesser et al., 15].
Designing market mechanisms in this domain is a complex task. The trade-off between
computational speed and allocative efficiency, which is already challenging in itself, is aggravated
by the introduction of economic design criteria such as truthfulness, budget-balance,
and individual rationality, and the interdependencies between these criteria.
The surprisingly high allocative efficiency of GreedEx in this experiment illustrates that
we are on the right track in this regard. A promising trade-off was achieved while sacrificing
only a minor fraction of efficiency. Market mechanisms may thus be a feasible means to allocate
utility computing resources in an efficient manner while allowing fully dynamic pricing
and the aggregation of demand and supply across multiple resource requesters and utility
computing providers.
4 Conclusion and Outlook
In the introduction to this paper, we motivated the use of utility computing as a promising
concept to increase the flexibility of IT systems while at the same time cutting down on IT
expenses. Utility computing markets may serve so as to aggregate demand and supply of
computer resources and to ease the discovery of appropriate utility computing providers. In
inter-organizational and dynamic settings with a large number of both resource requesters
and providers, however, the clearing of such markets becomes computationally demanding.
The current mechanisms presented in Section 1 are not fully able to cope with this setting. To
this end, in Section 2 we elaborated GreedEx consisting of a greedy allocation algorithm
and a pricing scheme. This market mechanism can be used to perform this clearing in a
scalable manner while at the same time satisfying basic economic design criteria. Besides its
computational speed, the main characteristic of GreedEx is its ability to account for the
inter-organizational nature of utility computing by generating truthful prices for resource
requests and approximately truthful payments to utility computing providers. In general,
however, the computational speed of heuristic comes at the expense of efficiency. Consequently,
in Section 3 we reported the results of a numerical simulation which illustrated
GreedEx’s speed-up as well as its surprisingly high efficiency.
These results strengthen our approach and point to several interesting avenues for future
research. These can be distinguished into two categories: Firstly, we will investigate whether
we can further improve GreedEx’s approximation of efficiency by means of alternative
norms in the sorting phase, look-ahead extensions, and/or randomization. However, this
is a tricky thing to do as the impact on the economic criteria needs to be considered as
well. Secondly, truthful payments to utility computing providers can only be approximated.
We intend to perform an in-depth analysis of alternative pricing schemes and their strategic
impacts.
90

References
[1] Rappa, M.A. (2004). “The utility business model and the future of computing services,”
IBM Systems Journal 43(1), 32–42.
[2] Carr, N.G. (2003). “IT Doesn’t Matter,” Harvard Business Review 811(5), 41–53.
[3] Amazon.com (2007). http://www.amazon.com, visited March 2, 2007.
[4] Sun Microsystems (2007). http://www.sun.com/service/sungrid/overview.jsp, visited
March 2, 2007.
[5] Waldspurger, C., T. Hogg, B.A. Huberman, J.O. Kephart, and W.S. Stornetta (1992).
“Spawn: A Distributed Computational Economy,” IEEE Transactions on Software
Engineering 18(2), 103–117.
[6] Chun, B., and D.E. Culler (1999). “Market-based proportional resource sharing for
clusters,” Millenium Project Research Report.
[7] Regev, O., and N. Nisan (1998). “The POPCORN Market – An Online Market for Computational
Resources,” Proceedings of the 1 st International Conference on Information
and Computation Economies, New York, USA, 148–157.
[8] Foster, I. (2005). “Globus Toolkit Version 4: Software for Service-Oriented Systems,”
Proceedings of the IFIP International Conference on Network and Parallel Computing,
Beijing, China, 2–13.
[9] Schnizler, B., D. Neumann, D. Veit, and C. Weinhardt (2006). “Trading Grid Services –
A Multi-attribute Combinatorial Approach,” European Journal of Operational Research,
forthcoming.
[10] AuYoung, A., B.N. Chun, A.C. Snoeren, and A. Vahdat (2004). “Resource Allocation
in Federated Distributed Computing Infrastructures,” Proceedings of the 1 st Workshop
on Operating System and Architectural Support for the On-demand IT Infrastructure,
Boston, MA, USA.
[11] Parkes, D.C., J. Kalagnanam, and M. Eso (2001). “Achieving Budget-Balance with
Vickrey-Based Payment Schemes in Combinatorial Exchanges,” IBM Research Report
RC 22218 W0110-065.
[12] Bapna, R., S. Das, R. Garfinkel, and J. Stallaert (2006). “A Market Design for Grid
Computing,” INFORMS Journal of Computing, forthcoming.
[13] Lehmann, D., L.I. O’Callaghan, and Y. Shoham (2002). “Truth Revelation in Approximately
Efficient Combinatorial Auctions,” Journal of the ACM 49(5), 577–602.
[14] Mu’alem, A., and N. Nisan (2002). “Truthful Approximation Mechanisms for Restricted
Combinatorial Auctions,” Proceedings of the 18 th National Conference on Artificial
Intelligence (AAAI-02), Edmonton, Alberta, Canada.
91

[15] Stoesser, J., D. Neumann, and A. Anandasivam (2007). “A Truthful Heuristic for Efficient
Scheduling in Network-Centric Grid OS,” Proceedings of the 15 th European Conference
on Information Systems (ECIS), St. Gallen, Switzerland.
[16] Degermark, M., T. Kohler, S. Pink, and O. Schelen (1997). “Advance Reservations
for Predictive Service in the Internet,” ACM/Springer Verlag Journal on Multimedia
Systems 5(3), 177–186.
[17] Smith, W., I. Foster, and V. Taylor (1998). “Predicting Application Run Times Using
Historical Information,” Proceedings of the IPPS/SPDP 1998 Workshop on Job
Scheduling Strategies for Parallel Processing, Orlando, FL, USA.
[18] MacKie-Mason, J.K., and M.P. Wellman (2006). Automated Markets and Trading
Agents,” In: Tesfatsion, L., and K.L. Judd (eds.), Handbook of Computational Economics,
vol. 2: Agent-Based Computational Economics, North-Holland.
[19] Neumann, D., S. Lamparter, and B. Schnizler (2006). “Automated Bidding for Trading
Grid Services,” Proceedings of the 14 th European Conference on Information Systems
(ECIS), Gothenburg, Sweden.
[20] Martello, S., and P. Toth (1990). “Knapsack problems,” John Wiley, Chichester, England.
[21] Böttcher, J., A. Drexl, R. Kolisch, and F. Salewski (1999). “Project Scheduling under
Partially Renewable Resource Constraints,” Management Science 45(4), 543–559.
[22] Feitelson, D.G. (2002). “Workload Modeling for Performance Evaluation,” Performance
Evaluation of Complex Systems: Techniques and Tools, Lecture Notes in Computer
Science 2459, Springer Verlag, 114–141.
[23] Sandholm, T., S. Suri, A. Gilpin, and D. Levine (2001). “CABOB: A Fast Optimal
Algorithm for Combinatorial Auctions,” Proceedings of the 17 th International Joint
Conference on Artificial Intelligence, Seattle, WA, USA.
92

Appendix – Numerical Results
Table 5: Numerical results w.r.t. computational tractability (in milliseconds)
Orders per side
Anytime[1800s]
Anytime[300s]
GreedEx – GreedEx –
Critical value K-Pricing
µ CT CV CT µ CT CV CT µ CT CV CT µ CT CV CT
10 42.40 1.02 52.77 0.88 5.77 1.49 4.23 2.34
20 193.30 0.54 211.47 0.52 5.20 1.95 1.07 3.74
30 960.43 1.09 1,254.73 1.39 7.23 1.07 3.60 1.81
40 2,070.30 0.68 2,790.67 1.09 14.10 0.33 5.73 1.80
50 6,000.70 0.90 6,498.37 1.11 27.20 0.42 9.37 1.11
60 11,431.73 0.69 14,047.73 0.64 44.23 0.26 14.53 1.00
70 28,573.00 1.19 36,452.57 1.45 59.33 0.20 13.50 1.22
80 37,959.37 0.58 44,692.70 0.73 85.90 0.22 12.20 0.80
90 57,150.47 0.49 56,592.63 0.68 131.87 0.20 12.47 0.50
100 106,325.63 0.57 111,070.87 0.59 170.97 0.22 22.40 0.84
110 141,735.97 1.02 112,234.80 0.49 216.17 0.14 21.30 0.69
120 199,872.83 0.66 171,713.60 0.44 280.77 0.15 31.27 0.74
130 503,131.70 0.95 229,594.20 0.34 340.70 0.14 25.60 0.37
140 502,005.73 0.83 267,041.10 0.21 452.57 0.11 42.17 0.62
150 991,292.77 0.58 292,886.40 0.10 538.60 0.14 33.77 0.36
160 1,200,359.93 0.44 300,532.90 0.03 635.93 0.12 33.33 0.16
170 1,334,458.33 0.40 302,720.20 0.00 762.90 0.14 39.07 0.34
93

Table 6: Numerical results w.r.t. allocative efficiency
Orders per side
Anytime[1800s] Anytime[300s] GreedEx
µ W V CV W V µ W V CV W V µ W V CV W V
10 23,372.53 0.31 23,377.97 0.31 22,403.13 0.32
20 46,669.43 0.23 46,637.00 0.23 43,209.23 0.23
30 76,111.47 0.20 76,116.83 0.20 71,035.03 0.20
40 102,620.77 0.14 102,572.77 0.15 94,277.03 0.16
50 134,355.93 0.13 134,494.70 0.14 124,980.30 0.14
60 161,050.57 0.09 161,076.40 0.09 151,352.00 0.10
70 190,288.40 0.12 190,329.83 0.12 178,100.40 0.12
80 210,268.83 0.13 210,307.70 0.13 198,849.70 0.13
90 245,961.03 0.09 245,829.93 0.09 229,105.50 0.09
100 268,946.90 0.09 268,940.60 0.09 251,325.97 0.09
110 302,435.27 0.07 302,229.43 0.07 284,234.43 0.07
120 332,844.77 0.10 331,924.17 0.10 314,384.40 0.10
130 354,773.97 0.09 350,776.97 0.09 333,256.77 0.10
140 394,100.43 0.09 389,306.87 0.09 371,391.67 0.10
150 415,811.30 0.08 391,304.48 0.14 395,062.93 0.07
160 443,498.97 0.09 376,813.14 0.23 420,429.00 0.09
170 464,621.13 0.06 390,701.57 0.21 443,740.97 0.06
Author Biographies
Jochen Stoesser finished his studies in Information Engineering and Management at Universität
Karlsruhe (TH) with a degree as Diplom-Informationswirt in 2006. During his studies
he spent two terms at Carleton University in Ottawa, Canada, and one term at The University
of New South Wales in Sydney, Australia. He joined the Institute of Information
Systems and Management at Universität Karlsruhe (TH) in September 2006 as Ph.D. student
and research assistant. Jochen’s main interest is in the design of market mechanisms
for an efficient allocation of computer resources.
Dirk Neumann received a degree as Diplom-Ökonom (economics) from Justus-Liebig University
in Giessen and a Master of Arts degree in Economics from University of Wisconsin,
Milwaukee. He received his doctoral degree from Universität Karlsruhe (TH) in 2004. Since
then, he has been working as an assistant professor at the Institute of Information Systems
and Management at Universität Karlsruhe (TH). Dirk’s research interests basically cover two
areas. The first topic is concerned with market engineering. In essence, market engineering
denotes the theoretically founded design of electronic markets. The second topic is concerned
with the management of self-organizing networks. In self-organizing networks it is extremely
difficult to influence the entire network or cliques, as those networks adapt quickly to the
new situation in unforeseen directions.
94

Market-Based Pricing in Grids: On Strategic Manipulation and
Computational Cost
Jochen Stößer, Dirk Neumann, Christof Weinhardt
Institute of Information Systems and Management (IISM)
Universität Karlsruhe (TH)
Englerstr. 14, 76131 Karlsruhe, Germany
lastname@iism.uni-karlsruhe.de
Abstract
In Grid computing, computer resources are shared across administrative boundaries. In this
dynamic and inter-organizational setting, scheduling becomes a key challenge. Market
mechanisms are deemed promising to revolutionize resource allocation in Grids. Current market
mechanisms, however, are not adequately able to account for dependencies between multiple
Grid resources in large-scale settings with strategic users who try to benefit from manipulating
the mechanism. In this paper, we design a market-based heuristic for clearing large-scale Grid
settings. At the core of this paper, we present alternative pricing schemes which may
complement this heuristic so as to induce resource requesters and providers to report their true
valuations and resource characteristics to the system. We provide an in-depth analysis of these
pricing schemes and outline basic design options of how the mechanism can be integrated into
current Grid schedulers.
Key words: Distributed Systems, Market-Based Scheduling, Heuristic, Pricing
Introduction
IT departments in business and academia are facing a challenging dilemma. On the
one hand, they need to support increasingly complex applications which require massive
amounts of computer resources, such as ERP systems, engineering applications, or
demand and weather forecasts. On the other hand, IT departments are under constant
pressure to cut down on hard- and software expenses. In order to be able to
95

accommodate peak loads, they must maintain large computing clusters which lie idle
most of the time, thus incurring tremendous costs. Results of a meta-study show that
corporate data centers are only using 10% to 35% of their available computing power
and that 50% to 60% of companies’ data storage capacity is being wasted [Carr, 2005].
Grid computing is a promising way out of this dilemma by allowing organizations to
respond quickly to peaks in demand for computer resources. It denotes a computing
paradigm where computer resources are shared across administrative boundaries, e.g.
between enterprises and/or scientific computing centers or among business units and
departments within one enterprise [Foster et al., 2001]. A recent study conducted by The
Insight Research Corporation projects a growth in worldwide Grid spending from US$
1.84 Billion in 2006 to US$ 24.52 Billion in 2011 [Anonymous, 2006]. Current scientific
Grids such as D-Grid (www.d-grid.org) and EGEE (www.eu-egee.org) or enterprise
Grids employed at GlaxoSmithKline [Anonymous, 2007a] and Sal. Oppenheim
[Anonymous, 2007b], however, suffer from two major drawbacks: Firstly, resources are
shared free of charge and users with idle resources do not have an incentive to
contribute these resources to the Grid. Within enterprises, for example, business units
are oftentimes unwilling to share idle resources with other units as they do not receive
monetary compensation or future access to Grid resources in return. Secondly, Grids
are highly dynamic in nature. Over time, participating organizations may switch from
being a resource requester to being a resource provider and vice versa, and new entities
may enter the system or existing entities may leave.
In this dynamic and inter-organizational setting, scheduling, i.e. the decision about
which user can access what resource at what time and at what cost, becomes a key
challenge. Current scheduling mechanisms are based on system-centric measures and
96

pre-defined policies and user priorities. The scheduler of Sun Microsystems’s N1 Grid
Engine, for example, uses a combination of a central queue and a technical scheduler,
as illustrated in Figure 1 [Anonymous, 2005].
Figure 1. Scheduler in Sun Microsystems’s N1 Grid Engine
Requests are placed in the central queue according to a priority value which is
computed by using a complex combination of policies. A sample policy mix may
comprise manually pre-defined shares for specific user groups, departments or projects,
a contribution for requests which are close to their deadline or which have been waiting
in the system for a long time, and users can sort their own requests according to their
importance. Following this queue, a technical scheduler assigns the requests to Grid
resources. The objective of technical schedulers is to maximize resource utilization and
balance the system load [Buyya et al., 2005]. Users cannot specify their valuation for
Grid resources. In settings with excess demand for resources, the resulting allocations
will thus typically not coincide with economically efficient allocations that try to
accommodate those requests with the highest valuations. Furthermore, pre-defined
policies and priorities which are manually administered fail in dynamic and large-scale
Grid settings with users frequently entering and / or leaving the system. In summary,
current technical scheduling mechanisms are not fully able to handle the interorganizational
and dynamic nature of Grids.
97

To this end, market mechanisms are deemed promising to revolutionize resource
allocation in Grids for two reasons: They dynamically and efficiently match resource
requests to resource offers based on the users’ valuations. In addition, they establish
previously unknown prices where these prices reflect the scarcity of the respective
resources, thus offering users an incentive to contribute their idle resources to the Grid
in return for the market price. Current market mechanisms, however, are not adequately
able to account for dependencies between multiple Grid resources (e.g. computing
power and memory are needed simultaneously) in large-scale settings with strategic
users who try to benefit from manipulating the mechanism.
Our paper is novel as:
• We design a scalable market-based heuristic for clearing large-scale Grid
settings.
• We present alternative pricing schemes which may complement this heuristic
so as to induce resource requesters and providers to report their true
valuations and resource characteristics to the system.
• We provide an in-depth analysis of these pricing schemes with respect to the
distribution of welfare among resource requesters and providers, the incentives
of single users to try to manipulate the mechanism, and the computational
impact of the pricing schemes on the mechanism.
• We outline basic design options of how the mechanism can be integrated into
current Grid schedulers.
This paper is structured as follows. We present related work in the field of market
mechanisms in Grids. We subsequently describe a scalable heuristic for clearing Grid
markets, before we present and analyze existing and elaborate novel pricing and
98

payment schemes which distribute the revenue collected on the demand side of the
market to the provisioning side while accounting for both computational and strategic
considerations. These dimensions are evaluated in more detail by means of a numerical
experiment. We finally summarize the paper and highlight future research directions.
Related Work
A conceptual view on the integration of market mechanisms into Grid systems is
provided by the G-commerce project [Wolski et al., 2001] and GridBus [Buyya et al.,
2002]. These projects propose and evaluate basic auction mechanisms. But as Wellman
et al. [Wellman et al., 2001] point out, scheduling problems often exhibit properties which
are not captured by such basic auction protocols, e.g. the presence of multiple resource
providers and time constraints. Instead, a suite of mechanisms is likely to emerge for a
broader range of scheduling problems. Research in market-based allocation of Grid
resources can be separated into two streams with respect to the trading object:
mechanisms which consider the trading of one type of resource only and mechanisms
which account for dependencies between multiple Grid resources.
In Spawn, workstations auction off idle computing time by means of a Vickrey auction
[Waldspurger et al., 1992]. In market-based proportional share, resource requesters get
allotted computing time proportional to their share in the total reported valuation across
all requesters [Chun and Culler, 1999]. The Popcorn market is an online market for
computing power which implements a Vickrey auction as well as two double auctions
[Regev and Nisan, 1998]. All of these approaches allow the specification and trading of
computing power only. But requesters require a bundle of resources such as computing
power, memory, and bandwidth. On the one hand, the approaches at hand thus lead to
inefficient allocations since requests with the same demand for computing power but
99

different memory requirements, for instance, are treated the same. On the other hand,
requesters are exposed to the risk of only being able to obtain one portion of the bundle
of required resources without the other (“exposure risk”).
Schnizler et al. [Schnizler et al., forthcoming] elaborate a Multi-Attribute Combinatorial
Exchange (MACE) which targets these deficiencies. Users are allowed to request and
offer arbitrary bundles of computer resources and can specify quality attributes on these
resources. The pricing scheme of MACE yields approximately truthful prices on both
sides of the market. The scheduling problem in this combinatorial setting is NP-hard and
the mechanism is thus not applicable to large-scale settings and interactive applications
which require the immediate allocation of resources. The Bellagio system also
implements a combinatorial exchange for computing resources [AuYoung et al., 2004].
Being an exact mechanism it shares the computational drawbacks of MACE. The work
of Bapna et al. [Bapna et al., forthcoming] is most relevant to the work presented in this
paper. In their model, multiple requesters and providers can trade both computing power
and memory for a sequence of time slots. First, an exact mechanism is introduced. By
introducing fairness constraints and imposing one common valuation across all resource
providers, a linear ordering across all jobs and time is introduced which reduces the
complexity of the allocation problem, which however still remains NP-hard. To mitigate
this computational complexity, a fast, greedy heuristic is proposed at the expense of
both truthfulness and efficiency.
In summary, the mechanisms proposed for Grids are not fully able to account for
dependencies between multiple Grid resources in large-scale settings with strategic
users. In the following section, we propose a scalable heuristic which allows the trading
of both computing power and memory. We present pricing schemes which may
100

complement this heuristic so as to induce strategic users to truthful behavior, before we
briefly discuss how this mechanism can be employed as economic Grid scheduler.
The Exchange
We will introduce the bidding language first which determines how resource requests
and offers are expressed. In order to enable true usability of market-based approaches,
human users may be released from the burden of having to issue requests and offers
manually. Instead, software agents may serve so as to hide the Grid’s complexity from
human users by automatically trading resources based on the current resource
consumption of applications and configurable bidding rules which automatically derive
corresponding valuations [MacKie-Mason and Wellman, 2006, Neumann et al., 2006].
The market mechanism is a centralized entity and clears periodically, i.e. it collects
resource requests and offers for a period of time before making its allocation decision.
The mechanism’s order book is closed in that market participants do not obtain
information about other participants’ requests and offers.
Expressing Requests and Offers
In our model, computing power is the central scarce resource to be traded. Let N be
the set of resource offers which the mechanism has collected over a period of time.
When submitting a resource offer (“node”) n ∈ N, resource providers report the tuple (r n ,
c¯n, m¯ n, ε n , λ n ) to the system. r n ∈ R + denotes the provider’s reservation price per unit of
computing power and time in some common monetary unit (MU), c¯n ∈ ℵ and m¯ n ∈ ℵ
specify the maximum available amount of computing power (e.g. number of CPUs or
CPU cycles) and memory (e.g. in Megabytes) respectively, while ε n ∈ ℵ and λ n ∈ ℵ
indicate the earliest and latest time slot during which these resources are available. We
101

assume that nodes can virtually execute multiple requests in parallel, e.g. by viewing
each node as a virtual machine which makes this node almost perfectly divisible
[Barham et al., 2003].
Let J be the set of resource requests. Resource requesters wanting to submit a
computational task (“job”) j ∈ J report the tuple (b j , c j , m j , s j , e j ) to the system where b j ∈
R + expresses the requester’s maximum willingness to pay per unit of computing power
and time, c j ∈ ℵ and m j ∈ ℵ specify the minimum required amount of computing power
and memory respectively, and s j ∈ ℵ and e j ∈ ℵ mark the jobs start and end time. A job j
can only be executed in its entirety, meaning it is only allocated if there are sufficient
resources available in each time slot t ∈ [s j , e j ]. A job can only be run on one node at a
time but can migrate across nodes over time.
Table 1: Sample Requests and Offers
Job j b j c j m j s j e j Node n r n c¯n m¯ n ε n λ n
J1 10 35 50 3 7 N1 8 123 98 1 8
J2 14 35 27 1 7 N2 11 98 82 1 9
J3 11 84 45 2 7
J4 16 71 48 2 7
J5 12 63 30 2 7
J6 17 55 20 2 7
Example: The sample requests and offers listed in Table 1 have been submitted to
the system. Job J1 is requesting 35 units of computing power and 50 units of memory
throughout timeslots 3 to 7. It is willing to pay up to 10 MU per unit of computing power
and timeslot, i.e. b j c j (e j – s j +1) = 1,750 MU in total. Node N1 requires a payment of at
least 8 MU per unit of computing power and timeslot while it is offering up to 123 units of
computing power and 98 units of memory throughout timeslots 1 to 8.
102

Allocating Requests to Offers
Let T be the scheduling horizon for a problem instance, i.e. the set covering all time
slots specified in resource requests and offers. We introduce the binary decision variable
x with x jnt = 1 if job j is allocated to node n in time slot t and x jnt = 0 else. Then the winner
determination problem which allocates resource requests to resource offers so as to
maximize allocative efficiency V, that is welfare (or overall “happiness” of the market
participants), can be formalized as a mathematical integer program:
maxV
: =
X
s.
t.
x
jnt
∑
∈
{ 0,1 },
∑ x
n∈N
jnt
∑ x
j∈J
jnt
∑ x
j∈J
jnt
e j
∑ ∑
≤ 1, s
c
c
m
u= s j n∈N
j∈J
j T∈t
n∈N
j
s
j
∑ ∑
j
x
( b − r )
≤ t ≤ e , ε ≤ t ≤ λ , b
≤ t ≤ e , j ∈ J
≤ c n , ε ≤ t ≤ λ,
n ∈ N
x
≤ m n , ε ≤ t ≤ λ,
n ∈ N
jnu
j
=
n
n
j
j
n
jnt
j
n
n
j
≥ r , j ∈ J,
n ∈ N
n
( C1)
( C2)
( C3)
( C4)
( e − s + 1) x , s t e , j J ( C5)
j j ∑ ≤ ≤ ∈
n∈N
jnt j
j
Constraint (C1) introduces the binary decision variable x and ensures that a job can
only be allocated to a node which is accessible during the right time slots and whose
reservation price does not exceed the job’s willingness to pay. Furthermore, a job can
only be allocated to at most one node at a time (C2). Constraints (C3) and (C4) specify
that the jobs allocated to one node are not allowed to consume more resources than are
available on this node. Constraint (C5) enforces atomicity, i.e. a job is either fully
executed or it is not executed at all.
This winner determination problem constitutes a special case of the NP-complete
Generalized Assignment Problem (GAP) [Chekuri and Khanna, 2000]. Consequently,
solving this combinatorial winner determination problem to optimality makes scheduling
in this model computationally intractable in practice. Numerical experiments indicate that
clearing order books with 170 resource requests and 170 resource offers already takes
103

about 22 minutes on average with CPLEX, a state-of-the-art commercial optimization
engine, with the solver not being able to find any feasible solution within 30 minutes in
14 out of 30 simulation runs [self-citing]. In practice, especially with interactive
applications which consist of rapid CPU bursts intermitted by I/O-requests, users
typically require the resources in a timely manner and the overhead for determining
allocations must be kept as low as possible. In interval allocation mode, the scheduler in
Sun N1 Grid Engine [Anonymous, 2005], for instance, is typically executed every 2 to 15
seconds, making the computation of optimal allocation schemes infeasible in practice.
This computational hardness forms the rationale for an allocation scheme based on a
greedy heuristic which, in its basic form, has been proposed by Lehmann et al.
[Lehmann et al., 2002] and Mu’alem and Nisan [Mu’alem and Nisan, 2002] for clearing
single-sided single-attribute combinatorial auctions. We extend this heuristic to the
setting of double-sided multi-attribute auctions. The greedy allocation scheme works as
follows:
(1) Sort jobs j ∈ J in non-increasing order of their reported willingness to pay b j , and
nodes n ∈ N in non-decreasing order of their reported reservation prices r n .
(2) Run sequentially through the job ranking, starting with the highest ranked job. For
each job j, try to allocate this job to the cheapest feasible node which has still idle
capacity left.
The basic idea behind this heuristic is that, by sorting the jobs and nodes with respect
to their reported valuation in Step (1), we try to greedily maximize the difference b j – r n in
the objective function of the exact formalization of the winner determination problem
above for each job assignment.
104

Example: Applying the greedy allocation scheme to the example generates the
allocation schedule depicted in Table 2 with welfare of V greedy = 6,570 MU, while the
optimum is 6,858 MU. In this example, job J6 is willing to pay up to 17 MU per unit of
computing power and timeslot, which is more than any other job is willing to pay.
Consequently, J6 is ranked first in the sorting phase, while J4 and J6 are ranked second
and third respectively. After having allocated these three jobs, the remaining jobs either
do not fit on nodes N1 and N2 at the same time or are not willing to pay these nodes
reservation prices and are thus not allocated.
Table 2: Sample Greedy Allocation Scheme
t 1 2 3 4 5 6 7 8 9
N1 J2 J2, J6 J2, J6 J2, J6 J2, J6 J2, J6 J2, J6
N2 J4 J4 J4 J4 J4 J4
This allocation scheme runs in polynomial time of the size of its input, i.e. the number
of resource requests and offers. The sorting phase (Step (1)) runs in O(|J|log|J|) and
O(|N|log|N|) while the allocation phase (Steps (2) to (4)) runs in O(|J||N|). This
computational speed is demonstrated by numerical experiments. In contrast to the 22
minutes of the exact integer program formalization, the heuristic clears order books with
170 resource requests and 170 resource offers in about 40 milliseconds on average
[self-citing].
The computational speed of heuristics generally comes at the expense of efficiency.
Let V* be the (optimal) efficiency generated by the exact mechanism and V greedy the
efficiency generated by the greedy heuristic.
Proposition 1: In the worst case, the performance ratio V greedy /V* can be made
arbitrarily close to zero.
See Appendix A for the proof. While this is clearly a negative result, even the
standard GAP is known to be APX-hard to approximate and only a 2-approximation
105

exists [Chekuri and Khanna, 2000]. Due to the multiple resources and time constraints,
the structure of our allocation problem is tremendously more complex than the standard
GAP, and these bounds thus probably do not hold for our model. Furthermore, the
approximation techniques used to obtain these bounds do not allow for constructing
truthful prices [Mu’alem and Nisan, 2002], a key feature of our mechanism as we will
present below. Those special worst-cases, however, are merely of theoretical value and
the heuristic’s efficiency needs to be evaluated in more realistic settings. Numerical
experiments show that the greedy heuristic approximates the optimal efficiency
generated by the exact integer program between 88-96% [self-citing].
Besides its computational speed and approximate efficiency, the heuristic allocation
scheme implements desirable strategic properties as it limits the potential gain of both
resource requesters and providers to manipulate the mechanism by reporting resource
characteristics, i.e. computing and memory requirements and time constraints, other
than their true characteristics. Truthfulness with respect to a job’s resource requirements
is straightforward. If a job’s resource requirements are understated, the job will not be
finished in time; it has to pay for the used resources but these are of no value.
Overstating a job’s requirements either increases the job’s payment or the job is not
scheduled at all. Analogously, resource providers clearly cannot gain from understating
the amount of available resources. Given sufficiently severe penalties, providers cannot
gain from overstating their available resources either, since they may then not be able to
execute allocated jobs.
Moulin [Moulin, forthcoming] introduces two further strategic properties of
mechanism’s which are relevant in this regard: A mechanism is said to be split-proof if
users cannot gain from splitting their jobs (nodes) into several smaller jobs (nodes). A
106

mechanism is said to be merge-proof if users cannot benefit from merging several jobs
(nodes) to one bigger job (node).
Proposition 2: The greedy heuristic is both split- and merge-proof.
See Appendix A for a sketch of the proof. Consequently, the strategy space of selfish
users is restricted to misstating the valuation for resources only, i.e. the maximum
willingness to pay respectively the reservation price. In the following section, we will
propose pricing schemes which further limit the mechanisms vulnerability in this respect.
Pricing the Outcome
Market mechanisms consist of two basic components: an allocation scheme and a
pricing scheme. While the allocation component (i.e. the winner determination problem)
generally accounts for the technical peculiarities of the Grid setting, the objective of the
pricing component is to induce the selfish agents in the market environment to act
towards the general objective, i.e. maximize efficiency. As Lai [Lai, 2005] points out, “the
attention of system designers, especially those designing scalable systems, should also
be balanced with equal concern given to strategic behavior”. To this end, in this section
we present two pricing schemes which may be used in conjunction with the greedy
heuristic and analytically derive both positive and negative results.
Proportional CV-Pricing
The concept of critical-value-based pricing (henceforth “CV-Pricing”) has been
introduced and studied by Lehmann et al. [Lehmann et al., 2002] and Mu’alem and
Nisan [Mu’alem and Nisan, 2002]. Let φ j R +
be the minimal valuation per unit of
computing power and time slot which the winning job j would have needed to report to
the mechanism in order to still remain in the allocation. Then, with CV-Pricing, we set the
107

price of job j to p j = φ j c j (e j – s j + 1) if j is in the allocation and p j = 0 else. In the remainder
of this paper, we assume resource requesters to have quasi-linear utility functions.
Proposition 3: The greedy heuristic with critical-value-based prices for resource
requests makes truthtelling a (weakly) dominant strategy.
Refer to Appendix A for the proof. CV-Pricing is a powerful tool as it tremendously
simplifies the strategy space of resource requesters: With truthful prices in dominant
strategies, there is no need to reason about the strategies of other users. The critical
value of a job essentially hinges on the competition of other jobs for the same resources.
If there is no competition, the job is only required to pay the reservation price. Otherwise,
the job at least needs to outbid these competing jobs. If there is sufficient competition
such that critical values are above reservation prices, CV-Pricing of resource requests
generates a surplus, i.e. overall payments exceeding the providers’ reservation prices.
The major drawback of CV-Pricing in our setting is the computational cost of
determining the critical values. In order to calculate the critical value for a specific
allocated request, it is not possible to simply take the valuation of the job with the
highest ranking which does not get executed since this job may not be executable in
general due to capacity constraints. Moreover, removing j from the winning allocation
might change the allocation of other jobs within the winning allocation as well.
Consequently, to determine the critical value for each winning job j, we need to
determine the allocation without j: We successively allocate all other jobs from the
ranking and, after having allocated a job, check whether j can still be accommodated.
Note that, while the mechanism seems to suffer from the same computational problems
than the prominent VCG mechanism, we can now compute the critical values in
polynomial time due to our greedy allocation scheme.
108

The subsequent question is, how the surplus generated on the demand-side of the
market (i.e. revenue exceeding reservation prices) can be distributed to resource
providers so as to also produce truthful payments. Unfortunately, CV-Pricing is not
applicable to the pricing of resource offers. It requires a binary decision in the sense that
a job is only allocated as a whole, or not at all. In our model, however, resource offers
are divisible. Worse, the following proposition holds:
Proposition 4: There is no payment scheme which can complement the greedy
heuristic so as to generate truthful payments to resource providers.
See Appendix A for a proof sketch. This is clearly a strong negative result as it
extends across any payment scheme for the greedy heuristic.
As a consequence of Proposition 4, we cannot design a payment scheme which
precludes any strategic behavior of resource providers but we can only try to limit these
strategic possibilities. A first approach may be to distribute any surplus according to the
providers’ contribution of computing power, the key resource in our model. Let
S : =
j∈J
jnt j
∑ p −
r
j∈J
j ∈ ∈
n
n
∑ ∑ ∑
n N t T
c
be this surplus. The first term captures the total revenue collected by CV-Pricing,
whereas the second term captures the providers’ reservation prices for the allocation
schedule at hand. Then, with proportional payments, provider n receives payments
amounting to
∑t
∈T
∑
x
c
x
c
∑t∈T
∑ j∈J
∑t
T ∑j
J∑
j∈J
jnt j
jnt j
r ⋅
+ S ⋅
n
,
c n
x
∈ ∈ m∈N
jmt
c
j
where the first summand captures n’s reservation price of the allocation schedule and
the second summand captures n’s proportional payments. The rationale for proportional
x
c
109

payments is that the share of the surplus which is allotted to n does not directly depend
on n’s reported reservation price. Instead, it only depends on the final allocation
schedule, which can only be influenced by n to a rather limited extent depending on the
competition on both sides of the market. The prices and payments generated by
Proportional CV-Pricing for the sample requests and offers are listed in Table 3.
Table 3: Sample Prices and Payments
Proportional CV-Pricing K-Pricing (k = 0.5)
Job j p j Node n p n Job j p j Node n p n
J2 2,940 N1 6,165.88 J2 2,695 N1 6,820
J4 5,112 N2 5,846.12 J4 5,751 N2 5,751
J6 3,960 J6 4,125
J1, J3, J5 0 J1, J3, J5 0
Σ 12,012 Σ 12,012 Σ 12,571 Σ 12,571
K-Pricing
K-Pricing is an alternative pricing scheme to Proportional CV-Pricing. It was
introduced in Schnizler et al. [Schnizler et al., forthcoming] for double-sided
combinatorial auctions. The basic idea is to distribute the welfare generated by the
allocation algorithm between resource requesters and resource providers according to a
factor k [0,1]. For instance, assume an allocation of resources from a specific provider
to a specific requester. The requester values these resources at 10 MU while the
provider has a reservation price of 5 MU. Then the (local) welfare of this transaction is
10 MU – 5 MU = 5 MU and k * 5 MU of the surplus is allotted to the requester – thus
having to pay 10 MU – k * 5 MU – and (1 – k) * 5 MU is allotted to the provider – thus
receiving 5 MU + (1 – k) * 5 MU.
K-Pricing has two main advantages. The distribution of welfare among resource
requesters and providers can be flexibly pre-defined by setting the factor k accordingly,
thus allowing for both fairness and revenue considerations. Prices can be determined in
110

polynomial runtime. On its downside, however, it only yields approximately truthful prices
and payments on both sides of the market.
Example: The prices and payments generated by Proportional CV-Pricing and K-
Pricing with k = 0.5 are listed in Table 3. By construction, both pricing schemes generate
budget-balanced prices and payments. In this example, K-Pricing produces higher
revenue (12,571 MU) than Proportional CV-Pricing (12,012 MU).
Evaluation
The analytic evaluation in the previous section allowed us to derive some general
insights into the strategic and computational properties of the presented pricing
schemes. These general properties, however, only give rather limited guidance to the
market operator who has to choose an “adequate” pricing scheme for the strategic
setting at hand. In this section, we thus employ a numerical simulation in order to obtain
more detailed insights into the properties of the presented pricing schemes, in particular
with respect to:
• The distribution of welfare among resource requesters and providers;
• The incentives of selfish users to misreport to the mechanism about their
valuations;
• The computational influence on the mechanism as a whole.
Data Generation
Based on our model and the corresponding bidding language, there is a set of
parameters upon which problem instances (henceforth “order books”) need to be
generated (cf. Table 4). The resource requirements of jobs were randomly drawn from a
lognormal distribution with mean 4 and variance .28 (in log space), whereas the
111

esource characteristics of nodes were drawn from a lognormal distribution with mean
4.5 and variance .16. The lognormal distribution is recommended by Feitelson
[Feitelson, 2002] for the creation of parameters such as file size. Representing a normal
distribution in log space, it is modal and positively skewed. The start times and end
times have been drawn from normal distributions while the maximum willingness to pay
and reservation prices were drawn from uniform distributions.
Data Analysis
Table 4: Simulation Settings
Parameter Resource Requesters Resource Providers
Computing power Lognormal (4, .28) Lognormal (4.5, .16)
Memory Lognormal (4, .28) Lognormal (4.5, .16)
Start time Normal (2, .5) Normal (2, .5)
End time Normal (6, .5) Normal (8, .5)
Valuation Uniform [10, 20] Uniform [7.5, 12.5]
The simulations were run on an Intel Pentium Xeon computer with 3.2 GHz and 2 GB
memory. We simulated four pricing schemes: K-Pricing with k = .7 (i.e. 70% of welfare is
allocated to the demand side), k = .5, k = .3, and Proportional CV-Pricing. We created
200 order books in total based on the distributions listed above. We then fed each
combination of the heuristic and these pricing schemes with the same set of order
books, i.e. we performed 200 runs per setting. The large number of runs is due to the
number of parameters and their interdependencies (e.g. recall that the total valuation of
a job (node) equals its reported valuation times its required computing power times its
time span) which may cause heavy statistical noise.
Welfare
In most market settings, the distribution of welfare (i.e. the sum over the true
valuations of all winning resource requesters less the sum over the true reservation
prices of all winning resource providers) among resource requesters and providers will
112

e the primary design consideration: Which side of the market receives what fraction of
the welfare that is generated by the allocation scheme? Table 5 lists the results of our
simulation with respect to this question. In a simulation with 20 requesters and 20
providers, the heuristic allocation scheme generates welfare amounting to 35,735.33 MU
on average across 200 runs. For K-Pricing, the answer to the above question is simple,
as the demand side receives the fraction k of the welfare whereas the supply side
receives 1 – k. The answer is more complex for Proportional CV-Pricing. In this
simulation, Proportional CV-Pricing assigns about 71% of welfare to the demand side.
Resource providers need to abandon a comparably large fraction of welfare to the
requesters in order to induce these to truthful reporting. The difference b j – φ j between
the reported valuation of request j and its critical value may thus be interpreted as a
discount. Recall from above that the critical values of jobs heavily depend on the
competition in the market and that the providers’ surplus will thus be higher in more
competitive settings.
Table 5: Distribution of Welfare
Pricing Scheme Requesters’ surplus Providers’ surplus
K-Pricing (k = 0.3) 10,720.6 (30%) 25,014.73 (70%)
K-Pricing (k = 0.5) 17,867.66 (50%) 17,867.66 (50%)
K-Pricing (k = 0.7) 25,014.73 (70%) 10,720.6 (30%)
Proportional CV-Pricing 25,353.20 (70.95%) 10,382.13 (29.05%)
The results listed in Table 5 assume truthful reporting of valuations by both resource
requesters and providers. Consequently, we want to examine the strategic properties of
the presented pricing schemes in more detail.
Manipulating requesters
In a first setting, we analyze to what extent a single requester can benefit from
reporting a valuation other than her true valuation. Let v j be the true valuation for request
113

j. Then we let j report b j = α * v j and vary α from .7 to .8 up to 1.3, that is b j is a
percentage bid of the true valuation. For j, there are intuitively two possible strategies of
how to deviate from truthful bidding:
α < 1, i.e. j understates its true valuation. With K-Pricing, there are two somewhat
opposite implications of this strategy compared to truthful bidding: On the one hand (the
adverse effect), j risks getting a lower ranking and may thus not be allocated at all, or it
may be allocated to a node n with a higher reservation price and may thus obtain a
smaller fraction of the resulting welfare spread b j – r n . On the other hand (the beneficial
effect), j may remain in the allocation and may instead even be able to obtain a higher
fraction of the welfare spread, e.g. if it is still allocated to the same node as with truthful
bidding. Which effect finally prevails depends on the individual setting.
α > 1, i.e. j overstates its true valuation. Equivalently, j may improve its ranking but
risk getting a lower fraction of the welfare spread. On the other hand, it may instead
even obtain a larger fraction.
Table 6 lists the average utility gain across all 200 runs for a single requester from
misstating compared to truthful reporting in each of the four pricing schemes for the
request side of the market.
Table 6: Average utility gain from misstating for a single
requester
α (b j = α* v j )
K-Pricing K-Pricing K-Pricing Proportional
(k = 0.3) (k = 0.5) (k = 0.7) CV-Pricing
.7 149.40 -146.96 -443.32 -405.20
.8 327.56 108.54 -110.48 -93.69
.9 177.86 52.41 -73.04 -32.48
1.1 -306.31 -191.12 -75.92 -28.71
1.2 -632.36 -397.81 -163.25 -53.81
1.3 -949.66 -598.58 -247.49 -53.81
In the simulation we found that with α = .9, the manipulating requester dropped out of
the allocation in 16.5% of the runs, with α = .8 in 25% of the runs, and with α = .7 the
114

equester dropped out of the allocation already in 47% of the runs. This component of
the adverse effect stems from the design of the ranking in the allocation heuristic and
thus holds for all pricing schemes.
Since we fed each pricing scheme with identical sets of 200 order books, we applied
one-tailed matched pairs t-tests to determine the statistical significance of these
deviations. The p-values are given in Table B1 in Appendix B. The test shows a
statistically significant utility gain from understating (α < 1) for K-Pricing with k = .3 and
for α = .8 and α = .9 with k = .5. For all other settings, the test shows a statistically
significant utility loss from misstating (i.e. both under- and overstating). In essence, the
larger the factor k – i.e. the larger the fraction of welfare which is allocated to resource
requesters – the less can these requesters gain from misstating. We hypothesize that
this effect can be explained with the two opposite effects pointed out above: On
average, the larger the factor k, the lower the beneficial and the higher the adverse
effect from understating the valuation for a specific job. On the other side, for overstating
resource requesters, we can see that on average, the larger k, the higher the beneficial
and the lower the adverse effect, whereas the latter still prevails.
As stated in Proposition 3, with Proportional CV-Pricing, requesters cannot gain from
misstating their valuations in any case, so the same holds for the averages across all
runs.
Manipulating Providers
In a second – and from the setup symmetric – setting, we analyze to what extent a
single provider can benefit from reporting a reservation price other than her true
reservation price. The implications from under- respectively overbidding are essentially
analogous to those on the demand side:
115

α < 1: As for resource requesters, there are two opposite effects of understating the
reservation price. Recall that the payment to a resource provider consists of two
components: its reservation price and its fraction of the surplus, i.e. total revenue less
total reservation prices. On the one hand, by understating its reservation price, the node
inherently lowers the first component. On the other hand, the node may achieve a higher
ranking and may thus be allocated more demand and thus a higher fraction of the
surplus.
α > 1: Again, the implications of overstating the reservation price are essentially the
reversion of the consequences of understating this valuation.
Table 7 lists the average utility gain across all 200 runs for a single requester from
misstating compared to truthful reporting in each of the four payment schemes. The p-
values for the one-tailed matched pairs t-tests for this setting are listed in Table B2 in
Appendix B.
Table 7: Average utility gain from misstating for a single
provider
α (b j = α* v j )
K-Pricing K-Pricing K-Pricing Proportional
(k = .3) (k = .5) (k = .7) CV-Pricing
.7 573.54 23.19 -527.16 -1026.51
.8 425.28 58.92 -307.45 -598.16
.9 -37.23 -125.91 -214.58 -260.26
1.1 -427.84 -248.67 -69.49 26.03
1.2 -744.00 -449.03 -154.06 15.18
1.3 -1067.02 -698.20 -329.38 -187.86
For K-Pricing with k = .3, the average utility gain from understating the reservation
price by 20% (α = .8) and by 30% (α = .7) is statistically significant with the p-values
being 9 * 10 –10 and 1 * 10 –12 respectively. While there is still a slight potential gain from
understating with k = .5, this gain is not statistically significant with p-values of .103 for α
= .8 and .325 for α = .7. For k = .7, there is no gain from both under- and overstating.
116

These results can again be explained by means of the two opposite implications of
understating respectively overstating. As discussed above, if the requester n
understates its reservation price r n , it directly lowers this component of its payment, but
may be able to extract a larger fraction of the spread b j – r n . However, since the
requester receives (1 – k) * (b j – r n ) of this spread, this beneficial effect of understating
gradually diminishes as k converges to 1.
Proportional CV-Pricing punishes the understatement of the reservation price more
than K-Pricing. On its downside, it offers potential gains from overstating. But these
gains only seem to support slight deviations. Looking at the p-values in Table B2, the
gain from overstating is not statistically significant (.235 for α = 1.1 and .382 for α = 1.2).
Scalability
As reported above, numerical experiments show that the greedy heuristic generates
highly efficient allocations at low computational cost. However, not only the allocation
scheme, but also the pricing scheme places a computational burden on the mechanism
as a whole. As presented above, with Proportional CV-Pricing, in order to determine the
critical value of a winning job, the allocation without this job needs to be determined.
In order to examine the impact of each pricing scheme on the mechanism’s
computational complexity in more detail, we varied the number of requests and offers in
the market from 20 orders per side, 40 orders per side, and so forth up to 200 orders per
side. We again randomly generated 200 order books for each of these settings, and
performed the heuristic with both K-Pricing and Proportional CV-Pricing. The average
runtimes across all 200 runs are listed in Table 8. These simulation results confirm the
computational superiority of the K-Pricing scheme. Complemented with K-Pricing, the
heuristic only takes about 90 milliseconds to clear comparably large order books with
117

200 requests and 200 offers. The computational cost of Proportional CV-Pricing is
illustrated by the heuristic with Proportional CV-Pricing, taking 2,400 milliseconds on
average for the largest order books.
Table 8: Average runtime in milliseconds
Orders per
market side
20 40 60 80 100 120 140 160 180 200
K-Pricing 2 7 16 26 41 48 52 58 70 87
Proportional
CV-Pricing
6 29 77 162 302 506 813 1,214 1,717 2,400
If employed in interval scheduling mode, Sun Microsystems’s N1GE scheduler is
typically executed every 15 seconds. We found that, if complemented with K-Pricing, the
heuristic can clear up to 2,500 orders per side on average across 200 order books within
this timeframe (12.7 seconds). If compared to the size of PlanetLab, the largest testbed
for networking and distributed computing which currently comprises about 700 nodes
[Peterson et al., 2006], these results underline our mechanism’s scalability and
applicability to practical scenarios.
Implications for Market-based Resource Allocation
From a welfare perspective, K-Pricing offers the nice feature that, given truthful
information from all users, welfare can be distributed among resource requesters and
providers according to the parameter k, which can act as adjusting crew. The welfare
implications of Proportional CV-Pricing are more involved, as the distribution of welfare
hinges on the competition on both sides of the market. In contrast to K-Pricing, there is
no direct adjusting screw for the market operator. However, via the critical values of
requests, Proportional CV-Pricing implements a desirable economic dynamicity in the
sense that prices and payments adequately reflect resource scarcity. If resources are
not scarce, critical values will be low, and resource providers may just receive their
118

eservation prices. With increasing scarcity, critical values converge to the requesters’
maximum willingness to pay, and resource providers extract a larger fraction of welfare.
Considering strategic users, Proportional CV-Pricing clearly shows its strengths, as
resource requesters cannot profit from misreporting and resource providers can only
slightly benefit from overstating their reservation prices. With K-Pricing, the freedom in
choosing k becomes somewhat restricted. On the one hand, if k – and thus requesters’
surplus – is small, both resource requesters and providers have a significant incentive to
understate their true valuations. On the other hand, if k is large, and hence providers’
surplus low, this reduces the incentive for resource providers to contribute their idle
resources to the Grid. With respect to computational considerations, K-Pricing clearly
outperforms Proportional CV-Pricing.
In large-scale settings, K-Pricing will generally be preferable to Proportional CV-
Pricing due to its computational simplicity. Moreover, as Jackson [Jackson, 1992]
shows, as market size increases, the potential gains from misreporting become smaller.
The increasing competition assumes the pricing scheme’s role of inducing users to
truthful behavior. In small-scale settings, the importance of computational considerations
naturally diminishes and, in line with Jackson’s [Jackson, 1992] results for increasing
market sizes, strategic considerations gain in importance as the market size decreases.
Proportional CV-Pricing may thus prove superior to K-Pricing in these settings as it not
only ensures truthful reporting on average, but in dominant strategies.
Returning to our discussion of Grid schedulers in the introduction, there are generally
two ways in which market mechanisms can be integrated into Grid schedulers.
119

Figure 2a. Market mechanisms to determine
priority values / shares
Figure 2b. Market mechanisms to determine
both priority values and allocations
Firstly, market mechanisms may be used to determine the priority values respectively
shares of requests in the Grid scheduler’s central queue (cf. Figure 2a). In conjunction,
as with Sun’s N1 Grid Engine, a technical scheduler can then subsequently be used to
allocate the requests in this queue to Grid resources. As proposed in [self-citing], the
pay-as-bid mechanism of Sanghavi and Hajek [Sanghavi and Hajek, 2004] or marketbased
proportional share [Chun and Culler, 1999] are two mechanisms which may be
suitable for this design option, which is particularly qualified to be employed as a so
called online mechanism in order to schedule interactive applications, meaning resource
requests and resources are allocated as soon as they arrive in the system. Secondly,
market mechanisms may assume the role of both establishing priorities and allocations,
thus also displacing the technical scheduler (cf. Figure 2b). This class of mechanisms is
typically executed periodically. Larger clearing periods allow for more efficient allocation
decisions since there are likely to be more requests and offers over which the
mechanism can try to optimize. On the other hand, in dynamic settings where users
require fast allocation decisions, the clearing period can be made as small as a couple
of seconds. The market mechanism presented in this paper falls into this category.
120

Concluding Remarks and Future Work
In the introduction, we discussed that technical schedulers based on system-centric
metrics – the current approach to allocating Grid resources – do not achieve efficient
allocations in the economic sense since the pre-defined and static policies and priorities
cannot cope with the dynamic and large-scale nature of Grids, do not allow requesters to
specify their valuations for resources, and do not provide immediate incentives to
contribute to the Grid. Market mechanisms are a promising approach to Grid scheduling
by explicitly addressing these deficiencies. We presented related work on market
mechanisms for Grid resource allocation and argued that these current approaches do
not fully consider dependencies between multiple Grid resources, scalability issues, and
strategic users which may try to manipulate the mechanism in their own benefit. The
contribution of this paper is the proposal of a mechanism which achieves a distinct
trade-off between allocative efficiency, computational tractability, and incentive
compatibility. To this end, we designed a greedy heuristic which achieves near-optimal
resource allocations at low computational cost. We subsequently presented alternative
pricing schemes which may complement this allocation scheme and derived some
general results on their strategic and computational properties. In order to examine
these properties in more detail, we employed a numerical experiment which we
described in our evaluation section. We compared these results for the presented pricing
schemes and derived some general recommendations and design options for integrating
market mechanisms into Grid schedulers.
In the future, we intend to perform further analyses of the presented pricing schemes,
e.g. by means of experiments and agent-based simulations. Moreover, we plan to
integrate a prototype of the market mechanism into MOSIX, a state-of-the-art Grid
121

system [Barak et al., 2005]. The heuristic’s approximation of efficiency needs to be
compared in detail to the efficiency produced by the exact formalization of the winner
determination problem. Extensions of the heuristic allocation scheme, such as the use of
more sophisticated norms in the sorting phase, as well as their impact on the
mechanism’s strategic properties, may be promising areas for future research.
Appendix A – Proofs
Proof of Proposition 1: W.l.o.g., we consider valuations and computing power only.
Suppose one resource offer with characteristics (r, c¯) = (0 MU, k), k ∈ ℵ, and two resource
requests (b 1 , c 1 ) = (2 MU, 1) and (b 2 , c 2 ) = (1 MU, k). The heuristic generates welfare of 2 MU
while the optimum is k MU. Clearly, V greedy /V* → 0 for k → ∞.
Proof sketch of Proposition 2: The sorting in Step 1 of the heuristic depends on the
reported valuations per unit of computing power. Consequently, both requesters and providers
can neither simultaneously improve the rankings of two requests (offers) by merging them, nor
can they simultaneously improve the rankings of two requests (offers) which result from splitting
a request (offer).
Proof of Proposition 3: We assume resource requesters to have quasi-linear utility
functions. Let u j be the resulting overall utility for request j, v j j’s true valuation for a unit of
computing power per time slot, and p j the resulting payment. Then u j = v j (e j – s j + 1)c j – p j if j is
allocated and u j = 0 else. Assume j has reported truthfully, i.e. b j = v j .
Suppose j has won, i.e. b j = v j > φ j and p j = φ j (e j – s j + 1)c j . Reporting a lower valuation b j < v j
either leaves j in the winning allocation while it does not change φ j and consequently it does not
change u j , or it leads to j being rejected and thus uj’ = 0 < (v j – φj)(ej – sj + 1)cj = uj. Reporting a
higher valuation b j > v j leaves j being accepted while it does not change φj and consequently it
does not change uj.
122

Now suppose j has lost, i.e. b j = v j < φ j and p j = 0. Reporting a lower valuation b j < v j leaves j
being rejected. Reporting a higher valuation b j > v j either leaves j being rejected (b j < φj), or it
leads to j being accepted (b j > φj > v j ) and thus uj’ = (b j – φj)(ej – sj + 1)cj < 0 = uj.
Proof sketch of Proposition 4: A key requirement for achieving truthfulness is designing a
monotonic allocation scheme in the sense that providers cannot do better by overstating their
reservation prices [Lehmann et al., 2002, Mu’alem and Nisan, 2002]. By means of a simple
example we will show that providers can potentially increase their load by overstating their
reservation prices.
W.l.o.g., we consider valuations and computing power only. Assume two resource requests,
(10 MU, 1) and (9 MU, 2), and two truthful resource offers, (1 MU, 2) and (2 MU, 2). Then the
first offer will be allocated one unit of computing power. Now assume the first provider had
reported (3 MU, 2) instead. Then she would have been allocated 2 units of computing power.
Appendix B – p-values
Table B1: p-values of one-tailed matched
pairs t-tests for the utility gains from
misstating for a single requester provider
α
K-Pricing K-Pricing K-Pricing Proportional
(k = 0.3) (k = 0.5) (k = 0.7) CV-Pricing
.7 .004 .007 3 * 10 –10 6 * 10 –15
.8 9 * 10 –15 .003 .005 1 * 10 –7
.9 1 * 10 –12 .036 .023 9 * 10 –5
1.1 3 * 10 –53 2 * 10 –30 2 * 10 –5 0.001
1.2 1 * 10 –57 3 * 10 –46 5 * 10 –14 4 * 10 –4
Table B2: p-values of one-tailed matched
pairs t-tests for the utility gains from
misstating for a single provider
α
K-Pricing K-Pricing K-Pricing Proportional
(k = 0.3) (k = 0.5) (k = 0.7) CV-Pricing
.7 1 * 10 –12 .325 4 * 10 –39 5 * 10 –73
.8 9 * 10 –10 .103 4 * 10 –22 3 * 10 –56
.9 .065 3 * 10 –11 1 * 10 –37 6 * 10 –32
1.1 1 * 10 –11 2 * 10 –8 .006 .235
1.2 1 * 10 –20 7 * 10 –16 6 * 10 –6 .382
1.3 2 * 10 –26 9 * 10 –23 2 * 10 –13 4 * 10 –4
1.3 1 * 10 –58 2 * 10 –52 4 * 10 –22 4 * 10 –4 123
Bibliography
Anonymous. (2005) Sun N1GE 6 User’s Guide. Sun Mircosystems, Inc.
Anonymous. (2006) Grid Computing: A Vertical Market Perspective 2006-2011. The Insight Research Corportation.
Anonymous (2007a) "GlaxoSmithKline: Implementing Breakthrough Technology for Clinical Development Modeling
and Simulation," United Devices, http://www.ud.com/company/resources/files/cs_gsk.pdf (May 3, 2007).
Anonymous (2007b) "Sal. Oppenheim claims first-mover advantage with a grid solution from IBM and Platform," IBM
Corporation,
http://www.platform.com/NR/rdonlyres/D0860543-DC4C-4D43-935F-
2A86387A224E/0/3597_SalOppenheim.pdf (May 3, 2007).
AuYoung, A., B. N. Chun, A. C. Snoeren, and A. Vahdat (2004) “Resource Allocation in Federated Distributed
Computing Infrastructures,” in Proceedings of the 1st Workshop on Operating System and Architectural Support for
the On-demand IT Infrastructure. Boston, MA, USA.
Bapna, R., D. Sanjunkta, R. Garfinkel, and J. Stallaert (forthcoming) "A Market Design for Grid Computing," INFORMS

Journal of Computing.
Barak, A., A. Shiloh, and L. Amar (2005) “An Organizational Grid of Federated MOSIX Clusters,” in Proceedings of the
5th IEEE International Symposium on Cluster Computing and the Grid (CCGrid). Cardiff, Wales.
Barham, P., B. Dragovic, K. Fraser, S. Hand et al. (2003) “Xen and the art of virtualization,” in Proceedings of the 19th
ACM symposium on Operating systems principles, pp. 164-177. Banff, Alberta, Canada.
Buyya, R., D. Abramson, J. Giddy, and H. Stockinger (2002) "Economic models for resource management and
scheduling in Grid computing," Concurrency and Computation: Practice and Experience (14) 13-15, pp. 1507-1542.
Buyya, R., D. Abramson, and S. Venugopal (2005) "The grid economy," Proceedings of the IEEE (93) 3, pp. 698-714.
Carr, N. (2005) "The End of Corporate Computing," Sloan Management Review (46) 3, pp. 67-73.
Chekuri, C. and S. Khanna (2000) “A PTAS for the multiple knapsack problem,” in Proceedings of the 11th annual
ACM-SIAM symposium on Discrete algorithms pp. 213-222. San Francisco, California, USA
Chun, B. and D. E. Culler. (1999) Market-based proportional resource sharing for clusters. Millenium Project Research
Report.
Feitelson, D. G. (2002) “Workload Modeling for Performance Evaluation,” in, vol. 2459 Performance Evaluation of
Complex Systems: Techniques and Tools, Lecture Notes in Computer Science: Springer Verlag, pp. 114-141.
Foster, I., C. Kesselman, and S. Tuecke (2001) "The Anatomy of the Grid: Enabling Scalable Virtual Organizations,"
International Journal of High Performance Computing Applications (15) 3, pp. 200-222.
Jackson, M. O. (1992) "Incentive compatibility and competitive allocations," Economics Letters (40), pp. 299-302.
Lai, K. (2005) "Markets are Dead, Long Live Markets," ACM SIGecom Exchanges (5) 4, pp. 1-10.
Lehmann, D., L. I. O’Callaghan, and Y. Shoham (2002) "Truth Revelation in Approximately Efficient Combinatorial
Auctions," Journal of the ACM (49) 5, pp. 577-602.
MacKie-Mason, J. K. and M. P. Wellman (2006) “Automated Markets and Trading Agents,” in L. Tesfatsion and K. L.
Judd (Eds.) Handbook of Computational Economics, vol. 2: Agent-Based Computational Economics: North-Holland.
Moulin, H. (forthcoming) "Proportional scheduling, Split-proofness, and Mergeproofness," Games and Economic
Behavior.
Mu’alem, A. and N. Nisan (2002) “Truthful Approximation Mechanisms for Restricted Combinatorial Auctions,” in
Proceedings of the 18th National Conference on Artificial Intelligence, pp. 379-384. Edmonton, Alberta, Canada.
Neumann, D., S. Lamparter, and B. Schnizler (2006) “Automated Bidding for Trading Grid Services,” in Proceedings
of the 14th European Conference on Information Systems (ECIS). Gothenburg, Sweden.
Peterson, L., A. Bavier, M. E. Fiuczynski, and S. Muir (2006) “Experiences Building PlanetLab,” in Proceedings of the
7th Symposium on Operating Systems Design and Implementation. Seattle, WA, USA.
Regev, O. and N. Nisan (1998) “The POPCORN Market – An Online Market for Computational Resources,” in
Proceedings of the 1st International Conference on Information and Computation Economies, pp. 148-157. New York,
USA.
Sanghavi, S. and B. Hajek (2004) “Optimal allocation of a divisible good to strategic buyers,” in Proceedings of the
43rd IEEE Conference on Decision and Control. Atlantis, Paradise Island, Bahamas.
Schnizler, B., D. Neumann, D. Veit, and C. Weinhardt (forthcoming) "Trading Grid Services – A Multi-attribute
Combinatorial Approach," European Journal of Operational Research.
Waldspurger, C., T. Hogg, B. A. Huberman, J. O. Kephart et al. (1992) "Spawn: A Distributed Computational
Economy," IEEE Transactions on Software Engineering (18) 2, pp. 103-117.
Wellman, M. P., W. E. Walsh, P. R. Wurman, and J. K. MacKie-Mason (2001) "Auction protocols for decentralized
scheduling," Games and Economic Behavior (35) 1-2, pp. 271-303.
Wolski, R., J. S. Plank, J. Brevik, and T. Bryan (2001) "Analyzing Market-Based Resource Allocation Strategies for the
Computational Grid," International Journal of High Performance Computing Applications (15) 3, pp. 258-281.
124

Trading Grid Services –
A Multi-attribute Combinatorial Approach
Björn Schnizler, Dirk Neumann, Daniel Veit, Christof Weinhardt
{schnizler,neumann,veit,weinhardt}@iism.uni-karlsruhe.de
Institute of Information Systems and Management, University of Karlsruhe (TH),
Germany
Abstract
The Grid is a promising technology for providing access to distributed high-end computational
capabilities. Thus, computational tasks can be performed spontaneously
by other resources in the Grid that are not under the user’s control. However, one
of the key problems in the Grid is deciding which jobs are to be allocated to which
resources at what time. In this context, the use of market mechanisms for scheduling
and allocating Grid resources is a promising approach toward solving these
problems. This paper proposes an auction mechanism for allocating and scheduling
computer resources such as processors or storage space which have multiple quality
attributes. The mechanism is evaluated according to its economic and computational
performance as well as its practical applicability by means of a simulation.
Key words: Auctions and Bidding, Integer Programming, Combinatorial
Exchange, Market Engineering, Grid
1 Introduction
The increasing interconnection between computers has created a vision for
Grids. Within these Grids, computing resources such as processing power,
storage space or applications are accessible to any participant. This accessibility
has major ramifications for organizations since they can reduce costs by
outsourcing nonessential elements of their IT infrastructure to various forms
of service providers. Such emerging e-Utilities – providers offering on-demand
access to computing resources – enable organizations to perform computational
jobs spontaneously through other resources in the Grid that are not
under the control of the (temporary) user (Foster et al., 2002).
Most Grid research has been devoted to the development of hard and software
infrastructures so that access to resources is dependable, consistent, pervasive
and inexpensive (Foster and Kesselman, 2004). The specification of open
Preprint submitted to Elsevier Science 29 June 2006
125

standards that defines the interactions between different computing resources
across organizational entities hallmarks a significant milestone in this development.
With the Open Grid Service Architecture (OGSA), which specifies
fundamental middleware components for the Grid, the Grid Community has
laid the foundation for future developments. OGSA defines computer and storage
resources as well as networks, programs, databases and the like as services,
i.e. network-enabled entities that provide some capability. Using resources as
service paves the road for interoperability among heterogeneous computing
and application environments (Foster et al., 2002). OGSA and one of its reference
implementations, Globus Toolkit 1 , provide the technical infrastructure
for accessing computational services over the Grid. Computers equipped with
a Globus installation can easily access computational resources over the Grid,
where access for the user is transparent. Resource owners can become providers
by offering their services using standardized interfaces in a global directory.
The global directory allocates the services to appropriate service consumers.
With the implementation of Grid middleware, institutional arrangements are
becoming increasingly important. In essence, current Grid middleware provides
insufficient incentive to participate in the Grid. This lack of incentive
stems from the fact that most Grid middleware service directories employ
batch algorithms which determine the allocation based on either naive heuristics
such as first-come first-serve or on idiosyncratic cost functions (Buyya
et al., 2002). From an economic point of view, however, these algorithms are
harmful, as they strive to maximize a system-wide performance objective such
as throughput or mean response time rather than maximize aggregate value.
These batch algorithms can not guarantee that the buyers who value some
services will really receive them.
Market-based approaches are considered to work well in Grid settings. By assigning
a value (also called utility) to their service requests, users can reveal
their relative urgency or costs to the service, which is subject to service usage
constraints. If the market mechanism is properly defined, users may be
provided incentives to express their true values for service requests and offers.
This in turn marks the prerequisite for attaining an efficient allocation
of services, which maximizes the sum of aggregate valuations (Schnizler et al.,
2005).
In recent times, the idea of incorporating market mechanisms into Grid technology
has increasingly gained attention. Among others, the proposals include
the establishment of Open Grid Markets, where either idle or all available resources
can be traded (Shneidman et al., 2005; Lai, 2005). Despite this recent
interest in market-based approaches, research regarding market mechanisms
for Grid resources remains in its infancy. The canon of available market mech-
1 See http://www.globus.org/ for details.
2
126

anisms only insufficiently copes with the requirements imposed by the Grid.
This paper seeks to supply the need for a market mechanism that is Gridcompatible
and can thus be used for resource management in federated environments
such as an Open Grid Market.
The remainder of this paper is structured as follows: the economic properties
which the proposed market mechanism should satisfy are presented in
section 2. In section 3, comparable efforts in designing mechanisms for Grids
and related domains are discussed. Since none of these mechanisms is directly
applicable, section 4 tailors a mechanism for the Grid. In section 5, the proposed
mechanism is evaluated according to the properties laid out in section 2.
Section 6 concludes with a brief summary and an outlook for future research.
2 Desirable Properties and Requirements
The objective of an adequate market mechanism for the Grid is the efficient
and reliable provision of resources to satisfy demand. Economic approaches
explicitly assume that users act strategically in a way that does not reveal
their true valuations for resources aggregated in services. Incentive compatible
mechanisms are accordingly advantageous, as they encourage users to report
their true valuations. This allows the mechanism to maximize allocative
efficiency (i.e. the sum of the individuals’ utilities).
A critical step in designing markets is to understand the nature of the trading
object. This paper considers services which respect resource functionalities
(e.g. storage) and quality characteristics (e.g. size), dependencies and time
attributes. Relying on services instead of computational resources removes
many technical problems. For instance, the resource CPU may technically not
be offered without an appropriate amount of hard disk space on the same
computer, while a computation service offering CPU cycles already includes
the complementary resources. In the remainder of the paper, it is abstracted
from those technical details by treating resources and services as synonyms.
The following presents the design objectives and the Grid specific requirements
for the market mechanism.
2.1 Design Objectives
The theoretical basis for designing mechanisms has emerged from a branch
of game theory called mechanism design (Milgrom, 2004). Within the scope
of practical mechanism design, the primary design objective is to investigate
a mechanism that has desirable properties. The following comprises common
economic properties of a mechanism’s outcome:
• Allocative efficiency:
3
127

An allocation is efficient if the sum of individual utilities is maximized. It
is assumed that utility is transferable among all participants. A mechanism
can only attain allocative efficiency if the market participants report their
valuation truthfully. This requires incentive compatibility in equilibrium.
• Incentive compatibility:
A mechanism is incentive compatible, if every participant’s expected utility
maximizing strategy in equilibrium with every other participant is to report
its true preferences (Parkes et al., 2001).
• Individual rationality:
The constraint of individual rationality requires that the utility following
participation in the mechanism must be greater or equal to the previous
utility.
• Budget balance:
A mechanism is said to be budget balanced if the prices add up to zero for
all participants (Jackson, 2002). In case the mechanism runs a deficit, it
must be subsidized by an outside source and is therefore not feasible per-se.
• Computational tractability:
Computational tractability considers the complexity of computing a mechanism’s
outcome. With an increasing number of participants, the allocation
problem can become very demanding and may delimit the design of choice
and transfer rules.
It is clear that the objective allocative efficiency meets the general design
goal that the mechanism designer wants to achieve, whereas the remaining
categories are constraints upon the objective (Neumann, 2004).
2.2 Domain-Specific Requirements
In addition to those mechanism properties pertaining to the outcome, the
mechanism must also account for the underlying environment. The constraints
of the market participants impose very rigid requirements upon the design:
• Double-sided mechanism:
A Grid middleware usually provides a global directory enabling multiple
service owners to publish their services and multiple service requesters to
discover them. Since a market mechanism replaces these directories, it has
to allow many resource owners (henceforth sellers) and resource consumers
(buyers) to trade simultaneously.
• Language includes bids on attributes:
Participants in the Grid usually have different requirements for the quality
characteristics of Grid services and require these in different time spans. For
example, a rendering job could require a storage service with at least 250 GB
of free space for four hours in any slot between 9 a.m. and 4 p.m. The Grid
community takes these requirements into account by defining service level
agreement protocols, e.g. WS-Agreement (Ludwig et al., 2004). To facilitate
4
128

the adherence of these agreement protocols, a market mechanism fitting the
Grid is required to support bids on multiple quality attributes of services
as well as their time objectives.
• Language includes combinatorial bids:
Buyers usually demand a combination of different Grid services as a bundle
in order to perform a task (Subramoniam et al., 2002). As such, Grid services
are complementarities, meaning that participants have super-additive
valuations for the services, since the sum of valuations for single services
is less than the valuation for the whole bundle (v(A) + v(B) ≤ v(AB)).
Suppose a buyer requires services for storage, computation and rendering.
If any service, e.g. the storage service, is not allocated to him, the remaining
bundle has no value for him. In order to avoid this exposure risk (i.e.
receiving only a subset of the bundle), the mechanism must allow for bids
on bundles.
The buyer may want to submit more than one bid on a bundle as well
as many that exclude each other. In this case, the resources for the bundles
are substitutes, i.e. participants have sub-additive valuations for the services
(v(A) + v(B) ≥ v(AB)). For instance, a buyer is willing to pay a high price
for a service during the day and a low price if the service is executed at
night. However, this service may only be computed once. To express this,
the mechanism must support XOR 2 bids to express substitutes. For the
sake of simplicity, a seller’s bid is restricted to a set of OR 3 bids. This
simplification can be justified by the fact that Grid services are non-storable
commodities, e.g. a computation service currently available cannot be stored
for a later time.
• Language includes co-allocation constraints:
Capacity-demanding Grid applications usually require the simultaneous allocation
of several homogenous service instances from different providers.
For example, a large-scale simulation may require several computation services
to be completed at one time. Research literature often refers to the
simultaneous allocation of multiple homogenous services as co-allocation. A
mechanism for the Grid has to enable co-allocations and provide functionality
to control it. In this context, two cases must be considered: Firstly, it
is desirable to limit the maximum number of service co-allocations, i.e. the
maximum number of service divisions. Secondly, it may be logical to couple
multiple services of a bundle in order to guarantee that these resources are
allocated from the same seller and – more importantly – will be executed
on the same machine.
An adequate market mechanism for the Grid must satisfy these requirements
stemming from the economic environment and ideally meet the design objectives.
2 A XOR B (A ⊕ B) means either ∅, A, or B, but not AB
3 A OR B (A ∨ B) means ∅, A, B, or AB
5
129

3 Related Work
For instance, Waldspurger et al. (1992) and Regev and Nisan (2000) propose
the application of Vickrey auctions for allocating homogenous computational
resources in distributed systems. Vickrey auctions achieve truthful bidding as
a dominant strategy and hence result in efficient allocations.
Buyya et al. (2002) were among the first researchers to motivate the transfer
of market-based systems from distributed systems to Grids. Nonetheless, they
propose classical one-sided auction types which cannot account for combinatorial
bids. Wolski et al. (2001) compare classical auctions with a bargaining
market, coming to the conclusion that the bargaining market is superior to
an auction-based market. This result is less surprising since the authors only
consider classical auction formats where buyers cannot express bids on bundles.
Eymann et al. (2003) introduce a decentralized bargaining system for
resource allocation in Grids. In their simulation the bargaining systems work
fairly well; however, bids on bundles are largely ignored.
Subramoniam et al. (2002) account for combinatorial bids by providing a
tâtonnement process for allocating and pricing Grid resources. Furthermore,
Ng et al. (2005) propose repeated combinatorial auctions as a microeconomic
resource allocator in distributed systems. Nonetheless, the resources are still
considered to be standardized commodities. Standardization of the resources
would either imply that the number of resources are limited compared to the
number of all possible resources or that there are many mechanisms which are
likely to suffer due to meager participation. Both implications result in rather
inefficient allocations.
Additionally, state-of-the-art mechanisms widely neglect time attributes for
bundles and quality constraints for single resources. Hence, the use of these
mechanisms in the Grid environment is considerably diminished. The introduction
of time attributes redefines the Grid allocation problem as a scheduling
problem. To account for time attributes, Wellman et al. (2001) model singlesided
auction protocols for allocating and scheduling resources under different
time constraint considerations. However, the proposed approach is single-sided
and favors monopolistic sellers or monopsonistic buyers in a way that allocates
greater portions of the surplus. Installing competition on both sides is deemed
superior, since no particular market side is systematically given an advantage.
Demanding competition on both sides suggests the development of a combinatorial
exchange. In research literature, Parkes et al. (2001) introduce the
first combinatorial exchange as a single-shot sealed bid auction. As payment
scheme, Vickrey discounts are approximated. The approach results in approximately
efficient outcomes; however, it neither accounts for time nor for quality
constraints and is thus not directly applicable to the Grid allocation problem.
6
130

Counteractively, Bapna et al. (2005) propose a family of combinatorial auctions
for allocating Grid services. Although the mechanism accounts for quality
and time attributes and enables the simultaneous trading of multiple buyers
and sellers, there is no competition on the sellers’ side as all orders are
aggregated to one virtual order. Moreover, the mechanism does not take coallocation
constraints into account.
In reviewing the related mechanisms according to the requirements presented
in section 2, it is revealed that no market mechanism installs competition on
both sides, includes combinatorial bids, allows for time constraints, manages
quality constraints, or considers co-allocation restrictions. This paper intends
to address these deficiencies by outlining the design of a multi-attribute combinatorial
exchange for allocating and scheduling Grid services.
4 MACE: A Multi-attribute Combinatorial Exchange
The design of a market mechanism mainly affects two components: (i) the
communication language which defines how bids can be formalized and (ii)
the outcome determination by means of winner determination and pricing
rules.
The following auction format follows common assumptions of mechanism design:
participants are assumed to be risk neutral, have linear utility functions,
as well as independent private valuations and reservation prices. Hence, the
sellers’ reservation prices can be linearly transformed to any partial execution
of any bundle.
Firstly, a bidding language will be introduced which fits the requirements
specified in section 2. Furthermore, a winner determination model (allocation
rule) and a family of pricing schemes will be proposed.
4.1 Bidding Language
Meeting the requirements specified in section 2 requires a formal bidding language
that enables buyers and sellers to submit combinatorial bids, including
multiple attributes and co-allocation constraints:
Let N be a set of N buyers and M be a set of M sellers, where n ∈ N defines
an arbitrary buyer and m ∈ M an arbitrary seller. There are G discrete resources
G = {g 1 , . . . , g G } with g k ∈ G and a set of S bundles S = {S 1 , . . . , S S }
with S i ∈ S and S i ⊆ G as a subset of resources. For instance, S i = {g k , g j }
denotes that the bundle S i consists of two resources g k and g j , where g k could
be a computation service and g j a storage service.
A resource g k has a set of A k cardinal quality attributes A gk = {a gk ,1, . . . , a gk ,A k
}
7
131

where a gk ,j ∈ A gk represents the j.th attribute of the resource g k . For instance,
in the context of a Grid service, a quality attribute can be the size of a storage
service.
A buyer n can specify the minimal required quality characteristics for a bundle
S i ∈ S with q n (S i , g k , a gk ,j) ≥ 0, where g k ∈ S i is a resource of the bundle S i
and a gk ,j ∈ A gk is a corresponding attribute of the resource g k . As such, the
minimal required size of a storage service can be denoted by q n (S i , g k , a gk ,j),
e.g. 200GB. Accordingly, a seller m can specify the maximum offered quality
characteristics with q m (S i , g k , a gk ,j) ≥ 0.
For each resource g k ∈ S i , a buyer n can specify the maximum number of coallocations
in each time slot with γ n (S i , g k ) ≥ 0. This means that a buyer n can
limit the number of partial executions for each resource g k . Let γ n (S i , g k ) = D,
if the resource g k has no divisibility restrictions, where D is a large enough
constant 4 . The coupling of two resources in a bundle is represented by the
binary variable ϕ n (S i , g k , g j ) where ϕ n (S i , g k , g j ) = 1 if resources g k and g j
have to be allocated from the same bundle bid of a seller and ϕ n (S i , g k , g j ) = 0
otherwise. It is assumed that all services offered in a bundle are located on
the same machine.
Resources in the form of bundles S i can be assigned to a set of maximal
T discrete time slots T = {0, . . . , T − 1}, where t ∈ T specifies one single
time slot. A buyer n can specify the minimum required number of time slots
s n (S i ) ≥ 0 for a bundle S i . The earliest time slot for any allocatable bundle
S i can be specified by e n (S i ) ≥ 0 for a buyer n and e m (S i ) ≥ 0 for a seller
m; the latest possible allocatable time slot by l n (S i ) ≥ 0 for a buyer n and by
l m (S i ) ≥ 0 for a seller m.
A buyer n can express the valuation for a single slot for a bundle S i by
v n (S i ) ≥ 0, i.e. the maximum price for which the buyer n is willing to trade.
The reservation price for allocating a single slot of a bundle S i is denoted by
r m (S i ) ≥ 0, which represents the minimum price for which the seller m is
willing to trade.
Subsequently, a buyer n can submit a set of XOR bundle bids as an order
B n = {B n,1 (S 1 ) ⊕ . . . ⊕ B n,u (S j )},
where u is the number of bundle bids in the order and B n,f (S i ) is defined as
4 The constant D has to be greater than the total number of seller bids.
8
132

the tuple
B n,f (S i ) =(v n (S i ), s n (S i ), e n (S i ), l n (S i ),
q n (S i , g 1 , a g1 ,1), . . . , q n (S i , g G , a gG ,A j
), γ n (S i , g 1 ), . . . , γ n (S i , g G ),
ϕ n (S i , g 1 , g 2 ), ϕ n (S i , g 1 , g 3 ), . . . , ϕ n (S i , g 1 , g G ), . . . , ϕ n (S i , g G , g G−1 )).
For example, the bid B n,1 (S 1 ) = {1, 4, 2, 10, (3000, 30), (2, 4), 0} with S 1 =
{g 1 , g 2 }, g 1 = {Computation} with one attribute A g1 = {Speed}, and g 2 =
{Storage} with one attribute A g2 = {Space} expresses that a buyer n wants
to buy the bundle S 1 and has a valuation of v n (S 1 ) = 1 per slot for it. The
requested computation service g 1 should at least be capable of providing 3000
MIPS 5 , and the storage service g 2 should have at least 30 GB of available
space. The computation service g 1 can split in two parts at the most, while
storage service g 2 can run on four different machines simultaneously. Furthermore,
the two services have no coupling requirements. The buyer requires four
slots of this bundle which must be fulfilled within a time range of slots two
and ten.
The sellers’ orders are formalized in a way similar to the buyers’ orders. However,
they do not include maximum divisibility and coupling properties and
assume that the number of time slots is equal to the given time range. An
order B m is defined as a concatenated set of OR bundle bids
B m = {B m,1 (S 1 ) ∨ . . . ∨ B m,u (S j )},
where u is the number of bundle bids. A single bid B m,f (S i ) for a bundle S i
is defined as the tuple
B m,f (S i ) = (r m (S i ), e m (S i ), l m (S i ), q m (S i , g 1 , a g1 ,1), . . . , q m (S i , g j , a gj ,A j
)).
In the following it is assumed that the bid elicitation has already taken place.
4.2 Winner Determination
Based upon this notation, the multi-attribute combinatorial winner determination
problem with the objective of achieving allocative efficiency can be
represented as a linear mixed integer program.
For formalizing the model, the decision variables x n (S i ), z n,t (S i ), y m,n,t (S i ),
and d m,n,t (S i ) have to be introduced. The binary variable x n (S i ) denotes
whether bundle S i is allocated to buyer n (x n (S i ) = 1) or not (x n (S i ) = 0).
Furthermore, the binary variable z n,t (S i ) is assigned to a buyer n and is associated
in the same way as x n (S i ) with the allocation of S i in time slot t. For
a seller m, the real-valued variable y m,n,t (S i ) with 0 ≤ y m,n,t (S i ) ≤ 1 indicates
5 Million instructions per second (MIPS) is a measure for a computer’s processor
speed.
9
133

the percentage contingent of bundle S i allocated to the buyer n in time slot
t. For example, y m,n,t (S i ) = 0.5 denotes that 50 percent of the quality characteristics
of bundle S i are allocated from seller m to buyer n in time slot t. In
the previous example including the storage service with 30GB of free space,
a partial allocation of 15 GB from seller m to buyer n in time slot t would
lead to y m,n,t (Storage) = 0.5. The binary variable d m,n,t (S i ) is linked with
y m,n,t (S i ) and denotes whether the seller m allocates bundle S i to buyer n in
time slot t (d m,n,t (S i ) = 1) or not (d m,n,t (S i ) = 0).
By means of these variables, the multi-attribute winner determination model
can be formulated as follows:
max ∑ ∑ ∑
v n (S i )z n,t (S i ) − ∑ ∑ ∑ ∑
r m (S i )y m,n,t (S i ) (1)
n∈N S i ∈S t∈T
m∈M n∈N S i ∈S t∈T
s.t. ∑
S i ∈S
x n (S i ) ≤ 1, ∀n ∈ N (2)
∑
z n,t (S i ) − x n (S i )s n (S i ) = 0, ∀n ∈ N , ∀S i ∈ S (3)
t∈T
∑
y m,n,t (S i ) ≤ 1, ∀m ∈ M, ∀S i ∈ S, ∀t ∈ T (4)
n∈N
The objective function (1) maximizes the surplus V ∗ which is defined as the
difference between the sum of the buyer’s valuations v n (S i ) and the sum of the
sellers’ reservation prices r m (S i ). Assuming bidders are truthful, the objective
function reflects the goal of maximizing economic social welfare. The first
constraint (2) guarantees that each buyer n can be allocated only to one
bundle S i . This constraint is necessary to fulfill the XOR constraint of a buyer
order. Constraint (3) ensures that for any allocated bundle S i , a buyer n
receives exactly the required slots within the time set T . The constraint has
thus to formulated as an equality in to order to be binding.
For each time slot t, constraint (4) ensures that each seller cannot allocate more
than the seller possesses. The constraints (2-4) consider the basic allocation
functionality of the exchange. In designing an adequate mechanism for the
Grid, quality characteristics and dependencies between resources must also to
be addressed:
∑
z n,t (S i )q n (S i , g k , a gk ,j) − ∑
S i ∋g k
∑
S i ∋g k m∈M
y m,n,t (S i )q m (S i , g k , a gk ,j) ≤ 0,
∀n ∈ N , ∀g k ∈ G, ∀a gk ,j ∈ A gk , ∀t ∈ T (5)
10
134

∑ ∑
d m,n,t (S i ) − ∑
γ n (S i , g k )z n,t (S i ) ≤ 0,
S i ∋g k m∈M
S i ∋g k
∀n ∈ N , ∀g k ∈ G, ∀t ∈ T (6)
∑
ϕ n (S j , g k , g l )( ∑
S j ∋g k ,g l
∑
ϕ n (S j , g k , g l )( ∑
S j ∋g k ,g l
∑
S i ∋g k m∈M
S i ∋g k
d m,n,t (S i ) − ∑
S i ∋g l
d m,n,t (S i )) = 0,
∀n ∈ N , ∀m ∈ M, ∀g k , g l ∈ G, ∀t ∈ T (7)
d m,n,t (S i )+ ∑
∑
S i ∋g l m∈M
d m,n,t (S i ) − 2z n,t (S j )) ≤ 0,
∀n ∈ N , ∀g k , g l ∈ G, ∀t ∈ T (8)
Constraint (5) guarantees that for any allocated bundle in an arbitrary time
slot t, all required resources have to be fulfilled in the same slot in at least the
demanded qualities. Constraint (6) ensures that a resource will be provided
by at most γ n (S i , g k ) different suppliers. For simplicity, it is assumed that a
good g k with restricted co-allocations is not part of further XOR concatenated
bids of the buyer n.
Constraints (7) and (8) account for the coupling of two resources. Constraint
(7) ensures that two resources must be provided by the same seller, in case they
should be coupled. This constraint alone does not suffice for coupling resources
requirements, since it would be possible for two sellers to co-allocate a coupled
computation service with 3000 MIPS and a storage service with 30 GB in
different quality characteristics, e.g. to allocate a computation service with
2998 MIPS and a storage service with 1 GB from one seller, and a computation
service with 2 MIPS and a storage service with 29 GB from another. To
exclude these undesirable allocations, constraint (8) imposes the restriction
that coupled resources cannot be co-allocated. Simplifying the model, this
also includes free-disposal resources. For instance, if the computation service
with 3000 MIPS and the storage service with 30 GB are allocated from one
particular seller as a bundle, another seller cannot allocate a bundle containing
a rendering service and another storage service to the same buyer. However,
the seller may allocate any bundle without a storage and computation service
to the buyer, e.g. the rendering service alone. Furthermore, it is assumed that
coupled resources are only part of one particular bundle bid B n,f (S i ) in case
a buyer submits two XOR concatenated bids containing coupled resources.
The time restrictions of the bids are given by:
(e n (S i ) − t)z n,t (S i ) ≤ 0, ∀n ∈ N , ∀S i ∈ S, ∀t ∈ T (9)
(t − l n (S i ))z n,t (S i ) ≤ 0, ∀n ∈ N , ∀S i ∈ S, ∀t ∈ T (10)
11
135

(e m (S i ) − t) ∑ y m,n,t (S i ) ≤ 0, ∀m ∈ M, ∀S i ∈ S, ∀t ∈ T (11)
n∈N
(t − l m (S i )) ∑ y m,n,t (S i ) ≤ 0, ∀m ∈ M, ∀S i ∈ S, ∀t ∈ T (12)
n∈N
Essentially, constraints (9)-(12) indicate that slots cannot be allocated before
the earliest and after the latest time slot of either buyer (constraint (9), (10))
or seller (constraint (11), (12)).
Finally, the establishment of the relationship between the real valued decision
variable y m,n,t (S i ) and the binary variable d m,n,t (S i ) needs to be addressed and
the decision variables of the optimization problem have to be defined:
y m,n,t (S i ) − d m,n,t (S i ) ≤ 0, ∀n ∈ N , ∀m ∈ M, ∀S i ∈ S, ∀t ∈ T (13)
d m,n,t (S i ) − y m,n,t (S i ) < 1, ∀n ∈ N , ∀m ∈ M, ∀S i ∈ S, ∀t ∈ T (14)
x n (S i ) ∈ {0, 1}, ∀n ∈ N , ∀S i ∈ S (15)
z n,t (S i ) ∈ {0, 1}, ∀n ∈ N , ∀S i ∈ S, ∀t ∈ T (16)
y m,n,t (S i ) ≥ 0, ∀n ∈ N , ∀m ∈ M, ∀S i ∈ S, ∀t ∈ T (17)
d m,n,t (S i ) ∈ {0, 1}, ∀n ∈ N , ∀m ∈ M, ∀S i ∈ S, ∀t ∈ T (18)
Constraints (13) and (14) incorporate an if-then-else constraint, i.e. if a
seller m partially allocates a bundle S i to a single buyer n (y m,n,t (S i ) > 0), the
binary variable d m,n,t (S i ) has to be d m,n,t (S i ) = 1 (constraint (13)); otherwise,
it has to be d m,n,t (S i ) = 0 (constraint (14)). Finally, the constraints (15) - (18)
specify the decision variables of the optimization problem.
As multiple optimal solutions may exist, ties are broken in favour of maximizing
the number of traded bundles and then at random.
Example (Winner determination): Suppose there are two buyers n 1 , n 2 ,
two sellers m 1 , m 2 and two services g 1 = {Computation} and g 2 = {Storage}
with G = {g 1 , g 2 } each with single attributes, namely a g1 ,1 = {Speed} and
a g2 ,1 = {Size}. The buyers and sellers can submit bids on the bundles S 1 =
{g 1 }, S 2 = {g 2 }, and S 3 = {g 1 , g 2 }. The bundles can be allocated within a
time range T = {t 0 , . . . , t 4 } of T = 5 slots. Each buyer submits a set of XOR
bids (shown in table 1) and each seller a set of OR bids (see table 2).
12
136

N S i v n (S i ) q n (S i , g k , a gk ,j) e n (S i ) l n (S i ) s n (S i ) γ n (S i , g k ) ϕ n (S i , g i , g j )
n 1 S 3 2 g 1 → 500, g 2 → 15 0 4 2 g 1 , g 2 → 1
n 2
S 1 3 g 1 → 400 1 4 3
S 3 2 g 1 → 300, g 2 → 25 0 4 2 g 2 → 1
Table 1. XOR orders of the buyers.
M S i r m (S i ) q m (S i , g k , a gk ,j) e m (S i ) l m (S i )
m 1 S 3 1 g 1 → 500; g 2 → 40 0 4
m 2
S 1 2 g 1 → 1000 0 3
S 3 2 g 1 → 700; g 2 → 20 1 4
Table 2. OR orders of the sellers.
For instance, buyer n 2 submits two XOR concatenated bids to bundles S 1 and
S 3 . The buyer has a valuation of v n2 (S 1 ) = 3 for the bundle S 1 which consists
of the computation service g 1 . The service must have at least 400 GB of free
space and can be allocated between the slots e n2 (S 1 ) = 1 and l n2 (S 1 ) = 4.
The buyer requires s n2 (S 1 ) = 3 slots of the service and has no co-allocation
restrictions.
An optimal solution for the winner determination problem is an allocation
of the bundles S 3 and S 1 to the buyers n 1 and n 2 with x n1 (S 3 ) = 1 and
x n2 (S 1 ) = 1. Seller m 1 is part of the allocation with bundle S 3 and seller m 2
with bundle S 1 . The maximized value V ∗ of the winner determination problem
is V ∗ = 8.6. The corresponding schedule for this allocation is given in table 3.
M S t 0 t 1 t 2 t 3
m 1 S 3 n 1 : g 1 → 500, g 2 → 40 n 1 : g 1 → 500, g 2 → 40
m 2 S 1 n 2 : g 1 → 400 n 2 : g 1 → 400 n 2 : g 1 → 400
Table 3. Allocation schema.
Buyer n 1 receives bundle S 3 = {g 1 , g 2 } from seller m 1 in time slots 0 and
1. The bundle is allocated from one seller and as such the buyer’s requested
coupling property is satisfied. Buyer n 2 receives requested computation service
S 1 = {g 1 } in time slots 1, 2, and 3.
Although n 1 does not require the entire allocated space of the storage service
g 2 , a partial execution of the bundle is not possible due to the computation
service g 1 requirements. Bundles can only partially executed as a whole. If an
isolated partial execution of single resources in a bundle would be possible,
these single resources would also have to be valued. However, as resources
may be complementarities or substitutes, valuation and reservation prices for
a single resource of a bundle do not exist (Milgrom, 2004). As such, resources
of a bundle cannot be partially executed.
13
137

The outcome of the winner determination problem is allocative efficient as long
as buyers and sellers reveal their valuations truthfully. The incentive to set
bids according to the valuation is induced by an adequate pricing mechanism
introduced in the following subsection.
4.3 Pricing
The question how to determine payments made by participants to the exchange
and vice versa after the mechanism has determined the winners is
referred to as pricing problem. With respect to the objective of achieving an
efficient allocation, a pricing scheme based on the well-known Vickrey-Clarke-
Groves (VCG) mechanism would attain this objective (Vickrey, 1961; Clarke,
1971; Groves, 1973). Moreover, VCG mechanisms are the only allocativeefficient
and incentive compatible mechanisms (Green and Laffont, 1977).
Hence, it is logical to turn to the VCG mechanism.
The basic idea behind a VCG mechanism is to grant any participant a discount
(called Vickrey discount) on that bid. The discount reflects the impact of that
bid on the social welfare. A VCG mechanism is efficient, incentive-compatible,
and individually rational for participants with quasi linear utility functions.
However, Myerson and Satterthwaite (1983) proved that it is impossible to design
an exchange which is incentive compatible, (interim) individually rational,
and budget balanced that achieves efficiency in equilibrium. The theorem is
comprehensive and also applies to the presented mechanism.
Obviously, the VCG pricing schema cannot be applied since it runs a permanent
deficit requiring outside subsidization. The VCG mechanism can, however,
serve as a benchmark.
VCG Pricing
Let ¯N be the set of buyers and ¯M be the set of sellers who are part of
the allocation (hence, x n (S i ) = 1 with n ∈ ¯N and y m,n,t (S i ) > 0 with m ∈
¯M). The union of both sets is defined as W = ¯N ∪ ¯M, where w ∈ W is a
participant that is part of the allocation. Let V ∗ be the maximized value of the
winner determination problem and (V −w ) ∗ be the maximized value without the
participant w. Thus, the Vickrey discount for a participant w can be calculated
by ∆ V ICK,w = V ∗ − (V −w ) ∗ . With regards to the Vickrey discount ∆ V ICK,w ,
the price p V ICK,n (S i ) for a bundle S i and a buyer n can be calculated by
p V ICK,n (S i ) = v n (S i )s n (S i ) − ∆ V ICK,n , (19)
and the price p V ICK,m (S i ) for a bundle S i and a seller m by
p V ICK,m (S i ) = r m (S i ) ∑ n∈N
∑
t∈T
y m,n,t (S i ) + ∆ V ICK,m
. (20)
α
14
138

N (V −n ) ∗ ¯v n (S i ) ∆ V ICK,n p V ICK,n (S i ) M (V −m ) ∗ ¯r m (S i ) ∆ V ICK,m p V ICK,m (S i )
n 1 6.6 4 2 2 m 1 7.6 2 1 3
n 2 2 9 6.6 2.4 m 2 4.75 2.4 3.85 6.25
Table 4. Vickrey discounts and prices.
As sellers submit OR concatenated bids, they can participate in the allocation
with multiple bundles. Because the Vickrey discount refers to the seller’s overall
impact, the discount has to be portioned among all of the seller’s successful
bids. Thus the discount is divided by α, where α is the number of bundles with
which a seller m is participating in the allocation.
Applying the VCG pricing scheme to the example presented above (V ∗ = 8.6)
results in the prices p V ICK,w (S i ) and the discounts ∆ V ICK,w shown in table 4
with ¯v n (S i ) = v n (S i )s n (S i ) and ¯r m (S i ) = ∑ t∈T ,n∈N y m,n,t (S i )r m (S i ).
As stated by the Myerson-Satterthwaite impossibility theorem, a VCG mechanism
is efficient and individually rational but non budget balanced in an
exchange. Aggregating the net payments of the example leads to a negative
value with 2 + 2.4 − (3 + 6.25) = −4.85. In this case, the auctioneer has to
subsidize the exchange. Naturally, such a situation cannot be sustained for a
long period of time, making the VCG mechanism unfeasible.
Relaxing the requirement of having an efficient allocation opens up the possibility
for a second-best mechanism that is budget balanced. These ideas gave
rise to the development of a k-pricing scheme which is presented in the next
section.
K-Pricing
The underlying idea of the k-pricing scheme is to determine prices for a buyer
and a seller on the basis of the difference between their bids (Sattherthwaite
and Williams, 1993). For instance, suppose that a buyer wants to purchase a
computation service for $5 and a seller wants to sell a computation service for
at least $4. The difference between these bids is π = $1, where π is the surplus
of this transaction that can be distributed among the participants.
For a single commodity exchange, the k-pricing scheme can be formalized as
follows: let v n (S i ) = a be the valuation of a buyer n and r m (S i ) = b be the
reservation price of the buyer’s counterpart m. It is assumed that a ≥ b, i.e. the
buyer has a valuation for the commodity which is at least as high as the seller’s
reservation price. Otherwise no trade would occur. The price for a buyer n and
a seller m can be calculated by p(S i ) = ka+(1−k)b with 0 ≤ k ≤ 1. In regard
to the above-mentioned bids for the computation service g 1 , the resulting price
for both participants is p(g 1 ) = 0.5 ∗ 9 + (1 − 0.5) ∗ 2.4 = 5.7 for k = 0.5.
15
139

In the following, it is assumed that the number of buyers equals the number
of sellers. Therefore, k is set to k = 0.5 since it favors neither the buyers’ side
nor the sellers’ side.
The k-pricing schema can also be applied to a multi-attribute combinatorial
exchange: in each time slot t in which a bundle S i is allocated from one or
more sellers, the surplus generated by this allocation is distributed among
a buyer and the sellers. Suppose a buyer n receives a computation service
S 1 = {g 1 } with 1000 MIPS in time slot 4 and values this slot with v n (S 1 ) = 5.
The buyer obtains the computation service S 1 = {g 1 } by a co-allocation from
seller m 1 (400 MIPS) with a reservation price of r m1 (S 1 ) = 1 and from seller
m 2 (600 MIPS) with r m2 (S 1 ) = 2. The distributable surplus of this allocation
is π n,4 (S 1 ) = 5 − (1 + 2) = 2. Buyer n gets k ∗ π n,4 (S 1 ) of this surplus, i.e.
the price buyer n has to pay for this slot is p k,n,4 (S i ) = v(S 1 ) − kπ n,4 (S 1 ).
Furthermore, the sellers have to divide the other part of this surplus, i.e.
(1 − k)π n,4 (S 1 ). This will be done by considering each proportion a seller’s
bid has on the surplus. In the example, this proportion o m,n,t (S i ) for seller m 1
is o m1 ,n,4(S 1 ) = 1 and for seller m 3
2 is o m2 ,n,4(S 1 ) = 2 . The price a seller m
3
receives for a single slot 4 is consequently calculated as
p k,n,4 (S i ) = r m (S 1 ) + (1 − k)π n,4 (S 1 )o m,n,4 (S 1 ).
Expanding this scheme to a set of time slots results in the following formalization:
let π n,t (S i ) be the surplus for a bundle S i of a buyer n with all
corresponding sellers for a time slot t:
π n,t (S i ) = z n,t (S i )v n (S i ) − ∑
m∈M
y m,n,t (S i )r m (S i ) (21)
For the entire job (i.e. all time slots), the price for a buyer n is calculated as
p k,n (S i ) = x n (S i )v n (S i )s n (S i ) − k ∑ t∈T
π n,t (S i ). (22)
This means that the difference between the valuation for all slots (v n (S i )s n (S i ))
of the bundle S i and the k-th proportion of the sum over all time slots of the
corresponding surpluses is determined.
The price of a seller m is calculated in a similar way: First of all, the proportion
o m,n,t (S i ) of a seller m allocating a bundle S i to the buyer n in time slot t is
given by
o m,n,t (S i ) =
y m,n,t (S i )r m (S i )
∑m∈M y m,n,t (S i )r m (S i ) . (23)
Having computed π n,t (S i ) and o m,n,t (S i ), the price a seller receives for a bundle
16
140

S i is calculated as:
p k,m (S i ) = ∑ ∑
y m,n,t (S i )r m (S i ) + (1 − k) ∑ ∑
o m,n,t (S i )π n,t (S i ). (24)
n∈N t∈T
n∈N t∈T
Applying this pricing scheme to the example presented above results in the
prices given table 5 with
¯v n (S i ) = v n (S i )s n (S i ) and ¯r m (S i ) =
∑
t∈T ,n∈N
y m,n,t (S i )r m (S i ).
In this case, the exchange does not have to subsidize the participants since it
fulfills the budget balance property in a way that no payments towards the
mechanism are necessary. Hence, the k-pricing schema qualifies as a candidate
pricing schema for the Grid. Since the Myerson-Satterthwaite impossibility
theorem is strict, the implications of the pricing schema on the allocation
need to be investigated.
N ¯v n (S i ) p k,n (S i ) M ¯r m (S i ) p k,n (S i )
n 1 4 3 m 1 2 3
n 2 9 5.7 m 2 2.4 5.7
Table 5. Prices using k-Pricing with k = 0.5.
5 Evaluation
The presented auction formats are evaluated by means of a stochastic simulation.
The winner determination problem is solved using the optimization engine
CPLEX 9.1. 6 The evaluation assesses the impact of the auction schemas
on the outcome requirements defined in section 2.
5.1 Data Basis
As a data basis, a random bid stream including Decay distributed bundles
is generated. The decay function has been recommended by Sandholm (2002)
because it creates hard instances of the allocation problem. At the beginning, a
bundle consists of one random resource. Afterwards, new resources are added
randomly with a probability of α = 0.75. This procedure is iterated until
resources are no longer added or the bundle already includes the same resource.
The effects that can be obtained by the Decay distribution will be amplified.
Hence, the Decay function is used to create a benchmark for upper bounds of
the effects.
6 CPLEX is commercial product and is currently the state of the art optimization
engine.
17
141

As an order, buyers and sellers submit a set of bids on 10 different bundles,
where a bundle is Decay-distributed from 5 possible resources. Each
resource has two different attributes drawn from a uniform distribution in line
with most of the quality characteristics of computer resources described by
Kee et al. (2004). The time attributes are each uniformly distributed where
the earliest time slot has a range of [1..3], the latest time slot is in [4..6],
and the number of slots lies in [1..3]. Forty percent of the buyers’ bids have
co-allocation restrictions. Sixty percent of the resources contained in these
bundles have uniform distributed restrictions on the maximum number of coallocations.
The number of resources which have to be allocated from the same
machine (coupling) is also drawn from an uniform distribution.
The corresponding valuations and reservation prices for a bundle are drawn
from the same normal distribution, whereas the number of resources in a
bundle and the quality characteristics affect the mean and variance of the
distribution. This means the probability is greater that a buyer has a higher
valuation for a storage service with 200 GB than for a storage service with 10
GB of space. In any problem instance, new orders for buyers and sellers are
randomly generated. Subsequently, demand and supply are matched against
each other, determining the winning allocation and corresponding prices.
In total, 120 different problem instances are generated, where a problem consists
of 10 or 20 participants. The proportion of buyers and sellers is equally
selected out of {10/10, 8/12, 12/8, 5/5, 1/4, 4/1}. In each instance, the buyers
and sellers each submit one single bid.
5.2 Truthful Bidders
In the first treatment, it is assumed that bidders are truthful. The average
results of the treatment regarding the budget balance and participant utilities
are shown in table 6.
Price Mechanism Budget Utility Buyers Utility Sellers Utility
VCG −3.9425 2.6484 4.0928 6.7412
k-price 0 1.3994 1.3994 2.7988
Table 6. Average Utility and Budget using Decay distributed bids.
As the Myerson-Satterthwaite theorem suggests, the budget using a VCG auction
is always negative. The mechanism requires subsidization from outside the
system. Note that these subsidies amount to half the total utility. Obviously,
incentive compatibility is extremely expensive, as these subsidizations are required
to motivate the buyers and sellers to truthfully report their valuations
and reservation prices.
By applying a k-price auction, the budget deficit can be converted into a mech-
18
142

anism, which is independent from outside subsidies. Regarding its formal construction,
the k-price auction is budget balanced; however, the average total
utility shrinks by the amount of the budget deficit (including a small approximation
increment). The share between buyers’ and sellers’ surplus remains
unchanged. In the k-pricing schema, however, the utilities of the buyers and
sellers are significantly smaller than in the VCG auction and denote a measure
for the incentive compatibility loss.
As table 6 shows, the sellers’ utilities in a VCG auction are higher than those
of the buyers. Regarding the data basis, sellers expect to offer more available
time slots than buyers request. Buyers submit bids on a number of time slots
s n (S i ) between a time range e n (S i ) and l n (S i ). The number of requested slots
is less than the number of allocatable slots, i.e. s n (S i ) ≤ l n (S i ) − e n (S i ). A
seller, however, offers a service for the entire time range. As the time ranges
of the buyers and sellers are drawn from the same distribution, the number
of tradeable slots a seller offers is greater than the buyer’s requested slots.
The greater the number of tradeable slots, the higher the impact of the bid in
the VCG mechanism. As such, the sellers’ average discounts are higher than
those of the buyers; therefore the sellers have a higher average utility than the
buyers.
Because the simulation assumes truthful bidding, this rather naive analysis
is only accurate for the VCG mechanism. If the pricing rule is changed – for
example from the VCG mechanism to the k-price auctions – strategic bidding
behavior may change since the k-price auction is not incentive compatible.
5.3 Manipulating Bidders
In a second treatment, it is assumed that buyers and sellers using the k-pricing
schema misrepresent their true valuation or reservation price respectively. Only
simple misrepresentations by β% are considered, where λ% of the buyers and
sellers reduce or increase their reported values by β%. Instead of observing only
symmetric Nash-equilibria as in the analysis by Parkes et al. (2001), where
buyers and sellers either misrepresent their valuation or reservation price by 0
or by β%, the ratio of misrepresenting buyers and sellers to the total number
of participants is also varied. A ratio of β = 50%, for instance, denotes that
50% of the buyers and sellers misrepresent their valuations and reservation
prices by λ%, while the other 50% report truthfully.
By exploring the joint strategy space (i.e. varying the share of misrepresentative
and truthful participants as well as the percentage of misrepresentation)
the average efficiency loss for the k-pricing schema can be analyzed: Let ¯n ∈ ¯N
be the buyers and ¯m ∈ ¯M be the sellers who are part of the allocation in a
truthful bidding scenario, i.e. λ = β = 0. Furthermore, let ˆn ∈ ˆN be the buyers
and ˆm ∈ ˆM be the sellers of the allocation having manipulating bidders
19
143

(λ ≥ 0, β ≥ 0). Subsequently, the efficiency loss EL can be calculated as the
difference between the true preferences of the participants who are part of the
allocation in both the truthful and manipulating scenarios:
∑
EL = s¯n (S i )v¯n (S i ) +
∑¯n∈ ∑ ∑ ∑ ∑
y ¯m,¯n,t (S i )r ¯m (S i )−
¯N S i ∈S
¯m∈ ¯M ¯n∈ ¯N S i ∈S t∈T
∑
( sˆn (S i )vˆn (S i ) +
∑ˆn∈ ∑ ∑ ∑ ∑
y ˆm,ˆn,t (S i )r ˆm (S i )).
ˆN S i ∈S
ˆm∈ ˆM ˆn∈ ˆN S i ∈S t∈T
Figure 1 shows the percentage efficiency loss and the percentage decrease
of the number of successful bidders with a varying number of manipulating
participants as a function of the manipulation factor. For instance, if β = 40%
of the participants manipulate their valuations and reservation prices by λ =
6%, the resulting efficiency loss is 15%. It is apparent that the efficiency loss
is higher when more participants use manipulation and thereby increase the
manipulation factor. With a manipulation factor greater than λ = 25%, the
efficiency curves stagnate. This stagnation results from the fact that either
non-manipulating bidders are the only participants of the allocation (in case
of β < 100%) or no one is allocated further (in case of β = 100%). The
valuation and reservation prices with a manipulation factor λ > 25% are so
low – respectively so high – that no bid from the manipulating participants is
allocated. This means that no manipulated offer for a bundle can be matched
with any manipulated request. This effect is supported by the percentage
decrease of the number of successful bidders by manipulating participants. For
instance, if β = 20% of the participants manipulate their bids by λ = 21%,
the number of allocated bids is 17% smaller than in the non-manipulating
scenario. Furthermore, if all participants are manipulating by λ = 30%, the
size decreases by 100%, i.e. the set of allocated participants is empty. The
curves recurrently stagnate with a manipulation factor greater than λ = 25%.
Effiency loss
100
90
80
70
60
50
40
30
20
10
0
0
0,1
0,2
0,4
0,6
0,7
0,8
1
0 0,03 0,06 0,09 0,12 0,15 0,18 0,21 0,24 0,27 0,3
Manipulation factor
Decrease of successful participants
100
90
80
70
60
50
40
30
20
10
0
0
0,1
0,2
0,4
0,6
0,7
0,8
1
0 0,03 0,06 0,09 0,12 0,15 0,18 0,21 0,24 0,27 0,3
Manipulation factor
Fig. 1. Efficiency loss (left) and decrease of the number of successful bidders (right)
by manipulating participants.
In another analysis, the aggregated average utility gain of manipulating partic-
20
144

ipants is measured. This gives information on whether or not the total utility
of participants can be improved through manipulation. The utility which participants
can gain by manipulation only depends on the prices they have to
pay and is thus calculated as
UG =
∑ ∑ˆn∈ ˆN S i ∈S
ˆpˆn (S i ) + ∑
∑
ˆm∈ ˆM S i ∈S
ˆp ˆm (S i ) − (
∑¯n∈ ¯N
∑
S i ∈S
p¯n (S i ) + ∑
∑
¯m∈ ¯M S i ∈S
p ¯m (S i )),
where ˆN and ˆM are the sets of buyers and sellers who are part of the allocation
with manipulating participants and ˆpˆn (S i ) and ˆp ˆm (S i ) are the corresponding
prices. Sets ¯N and ¯M denote the participants of the allocation in the nonmanipulated
scenario with p¯n (S i ) and p ¯m (S i ) as their prices.
The percentage utility gain of manipulating participants in the k-price auction
is shown in figure 2. The total utility of the participants can be increased
by manipulation; at the highest point where all participants manipulate by
λ = 6%, the utility gain is more than 26% compared to the non-manipulating
scenario. Moreover, small manipulation factors between λ = 1% and λ = 9%
always result in a positive utility gain. However, participants have a negative
utility gain when the manipulation factor is greater than λ = 13%. In the worst
scenario, the utility loss is 100% if all participants manipulate by λ = 30%.
Utility gain
40
20
0
-20
-40
-60
-80
-100
0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1
Participants manipulating
0,3
0,21
0,12
0,03
20-40
0-20
-20-0
-40--20
-60--40
-80--60
-100--
80
Manipulation
factor
Fig. 2. Utility gain using k-pricing.
In summary, the simulation has shown it reasonable to believe that participants
will not strongly deviate from their truth valuations and reservation
prices. Although participants’ average utility gain can be improved through
manipulation, the participants increasingly risk not being executed in the
auction (cf. figure 1). This risk actually increases the more participants use
manipulation. The simulation result suggests that the k-pricing schema has
accurate incentive properties resulting in fairly mild allocative efficiency losses.
As such, the pricing schema is a practical alternative to the VCG mechanism
and is highly relevant for an application in the Grid.
21
145

5.4 Computational Tractability
As denoted in section 2, a computational tractable outcome determination
is required. However, the multi-attribute combinatorial winner determination
problem (MWDP) is N P-complete: consider the special case of the MWDP
with multiple buyers, one seller, no quality attributes, one time slot, and
no coupling and maximal divisibility restrictions. In this case, the problem
is equivalent to the combinatorial allocation problem (CAP) formulated by
Rothkopf et al. (1998). CAP is equivalent to the set-packing problem (SPP)
on hypergraphs (Rothkopf et al., 1998) which is known to be N P-complete
(Karp, 1972). Since CAP can be reduced to the MWDP, the MWDP is also
N P-complete.
Due to the N P-completeness of the winner determination problem, the auction
schema is computationally intractable in large-scaled scenarios. For settings
with a limited number of participants, however, the auction schema
seems practicable. Ideally, an upper boundary of participants can be defined
for which the problem is still computationally tractable within a meaningful
timeframe. In the context of trading services, a meaningful timeframe is
consistent with the maximum time limit that an allocation process may last.
Experiences suggest that an allocation process that is shorter than 4 minutes
is adequate.
Preliminary studies with the auction schema confirmed that solving the winner
determination problem can become computationally intractable, even in
settings with less than 100 participants. This implies that exact solutions of
the winner determination problem may require too much computation time
and cannot be directly applied for the Grid market. To remedy this obstacle,
we have to rely on approximations. A common way is to employ algorithms
that compute solutions of the winner determination problem within a threshold
r away from optimality, say 2%. Such deviations from optimal solutions are
often conceived to be tolerable. However, in some cases even finding solutions
within the specified threshold of 2% becomes impossible within a meaningful
timeframe. In those cases the approximations need to have a termination
condition that forces the algorithm to return the best solution found within
a predefined time range. Combinatorial auction experiments have shown that
these so-called anytime algorithms perform fairly well. For instance, Andersson
et al. (2000) and Sandholm et al. (2005) report that feasible and approximate
solutions can be determined in a fraction of time that would be required
to find and prove an optimal solution.
The applied standard solver CPLEX provides these functionalities. For our
computational tractability analysis, we perform a series of simulations with
anytime approximations for computing the solutions. The 4 minutes introduced
above will serve as maximum timeframe. To compare the results, we
22
146

Participants
Time Limit 2min Time Limit 4min No Time Limit
µ 2 σ 2 CV 2 µ 4 σ 4 CV 4 µ σ CV
100 11.6s 25.34s 2.18 30.9s 103.28s 3.34 38.25s 143.89s 3.76
200 67.6s 50.09s 0.74 194.9s 178.78s 0.92 407.8s 598.4s 1.47
300 95.4s 42.74s 0.45 307.4s 184.118s 0.6 673.6s 809.74s 1.2
400 110.2s 24.06s 0.22 372.55s 165.07s 0.44 1536.8s 2117.91s 1.38
500 116.58s 13.73s 0.12 374.2s 203.72s 0.54 2146.7s 2593.21s 1.21
600 120s 0s 0 416.06 108.1s 0.26 3786.4s 3046.93s 0.8
Table 7. CPU time required to solve the problem.
also employ a simulation setting with a more optimistic limit of 2 minutes.
As the anytime algorithms diverges from the theoretical optimum, we need a
measure that captures the welfare loss imputed to the approximation. Nevertheless,
computing the theoretical optimum may take very long, so that we use
benchmark solution that lies within a 2% gap from the optimum. If CPLEX
cannot find any feasible solution within the specified timeframe, the problem
instance is rated as intractable. For instance, if no feasible solution is found
within 2 minutes, the problem is measured as an empty allocation with welfare
of 0 and runtime of 2 minutes. This way, we include those hard instances in
our analysis.
In order to perform the simulation, nearly the same random order flow is
used as described above. However, the time attributes are widened in order to
generate harder and more realistic settings. The earliest time slots are within
a range of [1..5], the latest time slots within [6..10], and the number of slots
within a range of [1..5]. In the simulation, the proportions of buyers and sellers
are equal and the number of participants is varied. In total, 20 treatments are
computed and the results are averaged.
Table 7 summarizes the runtime results of the simulations with 2 minutes
time limit, 4 minutes time limit, and benchmark solution with no time limit. 7
The table shows the mean µ, the standard deviation σ, and the coefficient of
variation CV = σ/µ of the CPU times required to solve the winner determination
problems. For example, with 200 participants the average runtime with
2 minutes time limit is µ 2 = 67.6 seconds, with 4 minutes limit µ 4 = 194.9
seconds, and without a time limit µ = 407.8 seconds. Note that the differences
in the runtime of settings with equal number of participants are caused
by the underlying time limits. For 200 participants, several instances require
more than 2 minutes, some even more than 4 minutes to find a solution within
the specified gap. As CPLEX terminates after the predefined timeframe, the
7 The analysis was performed on a Pentium XEON with 3.2GHz and 2GB ram
using CPLEX 9.1 with a MIP gap of 0.02.
23
147

Participants
Time Limit 2min
Time Limit 4min
µ 2 σ 2 CV 2 µ 4 σ 4 CV 4
100 0.09% 0.33% 3.71 0.03% 0.24% 8.01
200 0.27% 0.57% 2.11 0.06% 0.23% 3.67
300 0.86% 0.94% 1.10 0.1% 0.36% 3.59
400 1.85% 3.71% 2 0.31% 0.35% 1.13
500 11.96% 31.14% 2.6 0.55% 0.74% 1.33
600 21.59% 39.27% 1.82 13.79% 33.82% 2.45
Table 8. Percentage Welfare Loss
average runtime in the 2 minutes setting is less than the runtime required for
finding the benchmark solution.
With an increasing number of participants, CPLEX’s average execution time
grows rapidly. This is most obvious in the benchmark settings with no time
limit. For 500 participants, CPLEX requires nearly 2 hours to find a solution
for some instances. In some settings with 600 participants, CPLEX even fails in
finding a benchmark solution within 2 hours. In those cases, the simulation is
stopped and the obtained results are neglected. More precisely, in our analysis
this status occurred in 8 out of 20 instances. As these cases do not produce
benchmark solutions, the obtained results with 600 participants may only have
marginal significance.
Overall, the runtime analysis shows that the winner determination problem is
computationally intractable without the introduction of time limits. Apparently,
delimiting the approximation produces solutions in a meaningful time,
i.e. less than 4 minutes.
In settings with limited time, however, the quality of the (suboptimal) solutions
and accordingly its impact on the welfare changes. In the following the
welfare effect due to the approximations is studied in detail. For this, we define
a welfare loss higher than 5% as unacceptable for our settings, as we already
have approximate solutions via the applied 2% gap.
Table 8 summarizes the welfare loss of the time limited scenarios compared
to the benchmark solution. For instance, for 300 participants the average welfare
loss in the 2 minutes setting is µ 2 = 0.86% and in the 4 minutes limit
µ 4 = 0.1%. With an increasing number of participants, the welfare loss becomes
higher. On the one hand, this is reasoned by the fact that the problem
instances become harder to solve within the given time frame. On the other
hand, feasible solutions cannot always be found in time which results in a welfare
loss for these infeasible instances of 100%. This explains the big increase
24
148

of the welfare loss from 400 to 500 participants in the 2 minutes settings –
respectively, from 500 to 600 participants in the 4 minutes settings. Regarding
the simulations with a limit of 2 minutes, the welfare loss exceeds our
5% threshold for more than 400 participants. With a 4 minutes limit, this
threshold is exceeded in settings with more than 500 participants.
The analysis of the welfare loss shows, that the anytime algorithm achieves
fairly efficient results: With a 2 minutes time limit, the welfare loss is acceptable
for scenarios with up to 400 participants. For settings with up to 500
participants, a time limit of 4 minutes entails reasonable results. However, the
applied time limits become useless for settings with more than 500 participants.
For such cases, the use of other approximations has to be considered.
To anchor the problem to real Grid applications and to provide insights to
practitioners, we studied the characteristics of PlanetLab 8 and compared
them to our computational results. Essentially, PlanetLab is a testbed for networking
and distributed computing among several scientific institutes worldwide.
The network is state-of-the-art and currently the largest infrastructure
for sharing computational Grid services. The user information was obtained
using the data provided by the All-Pairs-Pings project 9 . We collected and analyzed
the available data from November 2005 until March 2006. As a result,
PlanetLab has currently µ = 170.89 active machines on average which can be
used to perform computational jobs (with a standard deviation of σ = 21.33).
The coefficient of variation (CV) for the values CV = σ/µ = 0.12 shows that
the variability in reference to the size of the mean is low. Thus, the determined
average constitutes to be a stable value.
Transferring the PlanetLab characteristics to the proposed auction schema
implies 170 sellers on average. Assuming an equal number of buyers and sellers,
the average number of participants in the PlanetLab scenario is 340. With
respect to the above presented results, the application of a 2 minutes time
limit is sufficient to fulfill PlanetLab resource bids in a meaningful time. In
cases the boundary raises up to 500 participants, a time limit of 4 minutes
is preferable. For larger settings, however, alternative approaches have to be
evaluated.
6 Conclusion
The paper at hand proposes the derivation of a multi-attribute combinatorial
exchange for allocating and scheduling services in the Grid. In contrast to
other approaches, the proposed mechanism accounts for the variety of services
8 See http://www.planet-lab.org/ for details.
9 Every 15 minutes, all registered nodes are contacted to check their availability
status. See http://ping.ececs.uc.edu/ping/ for details.
25
149

y incorporating time and quality as well as coupling constraints. The mechanism
provides buyers and sellers with a rich bidding language, allowing for the
formulation of bundles expressing either substitutabilities or complementarities.
The winner determination problem maximizes the social welfare of this
combinatorial allocation problem. The winner determination scheme alone,
however, is insufficient to guarantee an efficient allocation of the services. The
pricing scheme must be constructed in a way that motivates buyers and sellers
to reveal their true valuations and reservation prices. This is problematic in
the case of combinatorial exchanges, since the only efficient pricing schedule,
the VCG mechanism, is not budget-balanced and must be subsidized from
outside the mechanism.
This paper develops a new pricing family for a combinatorial exchange, namely
the k-pricing rule. In essence, the k-pricing rule determines the price such that
the resulting surpluses to the buyers and sellers divide the entire surplus being
accrued by the trade according to the ration k. The k-pricing rule is budgetbalanced
but cannot retain the efficiency property of the VCG payments.
As the simulation illustrates, the k-pricing rule does not rigorously punish
inaccurate valuation and reserve price reporting. Buyers and sellers sometimes
increase their individual utility by cheating. This possibility, however, is only
limited to mild misreporting and a small number of strategic buyers and sellers.
If the number of misreporting participants increases, the risk of not being
executed in the auction rises dramatically. As a result, the k-pricing schema
is a practical alternative to the VCG mechanism and highly relevant for an
application in the Grid. The runtime analysis shows that the auction schema is
computationally very demanding. However, the use of approximated solutions
achieves adequate runtime results and fairly mild welfare losses for up to 500
participants. Comparing these results with an existing Grid testbed evinces
the practical applicability of the proposed auction.
This paper is a step towards understanding the effect and strength of different
multi-attribute combinatorial exchange mechanisms on efficiency. Contributions
include detailing a realistic market definition for Grid services, deriving
the domain-specific requirements from the Grid, defining a winner determination
scheme that meets these requirements, exploring adequate pricing schemes
that contain the right incentives, and evaluating the market mechanism by
means of a simulation.
Future research needs to consider alternative heuristics that simplify the winner
determination problem. Nature-inspired algorithms such as genetic algorithms
may be adequate for an application in the service market. A further
step towards a functioning Open Grid Market would be to confront the mechanism
with real data in a pilot run.
26
150

Acknowledgements
This work has partially been funded by EU IST progamme under grant 003769
“CATNETS”. The authors are grateful to the anonymous reviewers for the
constructive comments and suggestions that significantly improved the paper.
References
Andersson, A., Tenhunen, M., Ygge, F., 2000. Integer Programming for Combinatorial
Auction Winner Determination. In: ICMAS ’00: Proceedings of
the Fourth International Conference on MultiAgent Systems. IEEE Computer
Society, Washington, DC, USA, pp. 39–46.
Bapna, R., Das, S., Garfinkel, R., Stallaert, J., 2005. A Market Design for
Grid Computing. Tech. rep., Department of Operations and Information
Management, University of Connecticut.
Buyya, R., Abramson, D., Giddy, J., Stockinger, H., 2002. Economic Models
for Resource Management and Scheduling in Grid Computing. The Journal
of Concurrency and Computation: Practice and Experience 14 (13-15),
1507–1542.
Clarke, E., 1971. Multipart pricing of Public Goods. Public Choice 2, 19–33.
Eymann, T., Reinicke, M., Ardaiz, O., Artigas, P., de Cerio, L. D., Freitag,
F., Messeguer, R., L. Navarro, D. R., 2003. Decentralized vs. Centralized
Economic Coordination of Resource Allocation in Grids. In: Proceedings of
the 1st European Across Grids Conference.
Foster, I., Kesselman, C., 2004. The Grid 2 - Blueprint for a New Computing
Infrastructure. Vol. 2. Elsevier.
Foster, I., Kesselman, C., Nick, J., Tuecke, S., 2002. Grid Services for Distributed
System Integration. IEEE Computer 35 (6), 37–46.
Green, J., Laffont, J.-J., 1977. Characterization of Satisfactory Mechanisms
for the Revelation of Preferences for Public Goods. Econometrica 45 (2),
427–438.
Groves, T., 1973. Incentives in teams. Econometrica 41, 617–631.
Jackson, M. O., 2002. Mechanism Theory. In: Encyclopedia of Life Support
Systems. Eolss Publishers, Oxford ,UK.
Karp, R. M., 1972. Reducibility among Combinatorial Problems. In: Miller,
R. E., Thatcher, J. W. (Eds.), Complexity of Computer Computations.
Plenum Press, pp. 85–103.
Kee, Y.-S., Casanova, H., Chien, A., 2004. Realistic Modeling and Synthesis
of Resources for Computational Grids. In: ACM Conference on High
Performance Computing and Networking (SC2004). ACM Press.
Lai, K., 2005. Markets are Dead, Long Live Markets. Tech. Rep. 0502027, HP
Labs.
Ludwig, H., Dan, A., Kearney, B., 2004. Cremona: An Architecture and Library
for Creation and Monitoring of WS-Agreements. In: Proceedings of
27
151

the 2nd International Conference on Service Oriented Computing (ICSOC
2004). pp. 65–74.
Milgrom, P., 2004. Putting Auction Theory to Work. Cambridge University
Press.
Myerson, R. B., Satterthwaite, M. A., 1983. Efficient Mechanisms for bilateral
Trading. Journal of Economic Theory 28, 265–281.
Neumann, D., 2004. Market Engineering – A Structured Design Process for
Electronic Markets. Ph.D. thesis, Economics and Business Engineering.
Karlsruhe, Germany, University of Karlsruhe (TH).
Ng, C., Buonadonna, P., Chun, B. N., Snoeren, A. C., Vahdat, A., 2005. Addressing
Strategic Behavior in a Deployed Microeconomic Resource Allocator.
In: In Proceedings of the 3rd Workshop on Economics of Peer-to-Peer
Systems.
Parkes, D. C., Kalagnanam, J., Eso, M., 2001. Achieving Budget-Balance with
Vickrey-Based Payment Schemes in Exchanges. In: Proceedings of the Seventeenth
International Joint Conference on Artificial Intelligence. pp. 1161–
1168.
Regev, O., Nisan, N., 2000. The POPCORN Market: Online Markets for Computational
Resources. Decision Support Systems 28 (1-2), 177–189.
Rothkopf, M., Pekec, A., Harstad, R., 1998. Computationally Manageable
Combinational Auctions. Management Science 44, 1131–1147.
Sandholm, T., 2002. Algorithm for Optimal Winner Determination in Combinatorial
Auctions. Artificial Intelligence 135 (1–2), 1–54.
Sandholm, T., Suri, S., Gilpin, A., Levine, D., 2005. CABOB: A Fast Optimal
Algorithm for Winner Determination in Combinatorial Auctions. Management
Science 51 (3), 374–390.
Sattherthwaite, M. A., Williams, S. R., 1993. The Bayesian Theory of the
k-Double Auction. In: Friedman, D., J.Rust (Eds.), The Double Auction
Market - Institutions, Theories, and Evidence. Addison-Wesley, Ch. 4, pp.
99–123.
Schnizler, B., Neumann, D., Veit, D., Weinhardt, C., 2005. A Multiattribute
Combinatorial Exchange for Trading Grid Resources. In: Tan, Y. (Ed.), Proceedings
of the 12th Research Symposium on Emerging Electronic Markets
(RSEEM), Amsterdam, Netherlands. pp. 220–240.
Shneidman, J., Ng, C., Parkes, D. C., AuYoung, A., Snoeren, A. C., Vahdat,
A., Chun., B., 2005. Why Markets Could (But Don’t Currently) Solve Resource
Allocation Problems in Systems. In: Proceedings of Tenth Workshop
on Hot Topics in Operating Systems.
Subramoniam, K., Maheswaran, M., Toulouse, M., 2002. Towards a Micro-
Economic Model for Resource Allocation in Grid Computing Systems. In:
Proceedings of the 2002 IEEE Canadian Conference on Electrical & Computer
Engineering.
Vickrey, W., 1961. Counter Speculation, Auctions, and Competitive Sealed
Tenders. Journal of Finance 16, 8–37.
Waldspurger, C., Hogg, T., Huberman, B., Kephart, J., Stornetta, W., 1992.
28
152

Spawn: A Distributed Computational Economy. IEEE Transactions on Software
Engineering 18 (2), 103–117.
Wellman, M. P., Walsh, W. E., Wurman, P. R., MacKie-Mason, J. K., 2001.
Auction Protocols for Decentralized Scheduling. Games and Economic Behavior
35, 271–303.
Wolski, R., Plank, J., Brevik, J., Bryan, T., 2001. Analyzing Market-Based
Resource Allocation Strategies for the Computational Grid. International
Journal of High Performance Computing Applications 15 (3), 258–281.
29
153

A Discriminatory Pay-as-Bid Mechanism for Efficient Scheduling in the Sun N1
Grid Engine
Jochen Stößer † , Philipp Bodenbenner † , Simon See ‡ , Dirk Neumann †
† Institute of Information Systems and Management
Universität Karlsruhe (TH)
Englerstr. 14, 76131 Karlsruhe, Germany
lastname@iism.uni-karlsruhe.de
‡ Asia Pacific Sciene & Technology Center
Sun Microsystems, Inc.
50 Nanyang Avenue, Singapore 639798
simon@apstc.sun.com.sg
Abstract
Grid computing is a promising concept to increase the
efficiency of existing computing systems and to cut down on
IT expenses by allowing the dynamic access to computer
resources across geographical and organizational boundaries.
These inter-organizational settings require a scheduling
strategy for flexibly and efficiently matching resource requests
to idle resources. Market-based mechanisms promise
a good fit to grids’ strategic and dynamic nature by allowing
resource requesters to express valuations in addition
to technical metrics. The contribution of this paper
is twofold: We present a discriminatory pay-as-bid market
mechanism by Sanghavi and Hajek [14] and analytically
show that it outperforms market-based proportional share –
the currently most prominent grid market mechanism – with
respect to both provider’s surplus and allocative efficiency.
We further illustrate that this mechanism is not a purely theoretical
construct but that it can be integrated into the Sun
N1 Grid Engine, a state-of-the-art grid scheduler.
1 Introduction
Grid computing denotes a computing model that distributes
processing across an administratively and locally
dispersed infrastructure. By connecting many heterogeneous
computing resources, virtual computer architectures
are created, increasing the utilization of otherwise idle resources
[7]. A recent report by The Insight Research Corporation
projects an increase in worldwide grid spending
from $1.84 billion in 2006 to $24.52 billion in 2011 [1].
The Enabling Grids for E-sciencE (EGEE) project is an
intriguing example for the value of grid technology in science.
EGEE aims at developing a grid infrastructure for
more than 240 scientific institutions in 45 countries. The
EGEE grid currently consists of more than 36,000 CPUs
and 5 Petabytes of storage [6].
The business case for grids is illustrated by the example
of Synopsys [19]. Synposys is a world leader in integrated
circuit (IC) design and requires massive computer resources
for its computer-aided design processes, such as regression
testing. In the past, each division within Synopsys maintained
its own, separate computing cluster. The full computing
power of these systems, however, was only needed
in rare occasions to accommodate peak loads on the system.
Obviously, this mode of operation thus led to tremendous
inefficiencies. Synopsys leveraged Sun Microsystem’s
grid technology to connect computing resources across divisional
boundaries, thus creating a virtual pool of computing
resources which can be dynamically accessed on demand.
In consequence, the runtime of regression tests could be reduced
from about 12 hours to 2 hours.
While the grid scenario is closely related to older allocation
schemes for computer resources, such as mainframe
allocation, it is somewhat more general because the grid
resources might be owned by different organizations. In
such inter-organizational settings, scheduling of resource
requests becomes a key challenge. What is needed is a set
of mechanisms that enable users to discover, negotiate, and
pay for the use of grid resources. Classic technical scheduling
mechanisms such as first-come-first-serve or fair share
are solely built on system-centric measures. Consequently,
they do not take into account the strategic and dynamic situation
in grids:
• Scarce resources: By nature, the concept of grids is all
about sharing scarce resources. Excess demand has to
be distributed to these resources so as to maximize the
value provided by the grid system to its users.
• Decentralized control: The scarce resources are spread
across organizational boundaries. There is no central-
154

ized and complete knowledge about the state and the
availability of these resources but the system depends
on these organizations to report and act in a goodnatured
manner so as to be able to realize this value.
• Self-interested agents: The resource requesters and the
organizations contributing their idle resources to a grid
will generally try to selfishly maximize their individual
benefit from participating in the system.
In the light of these characteristics, market-based mechanisms
are deemed promising to provide a better fit to
grids’ strategic and dynamic nature by allowing resource requesters
to express valuations in addition to technical metrics.
Ultimately, prices are formed which help to balance
the dynamic demand and supply in grids. The system can
thus induce resource requesters to report truthfully and to
distribute excess demand over time. To this end, the contribution
of this paper is twofold:
• Mechanism design: We present a discriminatory payas-bid
market mechanism by Sanghavi and Hajek [14]
and analytically derive conditions under which it outperforms
market-based proportional share – the currently
most prominent grid market mechanism – with
respect to both provider’s surplus (Proposition 1) and
allocative efficiency (Proposition 2).
• Integration into Sun N1GE: We show that this mechanism
is not a purely theoretical construct but that it
can be integrated into state-of-the-art grid schedulers
to economically enrich the current allocation logics.
We illustrate the basic design considerations for the
case of the N1 Grid Engine (N1GE), the scheduler of
Sun Microsystem’s grid platform.
This paper is structured as follows. In Section 2, we discuss
related work in the field of grid scheduling, before we
introduce the pay-as-bid mechanism in Section 3. We provide
an in-depth analysis of this mechanism and compare it
to market-based proportional share. We present two design
options for how to integrate this mechanism in the Sun N1
Grid Engine in Section 4. We subsequently propose extensions
to the basic mechanism and discuss problems which
may emerge in this context. Section 5 concludes the paper
and points to future work.
2 Related Work
Current resource allocation schemes in grids can be
distinguished into technical and market-based schedulers.
Technical schedulers are based solely on system-centric
measures and aim at maximizing resource utilization and/or
balancing the system load. In contrast, market-based schedulers
introduce economic principles to grids in order to
maximize the economic value provided by such systems;
auctions or negotiations explicitly involve the users in the
allocation process. Such market mechanisms must be carefully
tailored towards the pecularities of the application environment
and the trading object. Consequently, a suite of
mechanisms has been proposed for a range of grid application
scenarios (see [13] and [20] for surveys).
In [15], [3] and [2], the scheduling problem in grids is
formalized as an NP-hard periodic combinatorial allocation
problem. In [3] and [17], heuristics are developed to migitigate
this computational complexity. While these mechanisms
account for dependencies between multiple grid resources
(e.g. CPU, memory and bandwidth), they rely on
strong technical information assumptions, such as knowledge
about the time constraints and the resource requirements
of applications.
A fundamentally different approach are mechanisms
which almost continuously assign resource shares to applications
based on one-dimensional input only, e.g. single
values which represent the users’ valuations. Market-based
proportional share is the currently most prominent proxy of
such mechanisms [5, 9, 16]. With an allocation rule purely
based on economic reasoning, e.g. the prominent Vickrey
mechanism, all available resources would be given to the
resource request with the highest valuation. From a technical
viewpoint, however, avoiding starvation 1 may be an
important consideration. Combining the economic and the
technical viewpoint, it may be desirable to give “better” service
to high-value processes but to also give at least “some”
service to low-value processes in order to avoid starvation.
In the remainder of this section, the Sun N1 Grid Engine
(N1GE), a state-of-the-art technical scheduler, and marketbased
proportional share are presented in more detail.
2.1 Sun N1 Grid Engine
The N1GE is a distributed resource management and
scheduling system developed by Sun Microsystems [18].
Being an extension of the Solaris operating system, it administers
and dynamically allocates the shared pool of heterogeneous
resources such as computing power, memory
and licensed software within an organization. The usage of
these resources is managed so as to best achieve the goals of
the organization, such as productivity, timeliness and level
of service. The N1GE has been employed for setting up
grids comprising a size of around 500-2,000 CPUs.
The N1GE scheduler consists of a waiting queue with
pending jobs and a technical scheduler which subsequently
assigns waiting jobs to idle resources (cf. Figure 1). The
user submits a job together with a specification of the tech-
1 In scheduling theory, starvation denotes the fact that low-priority processes
are prevented from doing any progress because all resources are
assigned to other higher-value processes.
155

their reported valuations’ fraction of the overall reported
valuation across all resource requesters: User i with reported
valuation w i will receive a fraction of of
w i Pn
j=1 wj
the available resource when a group of n users is competing
for resource access. Systems using proportional share as
allocation scheme have been proposed in [5], [9] and [16].
The problem with market-based proportional share is that
it remains allocatively inefficient which we will illustrate in
Section 3.3 by means of a numerical example.
In the following, an alternative mechanism is explored –
a so-called “discriminatory pay-as-bid mechanism” [14] –
that may improve on these present mechanisms.
Figure 1: Scheduling process in the N1GE
nical requirements of the job. After receiving the job requests,
the scheduler places the jobs in the waiting list of
pending jobs. The position of a job in the waiting list is
determined by the job’s priority. This priority value is calculated
by the scheduler using a pre-defined and static mix
of different policies. A sample policy mix may comprise
manually (by the administrator) set shares for individual
users, user groups, a department or a project (also called
Entitlement policy), an increase of priority for jobs which
will reach their deadline soon or that have been waiting for a
long time (Urgency policy). Additionally, users may be able
to sort their own jobs according (Custom policy, POSIX)
[4].
An example policy mix can look like this
P mix = W e ∗ P e + W u ∗ P u + W c ∗ P c
where P mix is the dispatch priority, P e is the normalized
entitlement priority (on an interval between 0 and 1) and W e
is the entitlement weighting factor. P u , W u , P c and W c are
defined accordingly for the urgency and custom priorities.
The key drawback of technical schedulers is that static
priorities are manually set up and thus do not reflect the
fluctuating demand in the system, thus leading to inefficient
allocations from an economic viewpoint. To alleviate
this problem and to increase efficiency, proportional share
mechanisms have been introduced to grid systems.
2.2 Proportional Share
Proportional share allows for resource distribution with
shares of unequal size for different users accounting for
varying importance among them. Whereas scheduling according
to pre-set, fixed shares for different users remains
technical, market-based proportional share mechanisms dynamically
base the resource share on the users’ reported
valuations, their “bids”. The total amount of available resources
is distributed among the requesters according to
3 A Discriminatory Pay-as-Bid Mechanism
Let vector w = (w 1 , . . . , w n ) represent the positive
bids of the users. The users receive shares x =
(x 1 , . . . , x n ), x j ∈ R 0 ∑ n
+,
j=1 x j = 1, of the perfectly
divisible good. These shares are calculated according to the
pre-specified allocation mechanism τ. Thus, x i = τ i (w) is
the quantity user i is allocated for a given bid vector w.
As is common in mechanism design, the users are assumed
to have quasi-linear utility functions: U i (x) =
v i x i − c(x i ), with v i ∈ R + and linear price function
c(x i ) = p i x i where p i is user i’s unit price, i.e. the price
user i would have to pay if she got the whole resource unit
(x i = 1). Let U P (x) be the provider’s utility function.
For evaluating and comparing market mechanisms, we
first need to define the user behavior, i.e. the users’ reporting
of w, and a metric. For the former, we choose the
widely used solution concept of Nash equilibria. In a Nash
equilibrium w NE , no user i can benefit by unilaterally deviating
from wi
NE . We interpret Nash equilibria as the final
outcome of an iterative process. After each stage, the requesters
can adjust their bids based on feedback about the
other requesters’ bids. Ultimately, we assume that every
user knows the vector v = (v 1 , . . . , v n ) consisting of all
users’ valuations.
A common metric for a mechanism’s performance is
its performance ratio in its Nash equilibrium. The performance
ratio of mechanism τ is defined as
∑
U(τ(w NE ))
i
U ∗ =
U i(τ(w NE )) + U P (τ(w NE ))
U ∗ ,
i.e. the worst-case ratio of the social welfare generated by
the specific mechanism if all bidders play their Nash strategy
wi NE divided by the theoretical optimum U ∗ .
A key issue when considering the use and (probably
more importantly) the usability of markets is the question
of how users come up with their valuation functions and interact
with the market, i.e. express their valuations in order
to eventually arrive at some sort of equilibrium, such as a
Nash equilibrium. Human users may be released from the
156

urden of having to issue requests and offers manually. Instead,
software agents may serve so as to hide the grids’
complexity from human users by automatically trading resources
based on the current resource consumption of applications
and configurable bidding rules which automatically
derive corresponding valuations [10, 12].
From a mechanism design perspective, we are looking
for a mechanism with maximum performance ratio, i.e. the
mechanism that maximizes the market’s (and thus the grid
system’s) value across all users. Sanghavi and Hajek [14]
propose an allocation mechanism τ sh and show that it generates
the optimal performance ratio. While this mechanism
has been proposed for the allocation of network bandwidth,
the setting essentially generalizes to the allocation of any divisible
resource. In the remainder of this section, we will introduce
this mechanism and compare it to the market-based
proportional share mechanism with respect to provider’s
revenue and overall welfare. In the following section, we
will then discuss possibilities of how such mechanisms can
be integrated into state-of-the-art grid schedulers.
3.1 Two Users
For a scenario with two bidders l (low bidder) and h
(high bidder), τ sh allocates shares of the perfectly divisible
resource as follows:
τl sh (w l , w h ) = w l
and τh
sh (w l , w h ) = 1 − w l
2w h 2w h
Allocation scheme τ sh is complemented by a so called payas-bid
pricing scheme such that c(x i ) = p i x i = w i , i =
l, h. For two buyers, the worst case performance ratio of τ sh
adds up to 87.5% when both buyers have linear valuation
functions [14]. In comparison the worst case efficiency of
the proportional share mechanism is 82.84% [8].
Assume the low bidding user l has a quasi-linear valuation
function U l (x l ) = v l x l − w l with v l ∈ R + . Further
assume the high bidding user h to be characterized by utility
function U h (x h ) = v h x h − w h with v h ∈ R + , v h ≥ v l .
Lemma 1. In the Nash equilibrium w sh of the pay-as-bid
mechanism τ sh , user l bids wl
sh = v2 l
2v h
and receives a share
of τl sh (w sh ) = v l
2v h
, whereas user h bids wh
sh = v l
2 , thus
receiving τh
sh(wsh
) = 1 − v l
2v h
.
Proof. In the Nash equilibrium with bid vector w sh ,
∂U i(τ sh (w sh ))
∂wi
sh
∂U l (τ sh (w sh ))
∂wl
sh
= 0, i = l, h. Consequently,
=
v l
2w sh
h
− 1 = 0 ⇔ w sh
h
= v l
2
and ∂U h(τ sh (w sh ))
= v hw sh
∂wh
sh
l
− 1 = 0 ⇔ w sh
2(wh sh)2
l
=
(
2 vl
) 2
v h 2 ⇔ w
sh
l
= v2 l
2v h
. Inserting w sh in τ sh directly
yields τl
sh (w sh ) = v l
and τh
sh(wsh
) = 1 − v l
2v h
2v h
.
The mechanism τ sh allocates the resource shares in such
a way that, in the Nash equilibrium w sh , the low-bidding
user l is pushed to zero utility:
U l (τ sh (w sh )) = v l τl
sh (w sh ) − wl sh v l
= v l − v2 l
= 0.
2v h 2v h
The high-bidding user h obtains utility of
U h (τ sh (w sh )) = v h τh
sh (w sh ) − wh sh = v h − v l ,
while revenue amounts to
r sh = w sh
l
+ w sh
h
= v2 l + v lv h
2v h
.
This constitutes the provider’s utility U P (.) assuming a
quasi-linear provider valuation function and zero reservation
prices.
The central result of our analysis is that, from a resource
provider’s point of view, the pay-as-bid mechanism dominates
market-based proportional share under certain conditions
both with respect to provider’s surplus and welfare.
To be able to compare the results of both mechanisms,
we first derive the provider’s surplus generated by marketbased
proportional share in the Nash equilibrium. In doing
so, we assume that, as the mechanism by Sanghavi and Hajek,
the proportional share allocation rule is also complemented
by the pay-as-bid pricing rule, as proposed in [5].
Lemma 2. If combined with a linear uniform pricing
scheme ∂ci(w)
∂w i
= 1 ∀i (e.g. pay-as-bid), in the Nash
equilibrium w ps of the proportional share mechanism τ ps ,
( ) 2
user l bids w ps
l
= v lv h
v l +v h
− v
vh
l v l +v h
and receives a
share of τ ps
l
(w ps ) = v l
v l +v h
, whereas user h bids w ps
h =
v l
(
)
vh
2,
v l +v h
thus receiving τ
ps
h (wps ) =
v h
v l +v h
.
Proof. With proportional share, τ ps
l
(w) =
τ ps
h (w) = w h
w l +w h
.
Thus, in the Nash equilibrium w ps ,
w
v ps
h l
(w ps
h +wps
w ps
l
w l
w l +w h
∂U l (τ ps (w ps ))
∂w ps
l
and
− 1 = 0 ⇔ √ v
l ) 2 l w ps
h
= w ps
l
+ w ps
h
⇔
= √ v l w ps
h
− wps h . Analogously, ∂U h (τ ps (w ps ))
=
∂w ps
h
w
v ps
l h
− 1 = 0 ⇔ v
(w ps
h +wps l ) 2 h w ps
l
= (w ps
l
+ w ps
h )2 ⇔
( √vl )
( )
v h w ps
h
− wps h
= v l w ps
h
⇔ w ps
h
= v
vh
2.
l v l +v h
Inserting w ps
h
above directly yields, w ps
l
= v lv h
v l +v h
−
( )
v
vh
2.
l v l +v h
Inserting w ps
h
and wps
l
into τ ps (w ps ) returns τ ps
l
(w ps ) =
√ √
vl w h −w
√ h
vl w h
= 1 −
wh
vl
= 1 − v h
v l +v h
= v l
v l +v h
and
τ ps
h (wps ) = 1 − τ ps
l
(w ps ) = v h
v l +v h
.
=
157

Consequently, market-based proportional share generates
revenue of r ps = w ps
l
+ w ps
h
= √ v l w h = v lv h
v l +v h
.
Based on Lemma 1 and Lemma 2, we can now state our
first central result:
Proposition 1. For two users with linear valuation functions
with slopes v l and v h (v l , v h ∈ R + and v l ≤ v h ),
in the unique Nash equilibria the discriminatory pay-as-bid
mechanism generates a larger provider’s revenue than proportional
share (r sh ≥ r ps ) iff
Proof.
( √ 2 − 1)v h ≤ v l ≤ v h .
r sh − r ps = v2 l + v lv h
2v h
− v lv h
v l + v h
≥ 0
⇔ (v l + v h )(v 2 l + v lv h ) − 2v l v 2 h
2v h (v l + v h )
≥ 0
⇔ (v l + v h ) 2 ≥ 2v 2 h
⇔ v l ≥ ( √ 2 − 1)v h
Furthermore, as defined earlier, v h ≥ v l and thus r sh ≥
r ps ⇔ ( √ 2 − 1)v h ≤ v l ≤ v h .
This result is of significant importance in the context of
enterprise/campus grid environments, which are the main
application domains of N1GE. Especially in those scenarios,
we hypothesize that the users can be assumed to have
rather similar valuations for grid resources. Finally, we can
use these results to also assess the allocative efficiency that
is generated in the Nash equilibrium for two jobs, and in
line with the results of Sanghavi and Hajek [14] we state
the following proposition:
Proposition 2. For two users with linear valuation functions
with slopes v l and v h (v l , v h ∈ R + and v l ≤ v h ) the
discriminatory pay-as-bid mechanism generates an equal
or larger total welfare in the Nash equilibrium compared to
the proportional share mechanism with pay-as-bid pricing,
that is U(τ sh (w sh )) ≥ U(τ ps (w ps )) for all combinations
(v l , v h ).
Proof. From our previous results, U(τ sh (w sh )) = U l (.) +
U h (.)+U P (.) = 0+v h −v l + v2 l +v lv h
2v h
= v2 l +2v2 h −v lv h
2v h
and
U(τ ps (w ps )) = U l (.) + U h (.) + U P (.) = (v l τ ps
l
(w ps ) −
w ps
l
) + (v h τ ps
h (wps ) − w ps
h ) + wps l
+ w ps
h
= v lτ ps
l
(w ps ) +
v h τ ps
h (wps ) = v2 l +v2 h
v l +v h
.
Thus, U(τ sh (w sh )) − U(τ ps (w ps )) ≥ 0 ⇔
v 2 l +2v2 h −v lv h
2v h
(v l − v h ) 2 ≥ 0.
− v2 l +v2 h
v l +v h
≥ 0 ⇔ v 2 l
+ v 2 h ≥ 2v lv h ⇔
Consequently, the discriminatory pay-as-bid mechanism
of Sanghavi and Hajek [14] not only provides us with a better
performance ratio (i.e. worst-case bound) than proportional
share, but it outperforms the latter independently of
the choice of v l and v h .
If considered separately from the allocation scheme,
pay-as-bid pricing is a uniform pricing rule as defined in
Lemma 2. However, if combined with allocation scheme
τ sh , the resulting mechanism as a whole produces discriminatory
unit prices; the buyer with a lower bid pays a higher
unit price than the high bidder. This volume discount encourages
high bidders to bid higher, and thus closer to
their true valuation, compared to a scenario with uniform
prices where users can potentially benefit from shading their
bids downwards.The discriminatory pay-as-bid mechanism
is not a “fair” allocation mechanism in that it does not allocate
the resource in proportion to the submitted bids but
subsidizes the high bidding users. However, this is justified
by the increase in overall efficiency.
3.2 n Users
An extension of the above mechanism from two to n
buyers was developed in [14]. This mechanism still has the
property of a “volume discount”, i.e. higher bidders pay
lower prices.
For n buyers and a given payment vector w =
(w 1 , . . . , w n ) , the following allocation rule is proposed:
τ sh
i (w) = w i
w max
∫ 1
0
∏
j≠i
(
1 − s w j
w max
)
ds
with at least two w i ≥ 0 and w max being the maximum
bid. This allocation rule simplifies to the optimal mechanism
for two buyers given above for n = 2.
In contrast to the case for two buyers, it is hard to determine
an exact value for the worst case efficiency for an unlimited
number of buyers. Instead, [14] calculate the interval
[0.8703, 0.875] as bounds for the worst case efficiency.
The proposed mechanism is still close to the theoretical
maximum worst case efficiency, i.e. 87.5%. But a guarantee
that mechanism τ ∗ is the optimal one can no longer be
given.
Even if machines are assumed to be obedient, a “lying
auctioneer” could be a problem in this mechanism (cf.
[11]). The allocation rule is based on the the highest bid,
w max , which is not publicly known and could thus be manipulated
by the provider to change the allocation in his favor.
Therefore, in practice it might be necessary to somehow
publish and verify w max .
158

3.3 Numerical Example
In this section, proportional share and the discriminatory
pay-as-bid mechanism are compared concerning their allocation
and the resulting efficiencies by means of a simple
numerical example. Table 1 provides a brief overview of
the results of this example.
There are two divisions within a company – divisions L
and H – which execute computational jobs on a shared pool
of computing resources.
Division H temporarily demands more resources than
division L, which is reflected in a higher valuation for
the computing resources: U L (x L ) = 2.1x L − w L and
U H (x H ) = 5x H − w H . The provider’s valuation function
is given by U P (x) = w L + w H . Each division sends one
resource request to the central market-based scheduler. Attached
to both jobs are the corresponding valuations. Now
we determine the allocation of resource shares to the jobs
for both market-based mechanisms.
For the two jobs and the given valuation functions, the
proportional share mechanism arrives at the Nash equilibrium
with bid vector w ps = (w ps
L , wps H
) ≈ (0.437, 1.041).
In this equilibrium point, x ps
L
= 0.296 is allocated to division
L and x ps
H
= 0.704 is allocated to division H. None
of the divisions has an incentive to unilaterally deviate from
these bids. The unit prices are p ps
L
= pps H
= 1.478. Hence,
the proportional share mechanism generates an overall social
welfare of U(x ps ) = 4.142. This corresponds to a
performance ratio of 82.84% compared to the optimal allocation.
In this optimal allocation, the high bidding user
is given a resource share of 1.0 whereas the low bidder receives
nothing. This allocation would create the maximum
social welfare of 5.
The discriminatory pay-as-bid mechanism reaches the
Nash equilibrium for a bid vector w sh = (wL sh,
wsh H ) =
(0.441, 1.05). With these bids a resource share of x sh
H =
L =
0.79 is assigned to division H and the remaining x ps
0.21 are allocated to L. This allocation generates a total
social welfare U(x sh ) = 4.391, which stands for a performance
ratio of 87.82%. Applying the pay-as-bid-rule to this
allocation exemplifies the “volume discount” for the high
bidding division H. It pays a unit price of p ps
H = 1.329
whereas division L has to pay a notably higher unit price of
p sh
L = 2.1.
Whereas in this example the provider’s revenue created
by the discriminatory pay-as-bid mechanism (r sh = 1.491)
is only slightly higher than with the proportional share
mechanism (r ps = 1.478), overall welfare increases by almost
5% from U(x ps ) = 4.142 to U(x sh ) = 4.391. While
here both the high bidder and the provider benefit from
switching to the pay-as-bid mechanism, as showed above,
this gain mainly comes at the expense of the low bidder,
whose utility becomes zero and who is thus indifferent to
Optimal Bids
Allocation
Unit prices
Utilities Users
Provider’s Revenue
Prop. Share
w ps
L
w ps
H
x ps
L
Pay-as-Bid
= 0.437 wsh L = 0.441
= 1.041 wsh H = 1.05
= 0.296 xsh L = 0.21
x ps
H = 0.704
p ps
L
= pps H = 1.478
U L(x ps
L ) = 0.185
U H(x ps
H ) = 2.479
xsh H = 0.79
p sh
L = 2.1
p sh
H = 1.329
UL(xsh L ) = 0
UH(xsh H ) = 2.9
r ps = 1.478 r sh = 1.491
Social Welfare U(x ps ) = 4.142 U(x sh ) = 4.391
Performance
ratio
82.84% 87.82%
Table 1: Numerical example
not participating at all.
In certain cases, the provider might be willing to actually
sacrifice revenue in return for a higher social welfare.
E.g. in the case of Synopsys, attributing the resources to the
division with the higher valuation may be more important
than creating revenue from internal sources. Especially in
cases when provider revenue increases, one option might be
to “subsidize” the lower bidder in order to convince this bidder
to participate in the pay-as-bid mechanism. It will be an
interesting avenue for future research to explore how such
an incentive may be implemented in the pay-as-bid mechanism
and how it changes the users’ strategic considerations.
4 Integration into the Sun N1 Grid Engine
The following section is dedicated to the integration of
the market-based mechanism into the technical scheduling
environment of N1GE. Two possible approaches are evaluated:
1) a modular extension of N1GE, with an additional
market-based policy and 2) the displacement of the current
technical scheduler with the discriminatory pay-as-bid
mechanism. The analysis is concluded with a comparison
of both approaches in Section 4.3.
4.1 The Pay-as-Bid Mechanism as Additional
Policy
In the following section, the application of the discriminatory
pay-as-bid mechanism as a new policy in N1GE is
discussed. The objective is to use the market-based mechanism
as an instrument for partioning of the priority value.
Analogous to the current N1GE system the users submit
their jobs along with a specification document to state the
job’s resource requirements. In addition the users send a
159

one-dimensional bid (a single real number) to signal their
valuation of the submitted job. These bids are then used by
the discriminatory pay-as-bid mechanism to allocate shares
of the priority value, which then again are used to determine
the order in the waiting list of pending jobs. After
sorting the jobs according to their assigned priority value
the technical scheduler traverses the waiting list and picks
a job for execution. Since the jobs can have different resource
requirements the first job that fits the specifications
of a currently available resource is chosen. Thus it could
happen that a job that is heading the waiting list is not chosen
due to its resource requirements. But still, taking up
one of the front slots in the waiting list increases the possibility
of early execution. Thus it is an incentive for jobs,
respectively their owners, to compete for a high share of the
priority value. As long as the job is not released for execution
by the technical scheduler, a user can always abort his
own jobs.
In addition to this new policy, the currently available
policies in N1GE will still be at the administrator’s disposal.
This can be rather easily be achieved by adding the
new discriminatory pay-as-bid policy to the existing policy
mix (see figure 2). The value which is determined by the
market-based mechanism is weighted and added to the total
priority value of a job:
ˆP mix = P mix + W sh ∗ N sh
Thereby, ˆP mix is the new dispatch priority, P mix the priority
value generated with the current N1GE policy mix
and N sh , W sh the discriminatory pay-as-bid priority (on
an interval between 0 and 1) respectively the corresponding
weighting factor for this new policy.
The main objective of the integration process is to preserve
the economic properties of the discriminatory payas-bid
mechanism. Therefore, the approach is to incorporate
the mechanism mostly unchanged into a N1GE environment
that is tailored to the needs of the mechanism. To
guarantee a sound concurrence of N1GE system and the discriminatory
pay-as-bid mechanism as a new policy, a number
of extensions and modifications have to be considered
on both sides. Adaptation of the discriminatory pay-as-bid
mechanism requires a number of considerations in different
areas. The major challenges are evaluated next and suitable
solutions are proposed. Issues concerning the payment and
the enforcement thereof, are not in the scope of this work
and thus it plays only a tangential role in the evaluation.
4.1.1 Re-evaluation of the Priority Value
Recalculation of the priority values of all pending jobs is
necessary whenever a new job enters the waiting list. Every
new bid changes the priority values of all waiting jobs.
In addition, a re-evaluation falls due for each updated bid
that is received in the system. A job that leaves the waiting
list for any reason does not require an update of the
waiting list. The re-evaluation is processed periodically according
to a pre-defined interval. A continous evaluation
of newly arrived jobs would put to much additional load on
the provider. Jobs, arriving at the waiting list during such
an interval, are gathered and inserted into the waiting list
within the next interval. In the worst case, utilisation of
this periodic allocation scheme delays all arriving jobs by
the pre-defined interval. This is undesirable especially for
time critical and urgent jobs. Currently, the bidding interval,
or scheduler interval as it is called in N1GE, is set to
2-3 seconds. The testbed environments work with a 15 second
interval on default. Consequently, these values seem to
be reasonable references for the re-evaluation interval.
4.1.2 Bid Updating
Figure 2: Modified scheduling process in N1GE with the discriminatory
pay-as-bid mechanism as additional policy
Since re-evaluation of the priority value is indispensable in
the N1GE scenario anyway, bid updating imposes no further
computational effort on the scheduler. But it tremendously
increases the communication effort. For each bid
that is updated an additional message, carrying the single
real value, has to be sent by the user and processed by the
provider. However, allowing bid updates cannot interfere
with the economic properties of the mechanism, since the
Nash equilibrium is considered to be the final point of repeated
plays. Thus, bid updates can even improve the speed
of convergence since users can always adjust their bids towards
the equilibrium. These adjustments are based on the
behaviour of the other jobs in the queue. The bid updates
can be interpreted as part of the ’myopic best response’ bidding
strategy which leads to the Nash equilibrium. A received
bid, respectively bid update, for a particular job is
valid until another modified bid for this job is submitted by
the user. Another feature of bid updates is the prevention
160

of job starvation. Whenever job starvation impends the job,
the user can intervene and update the bid.
4.1.3 Feedback
For a convergence to the Nash equilibrium point, the users
need feedback on their bids to see what resource shares they
received. Using this partial market information, they will
adjust their bids in ”myopic best responses” to finally reach
the Nash equilibrium point. Consequently, an accurate feedback
is essential to preserve this equilibrium concept. By
now, N1GE scheduler does not support this kind of direct
feedback. The reporting tool for the waiting list, named qstat
in N1GE, shows the priority value for each job. This
priority value is the accumulated value for all policies that
are part of the policy mix which is employed in the specific
scenario. Thus, the users are not able to determine the fraction
of the priority value that was calculated based on their
bid. The speed of convergence to the Nash equilibrium can
be further improved by publishing the whole waiting list
in an anonymized form. This enables the users to monitor
both the priority value and the corresponding position in the
waiting list.
4.2 The Pay-as-Bid Mechanism as Scheduler
In this scenario the discriminatory pay-as-bid mechanism
is applied as a scheduler to directly allocate the resources
to the bidding users. In contrast to the scenario assumed
so far, the mechanism does not calculate the respective
priority value for each bidding user, but it determines
the actual share of resources a user gets.
The major challenge, that has to be mastered in this scenario,
is the architectural limitation of N1GE, which only
allows a single task to be executed on each slot at a time.
Therefore resources can no longer be assumed as being perfectly
divisible. There are two solution concepts to handle
this restriction of N1GE. Firstly, the N1GE’s limitation
can be attacked from the mathematical side. This requires
a modification of the mechanism’s allocation rule to allow
discrete shares only. Likewise, the problem can be tackled
from the technical side. Employing virtualization tools
could restore the perfect divisibility of the resources. Both
solution concepts are discussed in the following – showing
advantages and disadvantages of both.
4.2.1 Discrete Resources
The main challenge is to modify the current mechanism in
such a way that it can be used for discrete resources. The
calculated shares may only equal a multiple of 1 m , where
m is the number of currently idle resources. By putting up
a linear integer program a completely new mechanism is
created. Since the discriminatory pay-as-bid mechanism is
already the optimal mechanism (at least for two users) the
linear integer program can not improve the current allocation
and will most likely result in losses in the user’s utilities
and overall efficiency. In addition, further issues have
to be resolved. Firstly, the impact on the computational
tractability has to be analysed. This is mainly dependent
on the complexity class of such a linear integer program.
Secondly, the convergence to a (Nash) equilibrium needs a
more detailed examination. Taking all the considerations
into account, this approach of restricting the mechanism to
discrete resources is not very promising. Solving the linear
integer program does hardly scale and will get computational
intractable for a large number of jobs.
4.2.2 Virtualization
This approach tries to solve the restriction from the technical
perspective. Virtualization could be the solution to
the ”indivisibility constraint”. It is a technique for hiding
the physical characteristics of the underlying computing
resources from the way in which other systems or end
users interact with those resources. Not only homogeneous
resources can be joined to form a virtual resource,
with virtualization any type of resource can be combined.
State-of-the-art virtualization tools, such as e.g. VMWare
(http://www.vmware.com), can reach an almost perfect divisibility.
Instead of directly addressing resources, the
scheduler communicates with a virtual machines which
again manages the resources as such. To the scheduler,
the available resources appear to be one big resource. This
would again fit the requirements of the discriminatory payas-bid
mechanism, which is designed for allocation of a single
unit of a perfectly divisble resource.
In the following, a ’virtual resource’ is referred to as
a single logical representation of multiple resources. Furthermore,
it is assumed that virtualization is employed as
a mean to enable the usage of the discriminatory pay-asbid
mechanism for scheduling in N1GE. The adaptation of
the mechanism to discrete resources is discarded due to the
drawbacks of such an approach discussed earlier. Figure 3
shows the modified N1GE scheduling process.
The challenges in this application scenario differ essentially
from the ones identified in the first scenario. For application
of the discriminatory pay-as-bid mechanism as a
direct scheduler, N1GE’s central architecture and workflow
has to be radically changed. The market mechanism is not
integrated as an modular extension but directly embedded
into N1GE’s core structure.
4.2.3 Mode of Allocation
The scheduler allocates the jobs for a single time slot only.
Consequently, a new allocation is calculated for all cur-
161

to a fixed price, per job and time slot, paid in advance.
4.2.5 Feedback
Finally, a comprehensive feedback is the base for sustaining
the convergence to the Nash equilibrium. The user needs to
be informed about the share that he receives. This enables
the user to see the impact of his bid on the resource share
that is finally allocated to him. Analogous to the ”policy
scenario” the provider has to put the level of allocated resource
share at the user’s disposal.
4.3 Implications
Figure 3: Modified scheduling process in N1GE with the discriminatory
pay-as-bid mechanism as direct scheduler
rently running jobs after each time slot. This mode of allocation
allows a flexible adjustment of short-time changes
in the demand and supply situation. None of the resources
is blocked for more than one time slot. The downside is
an increased effort, both on the users’ and provider’s side.
The users need to constantly monitor the allocation process
for their running jobs and have to submit bids for every allocation
interval. Otherwise, when no bid is received for
a particular job, this job is suspended. Furthermore, the
users are no longer able to determine the total execution
time and corresponding payments for their jobs in advance.
Moreover, the allocation for a single time slot burdens the
provider with migration and re-prioritization of (nearly) all
jobs after each allocation interval. However, this type of allocation
does not affect the economic properties of the market
mechanism. This can be easily proved since the method
equals the initially proposed scenario by [14], in which the
bandwidth of a network link was allocated for a single predefined
time slot.
4.2.4 Job Length
A further factor that is not directly employed in the allocation
scheme yet, is the job length. A major problem in this
context is the determination of the job length in advance,
i.e. before execution of the job. This is particularly a problem
in an environment with heterogeneous resources. As
already discussed in the previous section, the job length can
indirectly be incorporated by allocating the resource for a
single time slot. The total payment of the user is determined
by summing up the fractional payments made for each allocation
interval. Consequently, the job length does not need
to be known in advance. Furthermore, it can be guaranteed
that the user only pays for units of the resource that he really
used, which adds an additional notion of fairness compared
The two integration scenarios discussed in the previous
subsections represent different approaches of extending the
N1GE with economic features. Introducing the discriminatory
pay-as-bid mechanism as an additional policy constitutes
an modular extension of the N1GE. Most of the current
architecture and structure retains unchanged. Solely
the payment system requires major modifications. The policy
mix still covers the existing policies besides the newly
introduced market-based principle. In contrast, the direct
scheduling scenario requires modifications and replacement
of core components. Consequently, the implementation effort
is much higher with this approach. In addition, the allocated
resources are limited to a single type whereas the
modular extension supports heterogeneous resources. On
the other hand, giving the discriminatory pay-as-bid mechanism
direct access to the resources entails a multitude of advantages.
For instance, a waiting list is dispensable since all
jobs get a share of the resource. Furthermore, the additional
in-between scheduler can be left out which removes complexity
from the scheduling process. Both scenarios offer
promising enhancements of N1GE. The modular extension
offers a quick solution for incorporation of the discriminatory
pay-as-bid mechanism, whereas the higher flexibility
of the second approach comes at the expense of higher implementation
effort.
5 Conclusion
The extended model of the Sanghavi/Hajek pay-as-bid
mechanism is a promising addition to the N1GE scheduler.
Employing a market-based mechanism for resource allocation
in grids offers new possibilities on both sides, for
providers as well as for buyers. Current technical schedulers
require an administrator to specify user weights based
on these users relative importance, regardless of the dynamic
demand and supply situation, leading to inefficiencies.
To this end, the Sanghavi/Hajek pay-as-bid mechanism
allows flexible reactions to changes in the demand and
supply situation. Moreover, it offers an elaborated pricing
162

scheme where prices reflect the current market situation and
induce users to report their true valuations to the system.
The administrator no longer needs to adjust the weights
manually but the users can directly express the urgency of
their jobs. Furthermore, the market prices can be leveraged
for usage-based accounting of shared computer resources.
In comparison to other market-based mechanisms, the
discriminatory pay-as-bid mechanism scores with its easeof-use.
In addition, it has an increased worst case performance
ratio as compared to market-based proportional
share and is close to the theoretical maximum. Above all,
the extended model imposes a very low communicational
and computational burden on the scheduling process and
allows for real-time allocations.
Further work has to be done on analyzing the extensions
to the basic mechanism and their impact on the mechanism’s
economic properties. We explicitly mentioned the
need to incentivize the low bidding user to participate in the
grid. It would be very interesting to actually implement the
mechanism within the N1GE scheduler. This would allow
us to examine the mechanism’s behavior and performance
in practice. A decentralized version of the mechanism
would be desirable to support decentralized waiting queues
as well. This might be necessary to keep the N1GE scheduler
applicable for very large clusters (> 20, 000 cores),
which will be demanded in the near future.
Acknowledgements
We would like to thank the anonymous reviewers for
their valuable comments. This research was partially supported
by the EU IST programme under grant #034286
SORMA – “Self-Organizing ICT Resource Management”.
References
[1] Anonymous. Grid Computing: A Vertical Market Perspective
2006-2011. Technical report, The Insight Research Corportation,
Boonton, NJ, USA, 2006.
[2] A. AuYoung, B. Chun, A. Snoeren, and A. Vahdat. Resource
allocation in federated distributed computing infrastructures.
Proceedings of the 1st Workshop on Operating
System and Architectural Support for the On-demand IT InfraStructure,
2004.
[3] R. Bapna, S. Das, R. Garfinkel, and J. Stallaert. A Market
Design for Grid Computing. Technical report, Technical
report, Department of Operations and Information Management,
University of Connecticut, 2005.
[4] C. Chaubal. Scheduler Policies for Job Prioritization in
the Sun N1 Grid Engine 6 System. Technical report,
Sun BluePrints Online, Sun Microsystems, Inc., Santa
Clara, CA, USA. http://www.sun.com/blueprints/1005/819-
4325.pdf, 2005.
[5] B. N. Chun and D. E. Culler. Market-based proportional
resource sharing for clusters. Technical report, University
of California at Berkeley, CA, USA, 2000.
[6] Enabling Grids for E-sciencE (EGEE) project. Website,
2007. Available online at http://www.eu-egee.org/; visited
on August 29th 2007.
[7] I. Foster, S. Tuecke, and C. Kesselman. The Anatomy of
the Grid: Enabling Scalable Virtual Organizations. International
Journal of High Performance Computing Applications,
15(3):200–222, 2001.
[8] R. Johari and J. N. Tsitsiklis. Efficiency Loss in a Network
Resource Allocation Game. Mathematics of Operations Research,
29(3):407–435, 2004.
[9] K. Lai, B. Huberman, and L. Fine. Tycoon: A Distributed
Market-Based Resource Allocation System. Arxiv preprint
cs.DC/0404013, 2004.
[10] J. K. MacKie-Mason and M. P. Wellman. Automated markets
and trading agents. In L. Tesfatsion and K. L. Judd,
editors, Handbook of Computational Economics.
[11] R. T. Maheswaran and T. Basar. Coalition formation in proportional
fair divisible auctions. In Proceedings of the Second
International Joint Conference on Autonomous Agents
and Multiagent Systems, pages 25–32, 2003.
[12] D. Neumann, S. Lamparter, and B. Schnizler. Automated
bidding for trading grid services. In Proceedings of the European
Conference on Information Systems (ECIS), 2006.
[13] D. Neumann, J. Stößer, and C. Weinhardt. Bridging the
Adoption Gap – Developing a Roadmap for Trading Grids.
Electronic Markets, 2007, forthcoming.
[14] S. Sanghavi and B. Hajek. Optimal Allocation of a Divisible
Good to Strategic Buyers. 43rd IEEE Conference on
Decision and Control, 2004.
[15] B. Schnizler, D. Neumann, D. Veit, and C. Weinhardt.
Trading Grid Services-A Multi-attribute Combinatorial Approach.
European Journal of Operational Research, forthcoming,
2006, forthcoming.
[16] I. Stoica, H. Abdel-Wahad, K. Jeffay, S. Baruah, J. Gehrke,
and C. Plaxton. A proportional share resource allocation
algorithm for real-time, time-shared systems. IEEE Real-
Time Systems Symposium, 1996.
[17] J. Stößer, D. Neumann, and A. Anandasivam. A truthful
heuristic for efficient scheduling in network-centric grid os.
In Proceedings of the 15 th European Conference on Information
Systems (ECIS), 2007.
[18] Sun Microsystems, Inc. Sun N1 Grid Engine
6 User’s Guide. Santa Clara, CA, USA.
http://docs.sun.com/app/docs/doc/817-6117/, 2004.
[19] Sun Microsystems, Inc. Grid Computing Solutions.
Website, 2007. Available online at
http://www.sun.com/software/grid/; visited on August
29th 2007.
[20] R. Wolski, J. Plank, J. Brevik, and T. Bryan. Analyzing
Market-Based Resource Allocation Strategies for the Computational
Grid. International Journal of High Performance
Computing Applications, 15:258–281, 2001.
163

Decentralized Online Resource Allocation for Dynamic Web Service Applications
Jochen Stößer, Christine Rössle, Dirk Neumann
Institute of Information Systems and Management (IISM)
Universität Karlsruhe (TH)
Englerstr. 14, 76131 Karlsruhe, Germany
{stoesser,roessle,neumann}@iism.uni-karlsruhe.de
Abstract
Economic scheduling mechanisms introduce economic
principles to Grids and thus promise to increase the
systems’ overall utility by explicitly taking into account
strategic, inter-organizational settings. Current economic
scheduling mechanisms, however, do not fully satisfy the requirements
of interactive Grid applications. In this paper,
we introduce a video surveillance environment and show
how its functionality can be encapsulated in Grid services
which are traded dynamically using an economic scheduling
mechanism. We suggest a mechanism from the general
machine scheduling domain and showcase its applicability
to the Grid context. However, the mechanism suffers from
some limitations due to the distinct properties of Grids. We
point to these limitations and propose extensions to the basic
mechanism which may alleviate these drawbacks.
1. Introduction
Distributed computing environments such as compute
clusters are a powerful concept to tackle complex computational
problems. In recent years, the concept of Grids has
emerged, extending these computing environments to comprise
not only computing resources within an organization
but across administrative domains to further increase efficiency
and flexibility [6].
Currently, mainly physical resources such as processing
power, memory and storage are being shared in Grids.
However, to be deployed in Grids, applications must be able
to achieve Quality of Service assertions despite running on
heterogeneous Grid resources and native platforms [6]. The
Open Grid Services Architecture (OGSA) describes a shift
from physical resources to Web services being shared in the
Grid [7]. Web services are used to hide the underlying infrastructures’
heterogeneity from the Grid application/user
and aggregate basic physical resources to more complex and
elaborate services. As illustrated in [12], the technology gap
between the Grid community and the Web services community
has continuously been narrowed and recently resulted
in the OGSA WSRF Basic Profile 1.0 in which the two technologies
converge into so called Grid services or OGSA
services which can be seen as stateful extensions to classic
Web services [8]. A recent and prominent example for
the application of Web services in the Grid context is Amazon’s
Elastic Compute Cloud that allows Web service-based
deployment of applications on Amazon’s compute platform
[1].
A further example for such a Grid service-based application
with distinct properties and requirements is a video
surveillance environment as described in [15]. Imagine an
office building which accommodates several organizations,
each operating video surveillance cameras. The aim of
the surveillance system is to detect persons and track their
movement across the facility. The analysis of the resulting
video streams requires vast amounts of computing power.
However, to avoid each organization having to deploy its
own, separate cluster and to increase resource utilization,
the organizations deploy one common inter-organizational
compute Grid. In order to perform target separation, the
analysis of a video stream performs a set of tasks which run
sequentially in a pipeline: motion detection, body detection,
direction estimation – in which direction is the person moving?
– and field of vision detection – what is the identity of
the spotted person? Each of these tasks can be encapsulated
Figure 1. Grid service pipeline
164

in a Grid service which can then be instantiated multiple
times, possibly spread across multiple physical machines.
Each video stream runs sequentially through these Grid services
as illustrated in figure 1.
If computer resources are scarce, a scheduling strategy
is needed that decides which Grid service should be allocated
to which camera feed at what time. Classic technical
scheduling mechanisms such at first-come-first-serve or fair
share are solely built on system-centric measures. Consequently,
they do not take into account the strategic dimension
in Grids, i.e. self-interested organizations with heterogeneous
preferences and objectives [5]. For instance, during
some periods of time, the required security level may
vary within organizations. With technical scheduling mechanisms,
these organizations cannot express this variation
in their valuation of the shared Grid services. Economic
scheduling mechanisms are deemed promising to provide a
better fit to the Grid’s inter-organizational nature by allowing
both service requesters and providers to express valuations
in addition to technical metrics. Ultimately, prices are
formed which help to balance dynamic demand and supply
in the Grid. The system can thus induce service requesters
to distribute excess demand over time. In addition, the system
gives service providers incentives to contribute idle resources
to the Grid.
This paper is structured as follows. In section 2, the
requirements for an economic scheduling mechanism for
dynamic Web service applications are outlined. Section 3
discusses related work on this class of scheduling mechanisms.
In section 4 we introduce a mechanism for machine
scheduling proposed by Heydenreich et al. in [11] and subsequently
showcase its general applicability for interactive
Grid settings such as the video surveillance application at
hand in section 5. We will point to some limitations and extend
this mechanism in the Grid context in section 6 before
we recapitulate the main points in section 7.
2. Requirements
The video surveillance Grid service application poses a
number of requirements towards an economic scheduling
mechanism [16]:
• Immediacy: The timing of the allocation is crucial.
The allocation of services in this scenario is time critical
as the monitored objects are dynamic in relatively
constant movement and the workflow of services is
sensitive to the order of execution [15]. For instance,
the direction estimation process may not be executed
prior to the body detection process. Consequently, a
so called online mechanism is needed which allocates
requests as soon as they arrive in the system instead of
pooling and allocating requests and offers periodically.
• Allocative efficiency: This requirement addresses that
there should be no “waste” of resources; the system is
supposed to make optimal use of its resources.
• Decentralized scheduling removes the need for a central
bottleneck which may not scale sufficiently or
cause the whole system to fall over, an aspect that is
crucial for large scale applications.
• Incentive compatibility: participants cannot benefit
from cheating the mechanism, i.e. reporting any valuation
for a service other than their true valuation.
• A mechanism is said to be individually rational if its
users cannot suffer any loss in utility from participating.
• The payment scheme of a budget-balanced mechanism
is designed such that it does not require any subsidies;
the payments to service providers are covered by the
payments from service requesters. Any mechanism
must be both individually rational and budget-balanced
in order to be sustainable over time.
In the following, we are interested in economic scheduling
mechanisms that satisfy those requirements.
3. Related Work
There are two main classes of economic scheduling
mechanisms with respect to the mode of allocation: mechanisms
which execute periodically and mechanisms which
execute continuously.
Hitherto, periodic economic mechanisms have received
the most interest from the economic mechanism design
community. The prototype SPAWN implements a market for
computing resources in which each workstation auctions off
idle computing time to multiple applications by means of a
Vickrey auction [17]. All resources are allocated to at most
one application at a time regardless of this application’s actual
demand which yields low resource utilization. POP-
CORN is a Web-based market for computing power which
implements a Vickrey auction as well as two double auctions
[14]. While these two mechanisms focus on the trading
of processing power only, Schnizler et al. elaborate a
multi-attribute combinatorial exchange (MACE) in which
users are allowed to request and offer arbitrary bundles of
Grid resources and can specify quality attributes on these
resources [16]. MACE implements an exact mechanism.
The scheduling problem in this combinatorial setting is NPhard,
the pricing scheme of MACE yields approximately
incentive compatible prices.
The BELLAGIO system also implements an exact combinatorial
exchange for computing resources [3]. Being an
exact mechanism it shares the computational drawbacks of
165

MACE. In the mechanism proposed by Bapna et al., multiple
requesters and providers can trade both computing
power and memory for a sequence of time slots [4]. First,
an exact mechanism is introduced. By imposing both fairness
constraints and one common valuation across all resource
providers, the search space in the underlying combinatorial
allocation problem is structured as to establish one
common, incentive compatible price per time slot for all
accepted requests. Additionally, this introduces a linear ordering
across all jobs and time which reduces the complexity
of the allocation problem, which, however, still remains
NP-hard. To mitigate this computational complexity, a fast,
greedy heuristic is proposed at the expense of both incentive
compatibility and efficiency. All of these approaches may
be applicable for batch applications where immediacy can
be sacrificed in return for increased economic efficiency.
However, due to periodic allocation, the complexity of the
underlying allocation problem and the necessary time to actually
compute the allocation, all of these mechanisms are
inappropriate for interactive Grid settings and do not satisfy
the required immediacy needed for the video surveillance
Grid service application at hand. Consequently, a continuous
economic mechanism is required.
The most prevalent continuous economic mechanisms –
also called online mechanisms – are based on the concept
of market-based proportional share as proposed in [5] and
[13]. Each requester receives a fraction of the total available
resources according to her share in the total reported
valuation across all requesters; i.e., if requester i out of n
requesters reported a valuation of v i , she will receive a fraction
amounting to v i / ∑ n
j=1 v j of the resource. The problem
with market-based proportional share is that it does
generally not support Quality of Service assertions. The
mechanism holds the risk that requesters are not able to
obtain the necessary resources [13]; the share of resources
provided to a requester, i.e. the actual service level returned
to the requester, will generally be below or above this requester’s
required service level. More importantly, the basic
form of market-based proportional share as suggested in
[5] does not support advance reservation and may result in
high latency which is fatal for interactive applications [13].
The AUCTION SHARE mechanism proposed by Lai et al.
extends the basic market-based proportional share mechanism
as to support low latency, advance reservations and an
incentive compatible pricing scheme [13].
In summary, the employed mechanisms proposed for
Grids do not fully satisfy all requirements stated in section
2.
4. A Decentralized Local Greedy Mechanism
A Grid environment consists of a set of individual users
(agents) which we will assume to act rationally and selfishly,
meaning the agents intend to maximize their individual
benefit from participating in the Grid. In the following,
this setting will be introduced as an instance of the class of
machine scheduling problems P| r j | ∑ w j C j [9].
Agents submit (computational) tasks to the system, so
called jobs. A job j is specified by the tuple t j =
(r j , p j , w j ) where r j ∈ R + 0 denotes the job’s release date,
p j ∈ R + the job’s processing time and w j ∈ R + the job’s
weight, i.e. the job’s cost of waiting for one additional unit
of time. When requesting resources for the execution of a
job, the selfish agent 1 j may strategically submit the request
(its type [10]) ˜t j = (˜r j , ˜p j , ˜w j ) ≠ t j to the system.
The Grid infrastructure consists of m identical, parallel
machines M = {1, ..., m}. We restrict the strategic analysis
of this setting to selfish behavior on the demand side only
and consequently assume that machines always report true
information about their status to the system.
Each of the jobs J = {1, ..., n} can be executed on any of
the m machines. Jobs are assumed to be numbered in order
of their release dates, i.e. j < k ⇒ ˜r j ≤ ˜r k . Each machine
can process one job at a time. Preemption of jobs is not
allowed, meaning once a job has been started, it cannot be
interrupted externally.
For this setting, Heydenreich et al. [10] propose a DE-
CENTRALIZED LOCAL GREEDY MECHANISM (DLGM)
which aims at maximizing the agents’ overall “happiness”
by minimizing the sum of the jobs’ weighted completion
times ∑ j∈J w jC j – the costs induced by the system –
where C j denotes j’s actual completion time.
DLGM comprises the following steps:
1. At its chosen release date ˜r j , job j communicates
˜w j and ˜p j to every machine m ∈ M. It is assumed
that a machine can postpone its release date, therefore
˜r j ≥ r j . The processing time can only be overstated
(˜p j ≥ p j ), e.g. by adding unnecessary calculations,
but not understated since this could easily be detected
and punished by the mechanism [10]. The weight ˜w j
can be reported arbitrarily.
2. Based on the received information (˜r j , ˜p j , ˜w j ), the machines
communicate a (tentative) completion time Ĉj
and a (tentative) payment ˆπ j to the job. The tentativeness
is due to the fact that later arriving jobs might
overtake job j. This leads to a final (ex-post) completion
time C j ≥ Ĉj and a final (ex-post) payment
π j ≤ ˆπ j as compensation payments by overtaking
jobs might occur (see below). The local scheduling
on each machine follows the ”weighted shortest processing
time first” (WSPT) policy. Jobs are assigned
a priority value according to their ratio of weight and
processing time: job j has a higher priority than job k
1 In the following we will use the terms “agent” and “job” interchangeably.
166

⇔ ˜w j /˜p j ≥ ˜w k /˜p k and is thus inserted in front of k
into the waiting queue at the machine. Ties may be resolved
by time of arrival or randomly, for instance. Let
H(j) denote the set of jobs with higher priority than
j and L(j) be the set of jobs with lower priority than
j, respectively. Furthermore, let S j denote the point in
time when job j is started to be processed and b i (t) the
remaining processing units of the currently executed
job at time t on machine i. j → i indicates that job j
is assigned to machine i. Then (tentative) completion
times Ĉj(i) and payments ˆπ j (i), j ∈ J and i ∈ M, can
be computed as
∑
Ĉ j (i) = ˜r j + b i (˜r j ) +
˜p k + ˜p j
k∈ H(j),k→i,k

formance can be assured; the objective value ∑ j∈J w jC j –
which is to be minimized – resulting from the online mechanism
where information about jobs is gradually becoming
available is compared to the the optimal solution by an
offline mechanism whose computations are based on prior
knowledge of the entire information space. Heydenreich et
al. [10] derive a value of ρ = 3.281 for DLGM, meaning the
objective value of DLGM is guaranteed to be no more than
ρ = 3.281 times the objective value of an optimal offline
mechanism.
5. Sample Scenario
In this section, we provide a numerical example to
demonstrate the applicability of the DLGM to the video
surveillance environment introduced in section 1. In our
example setting, there are five cameras performing video
stream analysis with the aim of target separation. A job
can in this context be interpreted as a segment of a video
stream which needs to be analyzed. According to the Grid
service pipeline illustrated in figure 1, each job requests to
be assigned sequentially to the four task types. For the sake
of simplicity, we confine our example to the first two tasks
in the pipeline, i.e. motion detection (MD) and body detection
(BD) (cf. figure 1). Furthermore, the critical job
restriction is neglected. Each of the tasks is encapsulated
in a corresponding Grid service. We assume that we have
two machines for each Grid service, i.e. MD 1 and MD 2
for motion detection and BD 1 and BD 2 for body detection.
The different job types are generated sequentially in
the given order at each camera: only when a MD-job has
been processed and the result has been transferred back to
the camera, a BD-job can be generated. This illustrates
the interactive processing mode of this Grid service application.
We can partly avoid combinatorial problems since
it is not possible that e.g. a direction estimation job is executed
without the underlying motion detection job being
completed. Note, however, that it is still possible that the
MD-part of a camera feed is being analyzed but the camera
is not able to subsequently obtain the BD-service. A job
which is released at time t can be started being processed at
time t + 1. The order in which jobs with identical release
dates are being considered by the machines is determined
randomly; the same holds for the cameras’s choice between
two machines if both yield the same utility. Table 1 lists
all the generated jobs with the respective camera and service
type as well as their type (˜r, ˜p, ˜w) and the priority ratio
˜w/˜p.
The waiting queues for each service and timestep are illustrated
in figure 2; empty queues have been omitted, active
jobs are highlighted. At time r = 0, cameras 1 to 3
each generate a MD-job. All jobs report to the two MDmachines
which compute tentative completion times and
job camera service ˜r ˜p ˜w ˜w/˜p
1MD 1 MD 0 2 6 3
2MD 2 MD 0 4 5 1.25
3MD 3 MD 0 1 6 6
4MD 4 MD 1 1 10 10
5MD 5 MD 2 2 7 3.5
3BD 3 BD 2 2 11 5.5
1BD 1 BD 3 2 5 2.5
4BD 4 BD 3 1 8 8
Table 1. Job generation at the cameras in
chronological order
Figure 2. Sample waiting queues
job Ĉ j (MD 1 ) ˆπ j (MD 1 ) û j (MD 1 ) choice
Ĉ j (MD 2 ) ˆπ j (MD 2 ) û j (MD 2 )
1MD 0+0+0+2=2 0 -12 MD 1
0+0+0+2=2 0 -12
2MD 0+0+2+4=6 0 -30 MD 2
0+0+0+4=4 0 -20
3MD 0+0+0+1=1 1*6=6 -6-6=-12 MD 2
0+0+0+1=1 1*5=5 -6-5=-11
4MD 1+1+0+1=3 0 -30 MD 2
1+0+0+1=2 1*5=5 -20-5=-25
5MD 2+0+0+2=4 0 -28 MD 1
2+0+0+2=4 1*10=10 -28-10=-38
Table 2. Motion Detection (MD) service
168

job Ĉ j (BD 1 ) ˆπ j (BD 1 ) û j (BD 1 ) choice
Ĉ j (BD 2 ) ˆπ j (BD 2 ) û j (BD 2 )
3BD 2+0+0+2 = 4 0 -44 BD 1
2+0+0+2 = 4 0 -44
1BD 3+1+0+2 = 6 0 -30 BD 2
3+0+0+2 = 5 0 -25
4BD 3+1+0+1 = 5 0 -40 BD 2
3+0+0+1 = 4 1*5 = 5 -32-5 = -37
Table 3. Body Detection (BD) service
payments (cf. table 2). With no other jobs being present at
any machine, camera 1 selects MD 1 to execute job 1MD
since it yields a higher utility. Camera 2’s job 2MD is
being considered next; due to the occupation of MD 1 by
job 1MD, 2MD is scheduled on the other machine MD 2 .
Since both jobs are allocated to idle machines, they do not
need to compensate any other jobs. Next, job 3MD is evaluated.
Both MD-machines now already have one job waiting
to be executed from the next timestep on. But since
MD 3 ’ s priority value exceeds the other two jobs’ priority,
it can overtake them. With a payment of $5, job 2MD is
”cheaper” to compensate and thus 3MD is scheduled on
machine MD 2 in front of 2MD. The execution of job
1MD on MD 1 and 3MD on MD 2 will start in the following
period r = 1. The next job being generated is 4MD by
camera 4 which shifts job 2MD at MD 2 further back in the
waiting queue due to its higher priority and consequently
pays $5 in compensation. MD 2 is chosen instead of MD 1
since it can guarantee an earlier completion time. Two new
jobs emerge at time r = 2, another MD-job at camera 5 as
well as the first BD-job 3BD arising as a result from the already
completed MD-job 3MD. Camera 5 chooses MD 1
since no compensation payments are due, while the completion
time is the same as for MD 2 . 3BD can choose from
the two empty BD-machines (cf. table 3). Concerning the
two BD-jobs 1BD and 4BD being generated in r = 3,
1BD is considered first and placed at the empty machine
BD 2 . It is, however, immediately being displaced by the
higher-prioritized job 4BD which gains more utility from
the earlier completion time feasible at BD 2 than from waiting
at BD 1 even though it has to compensate 1BD with
$5.
6. Limitations and Possible Extensions
Despite some striking features outlined in section 4,
DLGM also suffers from drawbacks. In this section, we
show a few of the identified limitations of the basic mechanism
and point to extensions that may alleviate the problems.
First, DLGM does not fully account for the interorganizational
nature of Grids; strategic and self-interested
behavior is modeled on the demand side only. Jobs have private
information about their type and may misreport about it
in hope for a utility gain. The supply side, i.e. the machines
offering their service for the execution of jobs, is assumed
to be ”obedient”. Since jobs are the ones determining the
allocation by choosing their favorite location for execution,
the machines’ action space is restricted to (truthfully) reporting
tentative completion times and payments to the jobs
upon request. Neither are they granted the right to select the
jobs they want to execute themselves, nor does the design
of the mechanism leave room for strategic behavior in terms
of misstating completion time and/or payments. Moreover,
with the basic setting, machines do not get any reward for
contributing to the system. Their behavior is assumed to
be beneficial and altruistic. Consequently, DLGM is only
suitable for scheduling scarce resources that are under centralized
control. The design of a mechanism which accounts
for strategic behavior on both sides of the market serves as
a goal for future research.
Another weakness of DLGM becomes observable when
taking a closer look at the example provided in the previous
section. Since job 2MD queues first at machine MD 2 , it
does not displace any present job and this exempts it from
having to make an immediate payment. However, it subsequently
receives compensation payments amounting to $10
from the overtaking jobs 3MD and 4MD. 2MD could
thus have exploited the payment scheme by adding unnecessary
computing instructions at the end of its regular code
or even creating a whole artificial job with entirely senseless
computations. A relatively simple remedy for preventing
jobs from exploiting the mechanism to earn money may
be to add a variable cost component to the payment scheme.
This can be done by introducing z i ∈ R + which equals the
operating costs per timestep for machine i. When extending
the scenario to feature strategically acting machines which
can misreport about Ĉj and ˆπ j , machines may be allowed to
individually report their operating costs z i . To incorporate
the operating costs in the mechanism, the tentative payment
ˆπ j has to be extended:
∑
ˆπ j = ˜p j (˜z i +
˜w k ).
k∈ L(j),k→i,kr j
The introduction of z (respectively ˜z) does on the one side
have a positive effect on jobs. They are provided with an
incentive to claim only as much processing time as needed
since every unit of useless computation now costs money.
On the other hand, it becomes possible to at least cover
the expenses the machines incur for executing the jobs and
computing the priority values, (tentative) completion times
and (tentative) payments as well as the continuous interjob
compensations. Assuring incentive compatibility on the
supply side is yet another issue to be tackled in future work.
169

Allowing preemptiveness of jobs may further increase
the efficiency of the mechanism. In DLGM in its current
form, a job which is being processed at a machine may not
be interrupted until it is completed. Grid applications like
the video surveillance scenario demand for the possibility to
freeze the execution of a running job if a job with a higher
priority enters the queue. As long as a job can be stopped
at any point in time, the procedure of the mechanism needs
to be adjusted only slightly: When a job j asks a machine
for Ĉj and ˆπ j , the machine also considers the priority value
of the currently executed job l . If the priority of the incoming
job is higher, i.e. ˜w j /˜p j > ˜w l /˜p l , the current job l is
“frozen”, its processing time set to the remaining processing
time, and it is put back into the queue. Enabling suspension
of jobs requires additional memory space for saving the
preliminary results of a job; they are transmitted to the job
initiator only when the job is completed to avoid communicational
overload. The formula for computing the tentative
completion time of a job j has to be changed to
∑
Ĉ j (i) = ˜r j +
˜p k + ˜p j .
k∈ H(j),k→i,k

in a scalable computing cluster. IEEE Trans. Parallel Distrib.
Syst., 11(7):760–768, 2000.
[3] A. AuYoung, B. Chun, A. Snoeren, and A. Vahdat. Resource
allocation in federated distributed computing infrastructures.
In Proceedings of the 1st Workshop on Operating
System and Architectural Support for the Ondemand IT InfraStructure,
October 2004.
[4] R. Bapna, S. Das, R. Garfinkel, and J. Stallaert. A Market
Design for Grid Computing. INFORMS Journal on Computing,
forthcoming, 2006.
[5] B. N. Chun and D. E. Culler. Market-based proportional
resource sharing for clusters. Technical report, Berkeley,
CA, USA, 2000.
[6] I. Foster, C. Kesselman, J. Nick, and S. Tuecke. The physiology
of the grid: An open grid services architecture for
distributed systems integration. Technical report, Globus
Project, www.globus.org/research/papers/ogsa.pdf, 2002.
[7] I. Foster, H. Kishimoto, A. Savva, D. Berry, A. Djaoui,
A. Grimshaw, B. Horn, F. Maciel, F. Siebenlist, R. Subramaniam,
J. Treadwel, and J. V. Reich. The open grid services
architecture, version 1.5. Technical report, Open Grid
Forum, July 24 2006.
[8] I. Foster, T. Maguire, and D. Snelling. OGSA WSRF basic
profile 1.0. Technical report, Open Grid Forum, May 1 2006.
[9] R. L. Graham, E. L. Lawler, J. K. Lenstra, and A. H. G. R.
Kan. Optimization and approximation in deterministic sequencing
and scheduling theory: a survey. Annals of Discrete
Mathematics, 5:287–326, 1979.
[10] B. Heydenreich, R. Müller, and M. Uetz. Decentralization
and mechanism design for online machine scheduling.
In L. Arge and R. Freivalds, editors, SWAT, volume
4059 of Lecture Notes in Computer Science, pages 136–147.
Springer, 2006.
[11] B. Heydenreich, R. Müller, and M. Uetz. Games and mechanism
design in machine scheduling : an introduction. 2006.
[12] J. Joseph, M. Ernest, and C. Fellenstein. Evolution of grid
computing architecture and grid adoption models. IBM Systems
Journal, 43(4):624–645, 2004.
[13] K. Lai, B. A. Huberman, and L. Fine. Tycoon: A distributed
market-based resource allocation system. Technical report,
April 5 2004.
[14] O. Regev and N. Nisan. The popcorn market - an online market
for computational resources. In ICE ’98: Proceedings of
the first international conference on Information and computation
economies, pages 148–157, New York, NY, USA,
1998. ACM Press.
[15] E. Rosenberg and A. Eilat. Video surveillance grid application,
version 1.0. Technical report, Correlation Systems
Ltd., December 17 2006.
[16] B. Schnizler, D. Neumann, D. Veit, and C. Weinhardt. Trading
grid services - a multi-attribute combinatorial approach.
European Journal of Operational Research, forthcoming,
2006.
[17] C. A. Waldspurger, T. Hogg, B. A. Huberman, J. O. Kephart,
and W. S. Stornetta. Spawn: A distributed computational
economy. IEEE Trans. Softw. Eng., 18(2):103–117, 1992.
171

Exploiting the Knowledge Economy: Issues, Applications, Case Studies
Paul Cunningham and Miriam Cunningham (Eds)
IOS Press, 2006 Amsterdam
ISBN: 1-58603-682-3
Self-Organizing ICT Resource Management
– Policy-based Automated Bidding
Dirk NEUMANN
Institute of Information Systems and Management (IISM), Universität Karlsruhe (TH),
Engler Str. 14, 76131 Karlsruhe, Germany
Tel: +49 721 608 8377, Fax: + +4 721 608 8399, Email: Neumann@iism.uni-karlsruhe.de
Abstract: For decades markets have been proposed for allocating computer
resources in distributed systems. None of these approaches has made it into practice.
One reason for the adoption failure that has been rarely discussed is the bidding
process. Theoretic approaches assume that the bidders know their bids exactly. This
includes the specific resource, which is needed at some time in the future, as well as
the price. This assumption simplifies reality and can thus not contribute to the
development of Grid markets. What is needed to establish a prospering Grid market
are rules how to conduct the bidding. Since demand & supply are dynamic, manual
bidding is too slow to accommodate abrupt shifts. This paper introduces a policy
based autonomous agent approach for automated bidding. By means of the policies
resource providers and consumers can specify how Grid resources are being traded.
1. Introduction
The vision of a complete virtualization of Information and Communication Technology
(ICT) infrastructures by the provision of resources like computing power or storage over
Grid infrastructures will make the development of Open Grid Markets necessary. Over the
Open Grid Market idle or unused resources (e.g. computational resources) can be supplied
(e.g. as services) and client demand can be satisfied not only within organization but also
among multiple administrative domains.
In conventional IT environments, the decision “who shares with whom, at what time” is
orchestrated via central server that uses scheduling algorithms in order to maximize
resource utilization. Those central scheduling algorithms, however, have problems when
organizational boundaries are crossed and information about demand and supply are
manipulated. In particular, if demand exceeds supply, the scheduling algorithms fail to
allocate the resources efficiently.
The further virtualization of ICT infrastructure and services requires technical solutions
that allow for an economically efficient allocation of resources. Economic mechanisms can
be the basis for technological solutions that handle these problems by setting the right
incentives to reveal information about demand and supply accurately. Market or pricing
mechanisms foster information exchange and can therefore attain efficient allocations.
The idea of introducing markets in distributed systems is certainly not new. Despite
previous efforts, none of the market-based approaches has made it into practice. The
reasons can briefly be summarized by the following three arguments: (1) inadequate market
design, (2) insufficient support to use the markets, and (3) improper coupling between the
market and state of the art middleware.
The SORMA project (Self-Organizing ICT Resource Management, www.sormaproject.eu)
will implement a comprehensive Open Grid Market by addressing all three
arguments. Firstly, the economic model provides an sound market structure. Secondly, the
self-organization model deals with interaction between the Grid-application and the market
Copyright © 2006 The Authors 172

y providing intelligent tools. Thirdly, the economic middleware model builds the bridge
between the self-organization, the economic model, and state of the art Grid infrastructure.
Integrating those three models into the SORMA system, the Open Grid Market is expected
to take off in practice and help realizing the benefits of Grid technologies. Potential target
groups of the Open Grid Markets range from server farms, over large companies to SMEs,
which strive for lowering the total cost of ownership for ICT resources.
The design of Open Grid Markets and related issues attempting to move business to
Grid are currently hotly debated under the label Grid Economics. In this context, this paper
concentrates on the self-organization model and shows how policies can be used to
automate the bidding process. This paper is unique as it proposes a novel approach on
bidding in Grid and reconciles research on Grid markets and on technical middleware.
The remainder of the paper is structured as follows. In section 2 the SORMA system,
enabling self-organized ICT resource management is illustrated. In section 3, the policybased
working of the bidding component is showed. Section 4 concludes with a summary
and a brief outlook.
2. The SORMA System
The lessons learned from the previous efforts to establish markets in distributed system
gave rise to the SORMA project. The overall objective of SORMA is the development of
methods and tools for establishing an efficient market-based allocation for ICT resources in
a more efficient way in order to enable resource accessibility for all users and to increase
user’s satisfaction, profit and productivity.
Grid Application
Payment Interface
Open Grid
Market
Resource
Monitoring
Clearing &
Billing
Security
Contract
Management
Market
Connector
Intelligent Tools
Economic Grid Middleware
Grid Market
Directory
Bidding
Component
Resource
Decomposition
Accounting
Resource
Auditing
Grid
Middleware
Extended
Resource
Management
Common Virtualization Middleware
(Globus, UNICORE,...)
Resource Fabrics
Figure 1: The SORMA System
SORMA uses concepts from autonomic computing [2, 9], such that the ICT resource
allocation process is automatically and autonomously conducted. Once the bidding process
is virtualized, the ICT system of an organization will self-organize its resource
management: overcapacity is offered on the market, while undercapacity initiates a buying
process. Markets and self-organization go thus hand in hand.
The SORMA system combines all those areas of self-organizing ICT resource
management in a holistic way. This is reflected by the high-level architecture, which
consists of the following building blocks: Open Grid Market, Intelligent Tools and
Economic Grid Middleware (Figure 1). In the following, the focus is on the bidding
component. Since bidding without a market is meaningless, the architecture of the Open
Grid Market is presented.
Copyright © 2006 The Authors 173

2.1. Open Grid Market
The Open Grid Market consists of the service consumers and providers, representing
demand and supply, and of the software system that implements the market functionalities.
Service consumers are those entities that request specific services. For example, a bionumerical
simulation application requests computation and storage services. As it is the
goal to have a self-organizing ICT resource management, the service consumer is
considered to be the specific Grid application, which provides runtime data of the
requestor’s system. Based on this data, service requests are initiated including detailed
service type requirements (e.g. storage service) and quality of service constraints (e.g. >
300GB free space for 4 hours). It is assumed that approximate quality and time constraints
of the requested services can be determined using prediction models [e.g. 1].
On the service provider side, computer nodes host a set of Grid services (e.g.
computation service or storage service) and provide information about their current quality
characteristics and availability. The node includes a resource manager that provides runtime
data about scheduling, allocating, and monitoring of local resources. This resource manager
makes use of prediction models in order to determine the approximate future availability of
the underlying computational resources.
The Open Grid Market component aggregates the bids from service providers and users
into an allocation of resources and corresponding prices. Crucial for the working of the
Open Grid Market is the clearing mechanism component, which implements the logic how
the market is resolved (i.e. auction formats). In the SORMA framework, the clearing
mechanism can be arbitrarily defined. The MACE auction is deemed promising for a use in
Grids, as it allows service requesters and providers the simultaneous submission of bids on
heterogeneous services expressing substitutes and complements [8]. Bids, encoded as WS-
Agreement offers, are submitted to an order book. The auction determines an allocation and
the corresponding prices either regularly or continuously.
The objective of the allocation process in MACE is the maximization of social welfare,
i.e. the difference between the buyers' valuations and the sellers' reservation prices. The
problem is formulated as a linear Mixed Integer Program and, thus, can be solved by
standard optimization solvers. The outcome of the winner determination model is allocative
efficient, as long as buyers and sellers reveal their valuations truthfully. The incentive to set
bids according to the valuation is induced by an efficient pricing mechanism. MACE
implements a k-pricing scheme and as such determines prices for buyers and sellers on the
basis of the difference between their bids. For instance, presume a buyer wants to buy a
computation service for 5 and a seller wants to sell a computation service for at least 4. The
difference between these bids is ζ =1, i.e. ζ is the surplus of this transaction and can be
distributed among the participants. This schema can be applied to MACE and results in an
approximate efficient outcome [8].
2.2. Bidding Component
When describing the Open Grid Market, it was mentioned that service consumer and
providers submit bids that express their demand and supply situation. This implies that the
Grid applications and the computing nodes have the capabilities to determine reasonable
bids. In Grid markets generating reasonable bids can become very complex for several
reasons: For example demand and supply fluctuates dynamically and is in most of the cases
unknown, valuations are unknown and bidding strategies have not yet developed.
The bidding component in SORMA consists of two elements: a Bidding Agent that
generates bids and forwards them to the clearing mechanism and an Administration
Interface that allows the control and configuration of the agent [7].
Copyright © 2006 The Authors 174

• The Administration Interface specifies user policies. Policies are mostly static and do
not change frequently. They define general rules under which condition, when and how
bids are being generated. These policies are either entered manually via a human control
interface or imported from other sources in a company.
• A Bidding Agent implements a bid generation process, which is based on policies as
well as runtime data. It is also responsible for submitting bids to the market mechanism.
The agent interacts with the Open Grid Market in order to gain auction feedback (e.g.
current highest price) which influences its bidding decisions. The agent specifies all
attributes of an order (e.g. price and bidding time) autonomously.
The Bidding Agent integrates policies as well as runtime information via an ontology.
Ontologies are logic-based representation languages that executable calculi. This is required
by the agent in order to query and reason about this information during run-time. To capture
system information as well as policies we rely on the Web Ontology Language (OWL) and
a decidable fragment of the Semantic Web Rule Language (SWRL) defined in [6].
3. Policy-based Bidding
This section introduces the reasoning behavior of the bidding agent, which autonomously
derives one or several bids based on system and market information. In essence, the
generation is guided by policies. Policies are considered as declarative rules defined by the
user to adapt and control the behavior of the system. Policies are thus adequate to represent
the logic (i.e. the business model) for bidding. Policies are specified by the administrator
using the human control interface. The bidding component receives system information
from the local resource manager of an application or computational nodes.
By means of policies this system information is interpreted and appropriate bidding
activities (e.g. submission or manipulation of bids) are initiated. The general bid generation
process performed by the bidding component is distinguished into four steps (c.f. Figure 2).
Demand
Curve
Market
Information
Demand
Curve
Grid
Ontology
System Information
Monitoring
Attribute
Determination
Bid expansion
Bid specification
Bidding
Policies
Pricing, QoS
Policies
Expansion
Policies
Bid Generation Process
User Policies
Figure 2: Bid Generation Process [7]
In a first step, monitoring information, derived from the resource manager, is explored
to assess whether a bid is submitted. This decision is based on (1) the current and on the
predicted utilization of the resource or (2) on the expected demand of the application and on
the preferences of the user.
In case demand or supply situation requires the submission of one or more bids, the
concrete properties of the bids are determined by taking information from the market
mechanism, as well as utilization, and user policies into account (second step).
In a third step, the bid is expanded by means of a domain ontology that contains
taxonomical and composition information. This step assures that different levels of
abstraction in descriptions of bids do not lead to inefficiencies in the market. In addition, it
reflects the fact that several Grid services can be offered and requested within one order.
Finally, before forwarding to the market the bid is serialized using a WS-Agreement
template to assure interoperability with the market mechanism.
Copyright © 2006 The Authors 175

3.1. Information Gathering
The bid generation process is triggered by a decision event. This decision event depends on
the actual and expected utilization of the ICT resource or the demand of the application,
respectively. The functions, which represent either the availability of the resources or the
demand of an application are denoted as load curves D(t). Load curves are obtained from
the resource manager by means of prediction models.
Based on the load curves and on the bidding policies defined by the user a decision is
made. For example, a simple bidding policy may specify that a bid must be submitted to the
Open Grid Market in case the demand is predicted to exceed a certain threshold υc within
the interval t 0 to t 1 . That means, a bid has to be submitted in case D(t) > υc holds at least for
one t with t 0 < t < t 1 . The policy specifying if a bid b has to be generated within a certain
planning interval p based on threshold value υ as well as the overall capacity c:
issueBid(?b,?p) ← PlanningInterval(?p), lowerBound(?p,?t1), upperBound(?p,?t2), Time(?t),
load(?t,?dt), greaterThan(?t,?t1), greaterThan(?t2,?t),hasThreshold(?b,?v), hasCapacity(?b,?c),
multiply(?v,?c,?e), greaterThan(?dt,?e)
The length of the interval should be determined according to the minimal slot that can
be requested from one single service provider or that can be provided to a single requester,
respectively. Long-range planning intervals may thereby entail inaccuracies due to load
changes during the interval. Thus, it is referred to mean values. For each interval this step
will be executed only once. Moreover, machine learning approaches can be applied in order
to adapting υ with respect to changing environments.
3.2. Bid Formulation
After the bid formulation has been initiated by the monitoring step the attribute values for a
concrete bid have to be determined. Within each bid the type of Grid service that is required
or offered, quality of service criteria that have to be fulfilled, a time interval in which the
resource is offered/requested, and a range of prices which are acceptable (e.g. maximal or
minimal price) has to be specified in a way that it is optimal with respect to the goals of the
agent. These goals are influenced by three aspects: The user preferences reflected by the
system policies, the system monitoring information, and the current market information. In
the following we show how these three information sources can be integrated in order to
derive attribute values for the individual attributes.
Static Storage Service
Storage Service
Dynamic Storage Service
f,g
1
HDD Service DVD Service RAM Service
0
0
Figure 3: Storage Service Ontology and Eagerness Policies
l(t 0 ,t 1 ), t 0 -t
Service Functionality: The service functionality determines which service is provided
by the computational node, or requested by an application. This information is directly
derived from the resource manager and can be modeled by referring to the corresponding
concept of a Grid ontology, which is based on the JSDL and domain-specific extensions as
shown for the storage service example in the left panel of Figure 3. In a subsequent step, the
relations between the different concepts in the ontology are exploited to improve the
matchmaking of services in the market.
Price: In literature, it is frequently assumed that bidders know their valuations. From
these valuations bidding strategies are derived that are either static such as so-called Zero-
Intelligence bidders or highly adaptive. Unfortunately, for Grid services it is difficult to
Copyright © 2006 The Authors 176

make this assumption that the valuations are known. Thus, we have to refer to a
constructive way to first approximate the valuations.
In our approach we assume that the price for bids is determined by three factors: (a) the
load curve and the resultant eagerness factor, (b) the current price of similar services in the
market and finally (c) the pricing policies of the user.
(a) In order to determine the load that is required from the application within the considered
time frame [t 0 ,, t 1 ] the equation l( t , t ) = max ( D( t)
) is applied to ensure that the demand for
0
1
[ t , t ]
t∈
0 1
that entire period can be fulfilled. For the provider, the minimum is vice versa used to
avoid overstraining the resource.
The eagerness factor µ specifies how quickly consumers or providers will make
concessions in the price. Thus, the eagerness factor can be represented by the equation µ
= f(l(t 0 ,t 1 )) + g(t 0 – t), where the functions f and g can be considered as eagerness
policies, which assign to each load D(t) and each time span t 0 - t a corresponding factor
µ. The right panel of Figure 3 reflects common eagerness policies for the storage
example, which rests on the intuition that the pressure to identify a corresponding
request increases with growing load and decreases with shortening remaining time. The
range of the eagerness factor is defined between zero and a positive rational number R.
According to the policies in our running example this might lead to f(300GB) = 0.3 and
g(60s-0s)=0.6. Consequently, we get an eagerness factor of µ=0.9.
(b) The eagerness factor µ is now used to adapt the current market price m of a service with
the same functionality and similar quality characteristics. This is achieved by means of
the following equation, where p represents the proposed price: p = m µ. Note that in
case µ is greater than one, the price of the bid might surpass the current market price. In
our running example this leads to 90% of the current market price.
However, since the price depends heavily on the quality of the service and it is not
always possible to find services with similar quality characteristics in the market, there
is a need for calculating prices for different quality levels. How this can be done based
on user policies is presented in [3].
(c) Assuring that the price does not adopt any undesired value, the provider might
additionally specify pricing policies, which define the valid range for the service prices.
Those pricing policies are not limited to linear prices but can also discriminatory prices
to skim off consumer surplus of heterogeneous users.
In this paper it is referred to rather simple pricing policies. The pricing policies (b) and
(c) can be expressed by the following statements:
price(?s,?p) ← Service(?s), minPrice(?s,?p min ), maxPrice(?s,?p max ), marketPrice(?s,?m),
hasEagernessFactor(?µ), multiply(?m,?µ,?p), greaterThan(?p,?p min ), smallerThan(?p,?p max )
price(?s,?p) ← Service(?s), minPrice(?s,?p), marketPrice(?s,?m), hasEagernessFactor(?µ),
multiply(?m,?µ,?e), smallerThan(?e,?p)
price(?s,?p) ← Service(?s), maxPrice(?s,?p), marketPrice(?s,?m), hasEagernessFactor(?µ),
multiply(?m,?µ,?e), greaterThan(?e,?p)
Quality of Service (QoS): The service offers and requests need to specify the guaranteed
or requested QoS level, since this information is crucial for the subsequent agreement.
However, QoS attributes are highly domain dependent and thus may vary from one service
to another. Our notion of QoS includes attributes ranging from the capacity of a storage
service to the response time or error rate of a computational service.
Information about quality requirements or quality guarantees can be automatically
derived from the computing node or from the application, respectively. For example, as
indicator for the capacity that can be offered or is required in the time interval [t 0 ,, t 1 ] the
value l(t 0 ,, t 1 ) can be used. For this purpose we use the formula cap = φ l(t 0 ,, t 1 ), where the
rate φ allows systematic over-/under-selling or over-/under-buying. In many scenarios φ is
required; especially when contractual penalties are high (typically φ < 1) or working to
Copyright © 2006 The Authors 177

capacity (typically φ > 1) is important. In case of a storage provider a rate of φ = 0.95
might be reasonable to reduce the risk of overselling. This means in the example we would
offer storage capability of cap = φ l(t 0 ,, t 1 ) = 0.95 * 300GB = 285GB.
Time Interval: Since the time interval [t 0 , t 1 ] represents the minimal period which can
be sold to a single consumer or bought from a single service provider, the interval [t 0 , t 1 ]
should also be the slot that is offered or requested in the bid. Moreover, the market
mechanism being used within the framework supports aggregation of quality criteria and
time. This means, several time slots can be aggregated in order to reach an agreement in the
market. Hence, the issue of choosing slots that are too short is not relevant in this context.
3.3. Bid Expansion
Having generated the initial bid, the aim of this section is to improve the probability for
finding a suitable request in the market by a semantic pre-processing of requests: On the
one hand, a Grid service taxonomy is used to introduce alternative bids that can also be
used to describe the corresponding service (vertical bid expansion). On the other hand,
equivalence relations explicitly modelled in an ontology are used to address requests that
could also be fulfilled by a combination of different services (horizontal bid expansion).
Vertical bid expansion: Since offers and requests might be defined on different levels
of abstraction pure syntactic matching may fail. Typically offers can be described very
detailed since providers know exactly what kind of service they provide (e.g. Hard Disk
Service), whereas customers usually define their requests on a rather abstract level (e.g.
Storage Service). Consequently, in the market there will be no match between Hard Disk
Service and Storage Service since the terms
are compared only using strings. Therefore,
we employ a Grid service ontology (c.f.
Figure 3) that serves as background
knowledge for adding additional XORrelated
bids. By introducing these alternative
bids according to Algorithm 1 we can
overcome the different levels of abstraction
while avoiding inappropriate matches.
Figure 3 shows a taxonomy for storage services which is used to exemplify how bids
can be vertically expanded. Note that since we use formal ontologies the taxonomic
relations either are modelled explicitly or they are derived by means of automatic
classification done by a reasoner. According to Algorithm 1 in case of offers we introduce
an additional XOR-related bid for each concept that is passed when traversing to the root of
the hierarchy. For instance, a HDD Service offer is expanded by adding an offer for a Static
Storage Service and an offer for a Storage Service. This is necessary since a HDD Service
also fulfills the functionality of a Static Storage Service and a Storage Service and thus
should be matched with a request for the latter. In case of a request we have to traverse to
the leafs of the taxonomy since more specific services also fulfill the functionality of the
more general services, e.g. a HDD Service meets the requirements for a Storage Service
request. However, since this expansion could lead to a high number of alternative bids it
might be necessary to limit the depth of such an expansion, which can be done by means of
an expansion policy.
Horizontal bid expansion: Often a requested Grid service can not only be fulfilled by
one single service, but also by a combination of different services. To account for this, the
Grid ontology allows expressing composite services and equivalence relations: A service is
either atomic or composite. Further, an equivalence relation can be defined between
different services. This allows the explicit specification which services provide the same
functionality. By means of the equivalent_to and contains relations new alternative bids can
Copyright © 2006 The Authors 178

e introduced automatically to improve the probability or finding a suitable match in the
market. For example, a mainframe computing service containing computation and storage
can be expanded by adding a XOR-related bundle request the two constituting services [4].
4. Conclusion
The presented SORMA project primarily aims at promoting the widest and best possible
use of Grid applications and services. This is attempted by devising incentive mechanisms
that ICT resources are offered and requested in Grid. This requires that the service
providers are compensated for offering their ICT resources by the service consumers. It is a
premise of this paper that markets can undertake this task better than any other coordination
mechanism. This is especially the case if the bidding process is fully automated, such that
the market itself will handle any unforeseen problem in the resource management process.
By bringing autonomic and Grid computing as well as economic principles closer
together, market-based approaches have the potential to achieve an economy-wide adoption
of Grid technology and will thus leverage the exploitation of potential Grid technology
offers. To achieve this potential, many Grid Economics related obstacles must be
overcome: markets can only work well if bids accurately reflect demand and supply. If this
is not the case, the market price cannot reflect the correct situation. Thus, the market price
looses its capacity to direct scarce resources to the bidders who value them most.
This paper is unique by recommending an autonomic computing approach for bidding
in Grid markets. By defining policies the business models of the service consumers and
providers can be represented. The proposed bid generation process is currently quite simple
and referring to observations and straightforward rules.
As the next steps, the bid generation process will be implemented and coupled with
available Grid markets and middleware extensions. Subsequently, validation and economic
testing will be performed, gradually improving the bid generation process (incorporating
minimum probabilities of receiving the resources).
References
[1] Kee, Y.-S., H. Casanova, and A. Chien. Realistic Modeling and Synthesis of Resources for
Computational Grids. in ACM Conference on High Performance Computing and Networking. 2004.
[2] Kephart, J.O. and D.M. Chess, The Vision of Autonomic Computing. Computer, 2003. 36(1): p. 41-50.
[3] Lamparter, S., A. Eberhart, and D. Oberle. Approximating Service Utility from Policies and Value
Function Patterns. in International Workshop on Policies for Distributed Systems and Networks. 2005:
IEEE Computer Society.
[4] Lamparter, S. and B. Schnizler. Trading Services in Ontology-driven Markets. in 21st Annual ACM
Symposium on Applied Computing. 2005. Dijon, France.
[5] Ludwig, H., A. Dan, and B. Kearney. Cremona: An Architecture and Library for Creation and
Monitoring of WS-Agreements. in 2nd International Conference on Service Oriented Computing
(ICSOC 2004). 2004.
[6] Motik, B., U. Sattler, and R. Studer, Query Answering for OWL-DL with Rules. Journal of Web
Semantics: Science, Services and Agents on the World Wide Web, 2005. 3: p. 41-60.
[7] Neumann, D., S. Lamparter, and B. Schnizler. Automated Bidding for Trading Grid Services. in
European Conference on Information Systems (ECIS). 2006. Goeteborg, Sweden.
[8] Schnizler, B., D. Neumann, and C. Weinhardt. Resource Allocation in Computational Grids - A Market
Engineering Approach. in WeB 2004. 2004. Washington.
[9] Sterritt, R., et al., A concise introduction to autonomic computing. Advanced Engineering Informatics,
2005. 19(3): p. 181-187.
Copyright © 2006 The Authors 179

A Decentralized Online Grid Market – Best myopic vs. rational response
Dirk Neumann 1 , Nikolay Borissov 1 , Jochen Stoesser 1 ,
and Simon See 2
1 Institute of Information Systems and Management, University Karlsruhe (TH), Germany
2 Sun Microsystems, Singapore
{neumann, stoesser, borissov}@iism.uni-karlsruhe.de; simon.see@sun.com
1. Introduction
Grid computing is a new computing paradigm where processing is distributed across locally dispersed
infrastructures. By connecting many heterogeneous computing resources, virtual computer architectures are
created, increasing the utilization of otherwise idle resources [1]. Grid computing extends the traditional parallel
and distributed computing paradigms as it deals with resources belonging to different administrative domains.
One of the key problems in Grids is the scheduling of jobs, i.e. the decision about what jobs are to be
executed on which machines at what time. Grid computing is hampered by the fact that different administrative
units are involved. Priorities of jobs are communicated by the respective agents, thus opening up possibilities for
strategic misreporting of true job priorities. In these heterogeneous and decentralized settings, markets are
supposed to work well as the pricing mechanisms of markets may force the participants to reveal their true
priorities [2]. Looking at existing market mechanisms, however, there are currently only few mechanisms which
perform real-time scheduling for Grid applications, while at the same time appropriately accounting for the
arising economic design challenges.
To this end, the contributions of this paper are threefold. Firstly, we present the Decentralized Local Greedy
Mechanism (henceforth “DLGM”) by Heydenreich et al. [3], a mechanism from the general machine scheduling
domain which exhibits several desirable properties, and propose its use for Grid environments. Secondly, the
theoretical analysis of DLGM is based on the assumption of simple agents behaving according to a so called
“myopic best response strategy”. By means of a simulation with learning agents, we show that an analysis using
myopic best response strategies overestimates the performance of the mechanism and is thus not an appropriate
solution concept for modeling real-world market-based scheduling mechanisms. As a byproduct of this analysis,
we illustrate the performance benefits of market-based schedulers as opposed to purely technical schedulers
which are solely based on system-centric measures. Thirdly, we point at limitations of the mechanism for the
practical use and suggest remedies that may help to mitigate these drawbacks.
1
180

This paper is structured as follows. In Section 2, we present related work on the design of Grid market
mechanisms. In Section 3, we present the DLGM mechanism and the concept of myopic best response equilibria
used to analyze its economic worst-case performance. In Section 4, we model agents with learning capabilities,
and contrast the DLGM for myopic best response agents with agents with learning capabilities by means of a
numerical simulation. In Section 5, we point to limitations and possible extensions of the basic DLGM
mechanism, before we conclude the paper in Section 6.
2. Related Work
Economic scheduling mechanisms can be distinguished into two classes according to their mode of
allocation: mechanisms which execute periodically, so called “offline mechanisms”, and mechanisms which
execute continuously, so called “online mechanisms” [4].
In classic mechanism design, offline mechanisms have received the most interest. Examples comprise the
simple Vickrey auction, English auction, Dutch auction, and double auctions [5, 6]. However, many modern
applications in E-Commerce, and in particular in Grid scheduling, involve complementarities between goods.
Examples of the resulting combinatorial mechanisms are MACE [7, 8], Bapna et al [9], and Stoesser et al [10,
11]. All these mechanisms involve a scheduling problem that is NP-hard.
Online mechanism design also allows agents to enter or leave the mechanism dynamically. Parkes et al.
propose an NP-hard online Vickrey-Clarke-Groves (VCG) mechanism [12]. In a subsequent work, Parkes and
Singh provide an approximation that retains the mechanism’s truth-telling property [13]. The computational
complexity of such mechanisms drives researchers and practitioners to look at simpler allocation models which
can be adapted to real-world scenarios and requirements. Prominent examples are proportional share
mechanisms [14], where the users receive a share of the computer resource proportional to their valuation’s
fraction of the overall valuation across all users. A related online mechanism is the so-called pay-as-bid
mechanism proposed in [15]. In pay-as-bid mechanisms, the agents submit only a single real value as bids which
at the same time make the agents’ final payments.
Heydenreich et al. propose a Decentralized Local Greedy Mechanism (DLGM) [3]. DLGM is a promising
economic mechanism for scheduling interactive Grid applications due to its polynomial runtime, budgetbalanced
payments and incentive compatibility. In the following section, we will introduce the DLGM
mechanism in more detail as we will use it to contrast the game theoretic solution concept of myopic best
response strategies with a learning-based approach.
2
181

3. A Decentralized Local Greedy Mechanism
3.1. The Setting
In the following, a Grid environment consists of a number of individual agents, which are assumed to act
rationally in the sense that they strive for maximizing their individual benefit from participating in the Grid.
Agents submit computational jobs to the Grid, where a job j is specified by the tuple (its “type”) t j = (r j , p j ,w j ).
The job’s release date is denoted by r j ∈
+
+
R o , the job’s processing time by p j ∈ R , and the agent’s cost of
+
waiting for one additional unit of time until the job is finished (its “weight”) by w j ∈ R
. The quasi-linear utility
function of agent j is defined to be
+
u ( C , π | t ) = −w
C − π , where C j ∈ R , C j ≥ p j , marks j’s completion
j
j
j
j
j
j
j
time, and π
j
∈
+
R o is j’s payment. When requesting resources for the execution of job j, the agent may
~
t =
~ r ,
~
p , w
~
≠ t
strategically choose to submit the request ( )
j j j j j
to the Grid. In the following, we will use thze
terms “job” and “agent” interchangeably. We consider a decentralized setting where the action space of agent j
~
does not only consist of reporting its type t j , but also of subsequently choosing a machine at which to queue.
We will henceforth denote agent j’s action, or strategy, by s j , and we will also write ( s,
t)
for agent j’s utility
given the strategy vector s = ( s ,..., ) and the true types t ( t ,..., )
1
s n
= .
The Grid infrastructure consists of m identical, parallel machines M = {1, ..., m}. The strategic analysis of
this setting is restricted to strategic behavior on the demand side only. Machines are assumed to report truthfully
about their status. Each job can be executed on any machine, but each machine can execute at most one job at a
time.
1
t n
u j
3.2. The Mechanism
For this setting, we propose the use of the DLGM mechanism suggested by [3]. The mechanism aims at
minimizing the total weighted completion time Σw j C j across all jobs. DLGM comprises the following three
steps:
Step 1 – Job submission
At release date
~
r j , j reports w ~ j and
~
p j to every machine m ∈ M.
Step 2 – Real-time planning
Based on this information, the machines perform real-time planning based on a local scheduling policy. Job j
3
182

has a higher priority value than k ⇔
w
~
~
p
j
j
≥
w
~
~
p
k
k
. Then j is scheduled before job k in the waiting queue.
Depending on the current local waiting queue which resulted from the previous strategies
s − j
of all other agents,
and j’s report ~ t , machine i reports the tentative completion time ˆ ~
~
)
j
C ( i | s− , t to agent j. Cˆ
( i | s
−
, t ) consists
of j’s own runtime plus the runtime of the job which is currently being executed on machine i (if any) and the
aggregated runtimes of the jobs which would be queuing in front of job j. Furthermore, the mechanism computes
a tentative payment ˆ
~
π ( i | s , t ) according to the prominent Vickrey principle. Agent j has to compensate all
j
− j
j
agents whose jobs are currently waiting at machine i and which are delayed due to j being assigned to i for their
additional waiting cost. These values are only tentative as later arriving jobs might displace job j. Consequently,
the tentative utility of agent j at machine i at time ~ ) ~ ) ~
~
r is u i | s , t , t ) = −w C ( i | s , t ) −
) π ( i | s , t )
j
Step 3 – Machine selection
j
j
j
j
(
− j j j
j j − j j j − j j
Upon receiving information about its tentative completion time and required payment from the machines, the
agent makes a binding decision to queue at a machine i, and pays ˆ
~
π ( i | s , t ) to the delayed jobs.
j
− j
j
j
j
j
.
3.3. Solution Concept and Theoretical Performance
When evaluating the performance of a market mechanism, we first need to fix a solution concept which
~
defines the agents’ behavior, i.e. their reports of t
j based on tj and their valuation for a given allocation.
In the model of Heydenreich et al. [3], agents do not remember the outcomes of earlier market interactions
but are somewhat “myopic” in the sense that they only consider the current situation. Agents choose the action
which maximizes their tentative utility, given the machines’ tentative feedback. Formally:
Definition 1 (“Myopic best response equilibrium”): A strategy profile s is called a myopic best response
equilibrium if, for all agents j ∈ J, all agent types t, all strategies
instead of s j , uˆ
(( s , s ), t)
≥ uˆ
((
~
s , s ), t)
j j − j
j j − j
.
s − j
, and all strategies
~
s
j which j could play
As shown in [3], without knowledge about the future, at time
~
rj
it is a utility maximizing strategy sj for agent
) ~
j ∈ J to report its true type t j to the system and to choose the machine i which maximizes u i | s , t , t ) . We
will henceforth refer to s j as being a “myopic best response strategy”.
A common metric for a mechanism's performance is its “performance ratio”:
(
j − j j j
Definition 2 (“Performance ratio”): Let Z A
(J)
be the objective value generated by an online mechanism A
4
183

for a given sequence J of jobs and let Z
* ( J ) be the (optimal) objective value generated by an omniscient offline
mechanism for this given job sequence. Then, for the class ψ of all possible job sequences, the performance
⎧Z A
A ( J ) ⎫
ratio of mechanism A is ρ = ρ Ψ
= sup ⎨ | J ∈ Ψ⎬
.
*
⎩ Z ( J ) ⎭
Assuming all agents play their equilibrium strategy, DLGM achieves a performance ratio of ρ = 3.281.
4. Evaluation
In this section, we examine two main questions which arise from the theoretical results in [3]:
Myopic best response vs. rational response: Given agents with learning capabilities, does the concept of
myopic best response equilibria adequately model agent behavior? Or do learning capabilities lead to some other
form of equilibrium, which may be called a “rational response equilibrium”?
Worst-case vs. average case: What is the average case performance of the DLGM as opposed to the worstcase?
How does it compare to FIFO as a purely technical allocation mechanism, and how does it change if we
introduce agents with learning capabilities?
4.1. Data Generation
For simplification, we assume that agents can only misreport about their weight w . As presented above, one
j
central result is that myopic best response agents then truthfully report
w
~
= w .
j
j
Our aim is to simulate rational agents with learning capabilities instead which may strategically misreport
about
w
j
based on their past behavior and payoffs. We will refer to these strategies as “rational response
strategies”. To model the learning capabilities of these agents, we chose a reinforcement learning-based
approach with a greedy selection policy [16]. This approach comprises the following two phases:
• Exploration phase: In this phase, the agents simultaneously explore the strategy space and heuristically try
to learn the expected payoffs of various strategies to finally decide on an expectedly optimal strategy. Within
this exploration phase, we perform 25 runs of the market. In each run k, every agent j reports
k
w ~ j
and p to
j
the market at time r j
and, upon having received the feedback about its tentative completion time and
) ~
payment from each machine, selects the machine i which maximizes its tentative utility u i | s , t , t ) ,
where
k
w ~
j
(
j − j j j
is an integer which is randomly drawn from the uniform distribution with support [0, 2w j ] in order
to model both the under- and overstating of weights. Subsequent to each run k, agent j determines its actual
5
184

k
ex-post utility (i.e. its “reward”) u ((
~
s , s ), t)
for round k. After all 25 runs, agent j greedily determines its
j j − j
best strategy * *
k
s as s = arg max u ((
~
s , s ), t)
. Since the strategy space of agent j consists of strategically
j j
j j − j
k
reporting
w ~
j
only, we can simplify this to the greedy optimal strategy being the report of
~ *
w which yielded
j
the maximum ex-post utility across all 25 runs.
• Exploitation phase: In this phase, the agents try to “exploit” the information which they obtained in the
previous exploration phase. Using a greedy selection policy, agent j assumes that, since the report of
~ *
w
j
maximized its ex-post utility given its true weight
w in the past, reporting a fraction of
j
~ for the weights
of all subsequent jobs will also maximize its future payoff. In order to examine this strategy, we also ran the
exploitation phase for 25 rounds. In each round, we generated one job per agent based on the same
distributions as in the exploration phase as specified in Table 1.
In order to compare the myopic best response strategy and the rational response strategy for the DLGM
mechanism, we ran six settings, whereas each setting comprised 5 machines. To increase the competition in the
market, across the course of these settings, we successively increased the Poisson-based arrival rate of jobs from
λ = .2 to λ = 1as listed in Table 1, which also specifies the random choices of the processing times and weights.
Each job sequence was then fed into both the DLGM mechanism with agents playing myopic best response
strategies and in the DLGM mechanism with agents playing what we call rational response strategies. Moreover,
we simulated a simple FIFO mechanism as an example of a technical scheduler to determine the benefit of
involving users into the scheduling process by enabling them to dynamically submit weights for jobs instead of
assigning static weights.
Setting S1 S2 S3 S4 S5 S6
Arrival time λ of the Poisson-based release dates r
j .2 .5 .6 .7 .8 1
Actual number of jobs 208 440 559 686 795 1050
Processing time p
j
Lognormal distribution with mean 9 and variance 3
Weights w
j
Normal distribution with mean 9 and variance 3
Table 1: Simulation settings
w *
j
w
j
4.2. Results
The results of the simulations are summarized in Table 2. As aforementioned, we tested a FIFO mechanism
as variant of a technical scheduler, the DLGM with myopic best response and the DLGM with learning agents in
six settings with varying competition. The utilization being defined as the number of jobs where the mechanism
6
185

is busy divided by the total number of time slots represents a good metric for expressing the level of competition.
As all three mechanisms do not impose reservation prices, utilization is identical for all three mechanisms. For
an arrival rate of λ = . 2 .with a job sequence of 208 jobs, the machines are only used 38 % of the time. When the
size of the job sequence increases to 559 jobs, the utilization rate grows up to 77 %, and the machines are fully
utilized for settings with more than 700 jobs.
When competition is low, technical schedulers achieve as good welfare (i.e. total weighted completion time)
as DLGM does. However, with increasing utilization, technical schedulers are starting to perform worse than
market mechanisms. When competition is extremely high, FIFO generates about 13 % higher overall waiting
costs than DLGM with agents playing rational response strategies. This result is rather straightforward, as
technical schedulers do not account for the relative urgency of jobs. When utilization is low, however, the
welfare loss incurred by technical schedulers is insignificant. In the absence of machine prices, the use of market
mechanisms for scheduling is accordingly only advantageous if excess demand for resources exist. Although
there are strong indicators that demand for Grid resources exceeds supply [17], most papers in the so-called Grid
economy overestimate the goodness of market mechanisms [18].
Furthermore, another important result of our simulation is that myopic best response is an adequate proxy for
rational bids only when utilization is low. Once competition increases, agents have an even stronger incentive to
manipulate their weights reported to the mechanism. More specifically, the jobs have an incentive to understate
their true weight, due to the fact that they have to compensate all other jobs in the queue for being scheduled
earlier. The longer the waiting queue becomes, the more expensive truthful reporting becomes, as all jobs in the
schedule needs to be compensated for their delay. This effect grows for the setting with more than 1,000 jobs, as
learning jobs report on average only 19 % of their true weight. Naturally, the DLGM looses much of its power to
prioritize jobs according to their urgency. However, when competition is high, the results of a realistic DLGM,
where jobs learn to adjust their weight reports, are still better than for technical schedulers but way less than the
concept of myopic-best response would suggest. This leads us to the conclusion that best myopic response
behavior is only an adequate concept when competition is low – in those cases, however, the use of market
mechanisms does not make sense, as technical schedulers are performing equally well.
Our simulation also yields that the fraction of jobs that receive an overall positive payment does not depend
on the learning capabilities of the jobs. When competition is low, the fraction of jobs earning positive
compensation payments is as low as 1 %, and grows up to almost 40 %, when competition gets fierce. The
employment of compensation payments opens up a way how to cheat the mechanism, as we will show in the
7
186

next section. This possibility does generally hold and is not restricted to myopic best response.
Another important result of our simulation is that the planning quality of the DLGM becomes negligible
when competition gets fierce. Essentially, the DLGM seems to be nice as it reports back a tentative completion
time. As competition becomes fierce, this tentative completion time is overly extended. The lateness metric
measures the time difference between ex-post and tentative completion time. As Table 2 shows, the lateness
extends considerably so that the tentative completion time loses its power for prediction. Jobs cannot rely on
those time reports, but should add considerably more time. Intuitively, lateness occurs in the same way for
learning jobs.
Metric Mechanism\Setting S1 S2 S3 S4 S5 S6
Utilization All mechanisms 38% 77% 91% 100% 100% 100%
%-bid of best
myopic
response weight
DLGM with learning agents 102% 90 % 65 % 25 % 22 % 19 %
Aggregated
waiting cost
Positive
Compensations
Lateness
FIFO -916,634 -1,993,694 -2,676,904 -3,633,538 -5,046,646 -8784801
DLGM with myopic best
-7042589
-916,701 -2,001,682 -2,650,274 -3,340,492 -4,370,345
response
DLGM with learning agents -916,717 -2,001,787 -2,663,577 -3,485,629 -4,676,194 -7802760
DLGM with myopic best
response
1% 7% 16% 18% 23% 34%
DLGM with learning agents 1% 8% 18% 20% 25% 36%
DLGM with myopic best
response
.002 .073 1.13 8.72 20.73 58.99
DLGM with learning agents .0021 .1056 1.198 8.6589 19.9821 58.79
Table 2: Simulation results
5. Limitations of the Mechanism and Possible Extensions
As the simulation results suggest, despite exhibiting some striking features, DLGM also suffers from various
drawbacks. In this section, we outline a few of the identified limitations of the basic mechanism and point to
extensions that may alleviate these problems.
First, DLGM does not fully account for the inter-organizational nature of Grids; strategic and self-interested
behavior is modeled on the demand side only. The supply side, i.e. the machines offering their service for the
execution of jobs, is assumed to be “obedient”. Moreover, with the basic setting, machines do not get any reward
for contributing to the system. The design of a mechanism which accounts for strategic behavior on both sides of
the market serves as a goal for future research.
Another weakness of DLGM is that jobs may be able to exploit its payment scheme. If a job does not displace
any present job, e.g. because it reported a low valuation, it does not have to make an immediate payment.
However, it may subsequently receive compensation payments from the overtaking jobs. A relatively simple
remedy for preventing jobs from exploiting the mechanism to earn money may be to add a variable cost
+
component to the payment scheme. This can be done by introducing z ∈R
i
which equals the operating costs
8
187

per timestep for machine i. When extending the scenario to feature strategically acting machines which can
misreport about
Ĉ j
and
πˆ
j
, machines may be allowed to strategically report their operating costs z i
~ . To
incorporate these operating costs in the mechanism, the tentative payment
ˆ π ( i | s
j
− j
~ ⎛
, t ) =
~
p ⎜ ~ z +
∑ w
~
j j i
k∈L(
j ), k→i,
k<
j,
⎝
k
S k > r
~ j
⎞
⎟
⎠
πˆ
j
has to be extended to
Allowing preemption of jobs may further increase the efficiency of the mechanism. In DLGM in its current
form, a job which is being processed at a machine may not be interrupted until it is completed. Grid applications
demand for the possibility to freeze the execution of a running job if a job with a higher priority enters the queue.
Deadline constraints are a further possible extension to the mechanism. With interactive settings,
applications / users need the jobs to be computed in a timely manner and the validity of the job is limited. An
extension to the basic mechanism is thus to expand a job’s (agent’s) type to t = r , p , w , d ) where
+
d ∈R
j
, d r + p
j j j
j
(
j j j j
~
~
≥ , is j’s deadline and agent j reports t = (
~
r ,
~
p , w
~
, d ) to the machines. This additional
j
j
j
j
j
information can then be used to adjust the priority of jobs when ordering the waiting queue such that jobs which
approach their deadline receive a higher priority. Assume a new job j reports to the system. Then, at time
~
rj
, the
priority values of all jobs have to be recomputed. A measure for the urgency of a particular job k is its distance to
the point in time at which it needs to be started at the latest in order to still be finished before its deadline, i.e.
~
d −
~
p −
~
r . The smaller this distance, the more the priority value of job k needs to be boosted. One possibility
k
k
k
is to use
~
p
k
w
~
k
~
( d −
~
p −
~
r )
k k k
as priority value for job k at time rj
~ w
~
instead of
k
~
pk
.
Overall, with such extensions, the DLGM mechanism is a promising approach towards scheduling in Grids.
6. Conclusion and Outlook
In Section 2, we briefly presented market mechanisms which have been proposed for the use in Grid
scheduling. In Section 3, we focused on the DLGM mechanism which exhibits several desirable features. In their
analysis, Heydenreich et al. [3] propose the use of the solution concept of myopic best response equilibria in
order to model the behavior of strategic agents. In order to examine the validity of this concept and the
performance of the mechanism in more realistic settings with learning agents, we performed an agent-based
simulation in Section 4. The central result of our analysis is that myopic best response strategies do not
adequately model more realistic agents. Furthermore, myopic best response strategies distort the performance
9
188

analysis of DLGM and yield overly optimistic results. In Section 5, we then pointed to several limitations of the
basic DLGM mechanism if considered for practical settings, and pointed at extensions to this basic mechanism
which may help to mitigate these drawbacks.
The presented results point to several interesting possibilities for future research. From the mechanism design
perspective, it would be interesting to evaluate the suggested extensions in more detail. We plan to implement
the mechanism in the Sun N1 Grid Engine, a state-of-the-art Grid scheduler by Sun Microsystems. This may
help in identifying both further limitations of this approach and possible remedies and extensions. The agent
perspective suggests two ways to extend on our research. Firstly, one fundamental question is how actual users
can discover their true valuations and interact with the mechanism. First research in this direction has been
presented in [19] and [20], who suggest some form of “intelligent” software agents as intermediaries which assist
the user in determining job priorities and communicating with the market. Secondly, the learning capabilities and
strategies of the (software) agents need further refinement and validation.
Bibliography
1. Foster, I., C. Kesselman, and S. Tuecke, The Anatomy of the Grid: Enabling Scalable Virtual Organizations.
International Journal of High Performance Computing Applications, 2001. 15(3): p. 200-222.
2. Sutherland, I.E., A futures market in computer time. Communications of the ACM, 1968. 11(6): p. 449-451.
3. Heydenreich, B., R. Müller, and M. Uetz., Decentralization and mechanism design for online machine scheduling., in
SWAT, L.A.a.R. Freivalds, Editor. 2006, Springer. p. 136-147.
4. Neumann, D., J. Stoesser, and C. Weinhardt, Bridging the Adoption Gap - Developing a Roadmap for Trading Grids.
International Journal of Electronic Markets, 2008. 18(1): p. forthcoming.
5. Grosu, D. and A. Das, Auctioning resources in Grids: model and protocols. Concurrency and Computation: Practice
and Experience, 2006. 18(15): p. 1909-1927.
6. Waldspurger, C.A., et al., Spawn: A distributed computational economy. IEEE Transactions on Software Engineering
Management Journal, 1992. 18(2): p. 103–117.
7. Schnizler, B., et al., Trading Grid Services - A Multi-attribute Combinatorial Approach. European Journal of Operation
Research, 2006: p. forthcoming.
8. Schnizler, B., D. Neumann, and C. Weinhardt. Resource Allocation in Computational Grids - A Market Engineering
Approach. in WeB 2004. 2004. Washington.
9. Bapna, R., et al., A market design for grid computing. INFORMS Journal of Computing, 2006: p. fortcoming.
10. Stoesser, J. and D. Neumann. GREEDEX – A Scalable Clearing Mechanism for Utility Computing. in Networking and
Electronic Commerce Research Conference (NAEC) 2007. Riva Del Garda, Italy
11. Stoesser, J., D. Neumann, and A. Anandasivam. A Truthful Heuristic for Efficient Scheduling in Network-Centric Grid
OS. in European Conference on Information Systems (ECIS 2007). 2007. St. Gallen.
12. Parkes, D.C., S. Singh, and D. Yanovsky. Approximately Efficient Online Mechanism Design. in 18th Annual
Conference on Neural Information Processing Systems (NIPS’04). 2004.
13. Parkes, D.C. and S. Singh. An MDP-based approach to Online Mechanism Design. in 17th Annual Conference on
Neural Information Processing Systems (NIPS’03). 2003.
14. Lai, K., et al., Tycoon: an Implementation of a Distributed, Market-based Resource Allocation System. Working Paper,
2004.
15. Hajek, S.S.a.B. Optimal allocation of a divisible good to strategic buyers. in IEEE Conference on Decision and
10
189

Control. 2004.
16. Kaelbling, L.P., M.L. Littman, and A.W. Moore, Reinforcement Learning: A Survey. Journal of Artificial Intelligence
Research, 1996. 4: p. 237-285.
17. Fu, Y., et al., SHARP: an architecture for secure resource peering. ACM SIGOPS Operating Systems Review, 2003.
37(5): p. 133-148.
18. Buyya, R., D. Abramson, and S. Venugopal, The grid economy. Proceedings of the IEEE, 2005. 93(3): p. 698-714.
19. MacKie-Mason, J.K. and M.P. Wellman, Automated Markets and Trading Agents, in Handbook of Computational
Economics, vol. 2: Agent-Based Computational Economics, L. Tesfatsion and K.L. Judd, Editors. 2006, North-Holland.
20. Neumann, D., S. Lamparter, and B. Schnizler, Automated Bidding for Trading Grid Services, in Proceedings of the
14th European Conference on Information Systems (ECIS). 2006: Gothenburg, Sweden.
11
190

Economically Enhanced Resource Management
for Internet Service Utilities
Tim Püschel 1,2 , Nikolay Borissov 2 ,MarioMacías 1 ,DirkNeumann 2 ,
Jordi Guitart 1 , and Jordi Torres 1
1 Barcelona Supercomputing Center - Technical University of Catalonia(UPC),
c/ Jordi Girona 29, 08034 Barcelona, Spain
{tim.pueschel,mario.macias,jordi.guitart,jordi.torres}@bsc.es
2 Institute of Information Systems and Management (IISM)
Universität Karlsruhe (TH), Englerstr. 14, 76131 Karlsruhe, Germany
{pueschel,borissov,neumann}@iism.uni-karlsruhe.de
Abstract. As competition on global markets increases the vision of utility
computing gains more and more interest. To attract more providers it
is crucial to improve the performance in commercialization of resources.
This makes it necessary to not only base components on technical aspects,
but also to include economical aspects in their design. This work
presents an framework for an Economically Enhanced Resource Manager
(EERM) which features enhancements to technical resource management
like dynamic pricing and client classification. The introduced approach
is evaluated considering various economic design criteria and example
scenarios. Our preliminary results, e.g. an increase in achieved revenue
from 77% to 92% of the theoretic maximum in our first scenario, show
that our approach is very promising.
1 Introduction
Many web applications have strongly varying demand for computing resources.
To fulfil these requirements sufficient resources have to be made available. In
some cases it is possible to run certain tasks at night to achieve a more even
usage, however in many cases it is not feasible to let users wait a long time
for results. This leads to a situation in which utilization is very high during
certain peak times while many resources lay idle during other times. Due to
high competition on global markets many enterprises face the challenge to make
use of new applications and reduce process times on one side and cut the costs
of their IT-infrastructures on the other side [5].
In light of this challenge the idea of utility computing gained interest. Utility
computing describes a scenario where computer resources can be accessed dynamically
in analogy to electricity and water [19]. The more resource providers
offer their resources or services, the more likely it is they can be accessed at
competitive prices. Therefore it is important to attract more providers.
However providers will only offer their services if they can realize sufficient
benefit. With state-of-the-art technology, this assimilation is hampered, as the
B. Benatallah et al. (Eds.): WISE 2007, LNCS 4831, pp. 335–348, 2007.
c○ Springer-Verlag Berlin Heidelberg 2007
191

336 T. Püschel et al.
local resource managers facilitating the deployment of the resources are not
designed to incorporate economic issues (e.g. price).
In recent times, several research projects have started to develop price-based
resource management components supporting the idea of utility computing.
Those approaches are devoted to scheduling by utilizing the price mechanism.
Clearly, this means that technical issues such as resource utilization are ignored
for scheduling. In addition, resource management is much more comprehensive
than just scheduling. For example Service-Level-Agreement (SLA) management
is also part of resource management that is often omitted in economic approaches.
This plays a role when deciding which already ongoing jobs to cancel
in overload situations to maintain system stability.
To improve performance in the commercialization of distributed computational
resources decisions about the supplied resources and their management
should be based on both technical and economic aspects [12]. Hence, this paper
is an interdisciplinary work taking into account aspects from computer science
and economics. We will explore the use of economic enhancements such as client
classification and dynamic pricing to resource management.
Technical resource management systems typically offer the possibility to include
priorities for user groups. In purely price-based schedulers it is not possible
to distinguish important from unimportant partners, as only current price matters
for the allocation. We will motivate that client classification should be integrated
into economically enhanced resource management systems. Essentially,
there are two main reasons to do so: First, client classification allows the inclusion
of long-term oriented relationships with strategically important customers
so-called credential components. Second, client classification can be used as an
instrument of revenue management, which allows skimming off consumer surplus.
The main contribution of this paper is to show how technical parameters can be
combined into an economically enhanced resources management that increases
revenue for the local resource sites.
2 Motivational Scenarios
2.1 Organisation Selling Spare Resource Capacity
An organisation offers different web services for internal use. However the utilization
of the resources is uneven. During certain times there is only a low load,
the resources are almost idle, and at other times users have to wait for their
results very long. Therefore the organisation decides to buy more capacity on
the market when needed and sell its spare capacity when the load is low. When
accepting jobs users from within the organisation should always be preferred.
To provide and communicate a good dynamic evaluation of resources, market
mechanisms are also used for internal users. This also gives internal users an
incentive to run their jobs during times of low utilization. Internal users also
receive a significant discount compared to external users.
192

Economically Enhanced Resource Management for Internet Service Utilities 337
2.2 Resource/Service Provider with Preferred Customers
A big service provider maintains a data center whose resources are sold. The
service provider already has a number of clients but still has a high spare capacity.
Therefore he joins a marketplace to find new clients and optimize capacity
utilization. To maintain the good relations with the current clients and encourage
regular use of its services the provider offers a preferred client contract.
Preferred clients receive a discount on the reservation price and soft preference
when accepting jobs. Soft preference means that their bids are increased by a
certain amount for the winner determination. When offering the same price the
preferred client is chosen over the external client, but the standard client has the
chance to outbid the preferred client. Prices should be calculated dynamically
based on utilization, client classification, estimated demand and further pricing
policies of the provider.
3 Economically Enhanced Resource Management
To improve performance in the commercialization of distributed computational
resources and increase the benefit of service providers we propose the introduction
of various economic enhancements into resource management. This chapter
first describes the objectives and requirements for the enhancements, then follows
a description of the key mechanisms that are to be integrated in the EERM.
The chapter ends with a description of the architecture and the components of
the EERM.
3.1 Objectives and Requirements
The main goals of these enhancements are to link technical and economical
aspects of resource management and strengthen the economic feasibility. This
can be achieved by establishing more precise price calculations for resources,
taking usage of the resources, performance estimations and business policies into
account. The introduced mechanisms should be be able to deal efficiently with
the motivational scenarios given earlier. This means they have to feature client
classification, different types of priorities for jobs from certain clients, reservation
of a certain amount of resources for important clients, and dynamic calculation
of prices based on various factors. In addition to these requirements the system
should also offer quality of service (QoS) and be able to deal with situations in
which parts of the resources fail. To adapt to different scenarios and business
policies of different situations it should be highly flexible and configurable via
policies.
When designing mechanisms various economic design criteria [4], [20] should
be considered. These following criteria apply to the respective features as well
as the overall system and the market mechanisms it is embedded in.
Individual Rationality. An important requirement for a system is that it is
individual rational on both sides, i.e. both providers and clients have to have a
benefit from using the system.
193

338 T. Püschel et al.
Simplicity and Computational Costs. While the enhancements introduce some
additional factors and they should not introduce any unnecessary complexity.
Similarly client classification, quality of service and dynamic pricing add
some additional computational complexity, however they should not add any
intractable problems and its benefits should outweigh its costs.
Revenue Maximization. A key characteristic for resource providers is revenue
maximization or more general utility maximization.
Incentive Compatibility. Strategic behaviour of clients and providers can be
prevented if a mechanism is incentive compatible. Incentive compatibility means
that no other strategy results in a higher utility than reporting the true valuation.
Efficiency. There are different types of efficiency. The first one considered here
is pareto optimimality. An allocation is considered pareto optimal if no participant
can improve its utility without reducing the utility of another participant.
The second efficiency criterion is allocative efficiency. A mechanism is called
allocative efficient if it maximizes the sum of individual utilities.
3.2 Key Features
The motivational scenarios and the further requirements lead to four key features
the of the EERM presented in this work.
Quality of Service. The first feature is quality of service. This can be broken
down into two aspects. The first aspect is to assure adequate performance during
normal operation of the resources. Overload situations can lead to reduced overall
performance [17] and thereby can result in breaking QoS agreements between
the provider and clients. Thus it is necessary to have a mechanism that ensures
that jobs will not be accepted if they result in an overload situation. The second
aspect of quality of service regards situations in which parts of the resources
fail. To be able to fulfill all SLAs even in situations of partial resource failure it
would be necessary to keep an adequate buffer of free resources. Where this is
not feasible there should at least be a mechanism that ensures that those SLAs
that can be kept with the available resources are fulfilled. This can be done by
suspending or cancelling those jobs that can not be finished in time due to the
reduced availability of resources.
Job Cancellation. Related to QoS is the feature automatic job suspension and
cancellation. It is needed to ensure quality of service in situations where problems
arise, i.e. parts of the resources fail or the estimations of the utilization were to
optimistic. Cancellation of lesser important jobs to free capacity for incoming
jobs with higher importance, i.e. jobs from a client with a higher classification
or a jobs that deliver significantly more revenue is also possible.
Dynamic Pricing. Another enhancement is dynamic pricing based on various
factors. [21] shows an approach for a pricing function depending on a base pricing
rate and a utilization pricing rate. However the price can depend not only on
current utilization but also on other factors such as projected utilization, client
194

Economically Enhanced Resource Management for Internet Service Utilities 339
classification, and projected demand. Pricing should also be contingent on the
demand on the market. This feature can be either implemented in the EERM
or in a dedicated component responsible for trading. This agent would request
a price based on factors including utilization of the resources and client classification.
from the EERM and then calculate a new price taking the situation on
the market into account.
Client Classification. Earlier giving clients different privileges was mentioned,
e.g. by discriminating on factors like price and quality of service. This part
describes the factors that can be used to differentiate various client classes.
Price Discrimination. Price discrimination or customer-dependent pricing is one
way to differentiate between different classes of clients. One idea to achieve this
is introducing Grid miles [16] in analogy to frequent flyer miles. Clients could be
offered a certain amount of free usage of the resources or a 10% discount after
spending a certain amount of money.
Reservation of Resources. For certain users it may be very important to always
have access to the resources. This class of users could be offered a reservation of
a certain amount of resources. One option is to reserve a fixed share of resources
for a certain class of users another possibility is to vary this share depending on
the usage of the system.
Priority on Job Acceptance. Another option is to use a client priority on job
acceptance. When the utilization of the system is low jobs from all classes of
clients are accepted but when the utilization of the resources rises and there is
competition between the clients for the resources, jobs from certain clients are
preferred. There are two types of priorities: strict priorities and soft priorities.
– Strict priority means that if a job from a standard client and a client with
priority compete for acceptance, the job from the client with priority always
wins. Jobs from clients with priority are always preferred, thus there is no
real competition between the different classes of clients.
– Soft priority means jobs from clients with priority are generally preferred but
standard clients have the chance to outbid clients with priority. Thus soft
priority is essentially a discount on the reservation price or bid that may only
apply in certain situation, i.e. when utilization exceeds a certain threshold.
Quality of Service. Another factor where differentiation for classes of clients
is possible is quality of service. For some classes of clients quality of service is
offered, for others not. Offering different levels of quality of service for different
classes of clients is also possible. An example for this would be offering different
risk levels [7].
4 Architecture of the EERM
This part describes the architecture of the EERM that was designed based on
the requirements and key features. It includes a description of the components
195

340 T. Püschel et al.
Fig. 1. EERM Architecture
with the aid of sequence diagrams. An overview of the architecture of the EERM
can be seen in Fig. 1.
The EERM interacts with various other components, namely a Grid Market
Middleware, a Monitoring component and the Resource Fabrics. The Grid
Market Middleware represents the middleware responsible for querying prices
and offering the services on the Grid market. The Monitoring is responsible for
monitoring the state and the performance of the resources and notifying the System
Performance Guard in case of problems. Additionally data collected by the
Monitoring is used by the Estimator component to for its predictions. Resource
Fabrics refers to Grid Middlewares such as Condor [15] or Globus [9].
Economy Agent. The Economy Agent is responsible for deciding whether
incoming jobs are accepted and for calculating their prices. These calculations
can be used both for negotiation as well as bids or reservation prices in auctions.
Figure 2 shows the sequence of a price request. First the Economy Agent receives
a request from a market agent. Then it checks whether the job is technically and
Fig. 2. Sequence Diagram Price Request
196

Economically Enhanced Resource Management for Internet Service Utilities 341
economically feasible and calculates a price for the job based on client category,
resource status, economic policies and predictions of future job executions from
the estimator component.
Estimator. The Estimator component calculates the expected impact on the
utilization of the resources. This is important to prevent reduced performance
due to overload [17], furthermore the performance impact can be used for the
calculation of prices. This component is based on work by Kounev et al. using
online performance models [13],[14].
System Performance Guard. The System Performance Guard is responsible
for ensuring that the accepted SLAs can be kept. In case of performance problems
with the resources it is notified by Monitoring. After checking the corresponding
policies it determines if there is a danger that SLAs cannot be fulfilled. It
then takes the decision to suspend or cancel jobs to ensure the fulfilment of the
other SLAs and maximize overall revenue. Figure 3 shows a typical sequence for
such a scenario. Jobs can also be cancelled when capacity is required to fulfil
commitments to preferred clients.
Policy Manager. To keep the EERM adaptable the Policy manager stores and
manages policies concerning client classification, job cancellation or suspension,
etc. Policies are formulated in the Semantic Web Rule Language (SWRL) [11].
All features of the EERM require the respective components to be able to communicate
with the Policy Manager and base their decisions on the corresponding
policies. A simple example from a pricing policy in SWRL is the following rule
which expresses that if the utilization is between 71% and 100% there is a surcharge
of 50:
Utilization(?utilization)∧InsideUtilizationRange(?utilization, ”71%−100%”)
⇒ SetSurcharge(?utilizationsurcharge, ”50”)
Fig. 3. Sequence Diagram System Performance Guard
197

342 T. Püschel et al.
Economic Resource Management. The Economic Resource Management is
responsible for the communication with the local resource managers and influences
the local resource management to achieve a more efficient global resource use.
5 Evaluation
The evaluation of this proposal consists of three parts. First an example scenario
is described, then the results are presented and last the economic design criteria
[4],[20] are considered.
5.1 Example Scenario
The first part of the evaluation considers an example scenario. For this evaluation
there were no actual jobs executed. The policies were manually applied and
revenue, utilization and utilities calculated. The job information given in Table 1
and the following assumptions were used.
The total capacity is 100 and the jobs given in Table 1 will be available during
the run. The system receives information about jobs one timeslot before they
become available. The Gold-client only uses the provider when he is assured
access to a certain capacity. In this case he is guaranteed a total capacity of 60
units, while being able to launch new jobs of up to 30 units per period.
Using this scenario as a basis four policies are evaluated. The following policies
represent the some of the basic options - no enhancements, client classification
with fixed reservation, dynamic pricing with reservation prices, and client classification
with strict priority:
– Case I is an example without any EERM. In this case any job is accepted if
there is enough capacity left to fulfil it.
– Case II features a simple form of client classification, there is a fixed reservation
of 60% of the capacity for the Gold-client. There is no System Performance
Guard.
– Case III features the EERM with utilization-based pricing and the System
Performance Guard, but without client classification. The policy is to only
accept jobs that offer a price higher than 1 currency unit per capacity unit
and time slot if the job results in utilization over 80%. The unit price is
determined by dividing the price by the total capacity used by a job over all
periods.
– In case IV the EERM with client classification and the system performance
guard is used. The policy is to accept only jobs from the Gold-client if the
job would result in utilization higher than 70%. This policy which can be
applied to the first motivational scenario when replacing ”Gold-client” with
”Internal”.
To evaluate the ability of the EERM to adapt to problems with the resources
there is also a second scenario. It includes a reduced capacity of 70 in t=7 and a
capacity of 60 in t=8 due to unpredicted failure of resources. In the cases without
198

Economically Enhanced Resource Management for Internet Service Utilities 343
Table 1. Example scenario
Job Start Time End Time Capacity/t Client Class Price Penalty
A 1 3 55 Standard 330 -82.5
B 1 5 24 Standard 180 -30
C 1 7 20 Standard 140 -17.5
D 2 4 20 Standard 120 -30
E 3 5 15 Standard 90 -22.5
F 3 8 20 Standard 120 -15
G 4 7 20 Standard 160 -40
H 4 9 15 Standard 135 -22.5
I 5 10 30 Standard 180 -22.5
K 5 8 30 Standard 240 -60
L 6 8 30 Standard 90 -11.25
M 6 9 12.5 Standard 50 -6.25
O 8 10 20 Standard 90 -15
P 9 10 21 Standard 84 -21
Q 9 10 30 Standard 90 -15
R 2 6 30 Gold 375 -187.5
S 5 8 30 Gold 300 -150
T 7 10 7.5 Gold 75 -37.5
U 7 9 20 Gold 150 -75
V 9 10 20 Gold 100 -50
the System Performance Guard it is assumed that the SLAs of jobs that end in
periods that are overloaded due to the reduced capacity fail and that jobs that
continue to run for more time can catch up and fulfil their SLAs. To assess the
benefits of the EERM for providers and clients we compare their utility without
and with EERM. Utility functions depend on the preferences of providers and
clients. For the provider revenue is obviously a key criteria. Another factor that
has to be considered is utilization. To allow for maintenance and reduce the
effects of partial resource failure it is not in the interest of the provider to have a
utilization that approaches 100%. In this example an average utilization of 80% is
considered optimal. The following utility function for the provider considers the
revenue per utilization while taking into account the optimal average utilization
of the resources:
u provider =
revenue
n ∗ averageutilization ∗ 1
1+|0.8 − averageutilization
100
|
For the client utility the idea is to weight the capacity needed by a job with
its importance. The importance of a job for a client is expressed in the price per
capacity unit the client is willing to pay for it. This results in the following client
utility function, where P is the set of prices per unit (i.e. 2.5, 2, 1.5, 1):
u client = ∑ l∈P
unitprice l ∗ allocatedjobcapacity l
totaljobcapacity l
199

344 T. Püschel et al.
Table 2. Result of the four cases
Case Completed
Revenue Avg Load u provider u client Revenue Proportion
I A, B, C, G, H, L, M, O , P, Q 1349 89.7 1.37 3.10 76.65%
II C, D, M, O, R, S, T, U, V 1400 71 1.81 3.33 79.55%
III A, B, D, G, H, K, M, O, P, Q 1479 84.7 1.67 3.25 84.03%
IV A, G, H, R, S, T, U, V 1625 73.5 2.08 3.87 92.33%
A* A, E, G, O, Q, R, S, T, U, V 1760 81 2.15 3.53 100%
Table 3. Result with partial resource failure
Case Completed
Failed Revenue Avg Load u provider u client Revenue Proportion
I A, B, H, L, M, P, Q C, G 901.5 79.7 1.13 2.28 56.56%
II D, P, R, M, T, U, V C, S 786.5 66.2 1.04 2.23 49.34%
III A, B, D, G, H, K, P, Q M 1332.75 74.95 1.69 3.05 83.61%
IV A, P, R, S, T, U, V G, H 1351.5 71.2 1.74 3.31 84.79%
A* A, E, P, Q, R, S, T, U, V 1594 71.2 2.06 3.11 100%
5.2 Results
Table 2 shows the results of applying these 4 cases to the scenario given in Table
1. A* is the allocation with maximum revenue that could be achieved if all job
information was available before the first period. The table includes information
about the completed jobs, the realized revenue, the average utilization of the
providers resources, the provider utility, the client utility, and the proportion of
the revenue in comparison with the A*. Cases II, III and IV perform significantly
better than the standard case I without any economical enhancements. Although
the average utilization is lower than in case I the yielded revenue as well the
provider’s and the client’s utility is higher.
Table 3 shows the results of applying the policies to the scenario with failure
of parts of the resources in t=7 and t=8. It can be seen that the improvement in
cases III and IV is even more significant. In this situation case II is worse than case
I regarding revenue, provider utility and client utility. This is caused by the lack of
the System Performance Guard and thereby higher cancellation penalties of the
jobs of the Gold-client. Case III delivers a significantly better utility for providers
and clients. However since it doesn’t include client classification it does not address
the needs of the Gold-client. Case IV again delivers the best results regarding revenue,
provider utility and client utility. Figure 4 shows the revenue proportion in
the standard case (a) and with partial resource failure (b). In both cases III and
IV the benefit of the System Performance Guard can be clearly seen, they deliver
of 83.61% and 84.79% compared to the 56.56% of the standard case I.
5.3 Economic Design Criteria
In cases III and IV of the example both sides have a clear benefit from using
the respective mechanisms, hence individual rationality is achieved in these
200

Economically Enhanced Resource Management for Internet Service Utilities 345
(a)
(b)
Fig. 4. Proportion of maximum revenue in the four cases
cases. The proposed mechanisms introduce some additional complexity but this
is encapsulated in the EERM. They, however, do not introduce any NP-hard
problems into the mechanism and the additional computational cost is limited.
Another key characteristic for resource providers is revenue maximization. Inthe
given example the policy of case IV delivers the highest revenue, provider utility
and client utility among the options.
Whether the criterion of incentive compatibility is fulfilled depends on the
market and classification mechanisms the EERM is embedded in. For the example
a simple pay as you bid scheme is assumed. As this means that clients set
their bids lower than their evaluation to achieve a benefit incentive compatibility
is not given in the example. If, however, embedded in an allocation and pricing
mechanism that is incentive compatible the EERM serves to give more precise
valuations for the jobs.
Efficiency also depends on the mechanisms the EERM is embedded in. For
the example we used a scenario where the provider decides whether jobs are
accepted. Obviously the provider is mainly concerned with maximizing its own
utility. Therefore it does not guarantee the maximum total utility. The EERM
can also be embedded in efficient market mechanisms serving to give a more
precise valuations of the jobs.
6 Related work
There is related work covering different aspects of our work. [8] discusses the
application of economic theories to resource management. [1] presents an architecture
for autonomic self-optimization based on business objectives. Elements
of client classification such as price discrimination based on customer characteristics
have been mentioned in other papers [16], [3]. They did however not
consider other discrimination factors. [6] describes data-mining algorithms and
tools for client classification in the electricity grids but concentrate on methods
for finding groups of customers with similar behaviour. An architecture for admission
control on e-commerce websites that prioritizes user sessions based on
predictions about the user’s intentions to buy a product is proposed in [18]. [2]
201

346 T. Püschel et al.
presents research on how workload class importance should be considered for
low-level resource allocation.
One approach to realize end-to-end quality of service is the Globus Architecture
for Reservation and Allocation (GARA) [10]. This approach uses advance
reservations to achieve QoS. Another way to achieve autonomic QoS aware resource
management is based on online performance models [13], [14]. They introduce
a framework for designing resource managers that are able to predict the
impact of a job in the performance and adapt the resource allocation in such a
way that SLAs can be fulfilled. Both approaches do not consider achieving QoS
in case of partial resource failure. Our work makes use of the second approach
and adds the mechanism of Job Cancellation.
The introduction of risk management to the Grid [7] permits a more dynamic
approach to the usage of SLAs. It allows modelling the risk that the SLA cannot
be fulfilled within the service level agreement. A provider can then offer SLAs
with different risk profiles. However such risk modelling can be very complex. It
requires information about the causes of the failure and its respective probabilities.
Clients need to have the possibility to validate the accuracy and correctness
of the providers risk assessment and risks have to be modelled in the SLAs.
7 Conclusion and Future Work
In this work various economical enhancements for resource management were
motivated and explained. We presented a mechanism for assuring Quality of
Service and dealing with partial resource failure without introducing the complexity
of risk modelling. Flexible Dynamic Pricing and Client Classification was
introduced and it was shown how these mechanisms can benefit service providers.
Various factors and technical parameters for these enhancements were presented
and explained.
Furthermore the preliminary architecture for an Economically Enhanced Resource
Manager integrating these enhancements was introduced. Due to the general
architecture and the use of policies and a policy manager this approach can
be adapted to a wide range of situations.
The approach was evaluated considering economic design criteria and using
an example scenario. The evaluation shows that the proposed economic enhancements
enable the provider to increase his benefit. In the standard scenario we
managed to achieve a 92% of the maximum theoretically attainable revenue
with the enhancements in contrast to 77% without enhancements. In the scenario
with partial resource failure the revenue was increased from 57% to 85%
of the theoretical maximum.
The next steps will include refinement of the architecture as well as the implementation
of the EERM. During this process further evaluation of the system
will be done, e.g. by testing the system and running simulations. Another issue
that requires further consideration is the generation of business policies for the
EERM. Further future work in this area includes the improvement of price calculations
with historical data and demand forecasting; the design of mechanisms
202

Economically Enhanced Resource Management for Internet Service Utilities 347
for using collected data to determine which offered services deliver the most revenue
and concentrate on them; the introduction of a component for automatic
evaluation and improvement of client classification.
Acknowledgments
This work is supported by the Ministry of Science and Technology of Spain
and the European Union (FEDER funds) under contract TIN2004-07739-C02-
01 and Commission of the European Communities under IST contract 034286
(SORMA).
References
1. Aiber, S., Gilat, D., Landau, A., Razinkov, N., Sela, A., Wasserkrug, S.: Autonomic
self-optimization according to business objectives. In: ICAC 2004. Proceedings of
the First International Conference on Autonomic Computing, pp. 206–213 (2004)
2. Boughton, H., Martin, P., Powley, W., Horman, R.: Workload class importance
policy in autonomic database management systems. In: POLICY 2006. Proceedings
of the Seventh IEEE International Workshop on Policies for Distributed Systems
and Networks, pp. 13–22 (2006)
3. Buyya, R.: Economic-based Distributed Resource Management and Scheduling for
Grid Computing. PhD thesis, Monash University (2002)
4. Campbell, D.E.: Resource Allocation Mechanisms. Cambridge University Press,
London (1987)
5. Carr, N.G.: The end of corporate computing. MIT Sloan Management Review
46(3), 32–42 (2005)
6. Chicco, G., Napoli, R., Piglione, F.: Comparisons among clustering techniques
for electricity customer classification. IEEE Transactions on Power Systems 21(2),
933–940 (2006)
7. Djemame, K., Gourlay, I., Padgett, J., Birkenheuer, G., Hovestadt, M., Kao, O.,
Voß, K.: Introducing risk management into the grid. In: eScience2006. The 2nd
IEEE International Conference on e-Science and Grid Computing, Amsterdam,
Netherlands, p. 28 (2006)
8. Ferguson, D.F., Nikolaou, C., Sairamesh, J., Yemini, Y.: Economic models for
allocating resources in computer systems, 156–183 (1996)
9. Foster, I., Kesselman, C.: Globus: A metacomputing infrastructure toolkit. International
Journal of Supercomputer Applications and High Performance Computing
11(2), 115–128 (1997)
10. Foster, I., Kesselman, C., Lee, C., Lindell, B., Nahrstedt, K., Roy, A.: A distributed
resource management architecture that supports advance reservations and
co-allocation. In: IWQoS 1999. Proceedings of the 7th International Workshop on
Quality of Service, London, UK, pp. 62–80 (1999)
11. Horrocks, I., Patel-Schneider, P.F., Boley, H., Tabet, S., Grosof, B., Dean, M.: Swrl:
A semantic web rule language combining owl and ruleml. Technical report, W3C
Member submission (2004)
12. Kenyon, C., Cheliotis, G.: Grid resource commercialization: economic engineering
and delivery scenarios. Grid resource management: state of the art and future
trends, 465–478 (2004)
203

348 T. Püschel et al.
13. Kounev, S., Nou, R., Torres, J.: Autonomic qos-aware resource management in
grid computing using online performance models. In: VALUETOOLS 2007. The
2nd International Conference on Performance Evaluation Methodologies and Tools,
Nantes, France (2007)
14. Kounev, S., Nou, R., Torres, J.: Building online performance models of grid middleware
with fine-grained load-balancing: A globus toolkit case study. In: EPEW
2007. The 4th European Engineering Performance Workshop, Berlin, Germany
(2007)
15. Litzkow, M.J., Livny, M., Mutka, M.W.: Condor - A hunter of idle workstations. In:
Proceedings of the 8th International Conference of Distributed Computing Systems
(1988)
16. Newhouse, S., MacLaren, J., Keahey, K.: Trading grid services within the uk e-
science grid. Grid resource management: state of the art and future trends, 479–490
(2004)
17. Nou, R., Julià, F., Torres, J.: Should the grid middleware look to self-managing
capabilities? In: ISADS 2007. The 8th International Symposium on Autonomous
Decentralized Systems, Sedona, Arizona, USA, pp. 113–122 (2007)
18. Poggi, N., Moreno, T., Berral, J.L., Gavaldà, R., Torres, J.: Web customer modeling
for automated session prioritization on high traffic sites. In: Proceedings of the 11th
International Conference on User Modeling, Corfu, Greece (2007)
19. Rappa, M.A.: The utility business model and the future of computing services.
IBM Systems Journal 43(1), 32–42 (2004)
20. Wurman, P.R.: Market structure and multidimensional auction design for computational
economies. PhD thesis, University of Michigan, Chair-Michael P. Wellman
(1999)
21. Yeo, C.S., Buyya, R.: Pricing for Utility-driven Resource Management and Allocation
in Clusters. In: ADCOM 2004. Proceedings of the 12th International Conference
on Advanced Computing and Communications, Ahmedabad, India, pp. 32–41
(2004)
204

SITUATED DECISION SUPPORT FOR MANAGING SERVICE LEVEL
AGREEMENT NEGOTIATIONS
Rustam Vahidov 1 , Dirk Neumann 2
1 John Molson School of Business
Concordia University, Montreal, Canada
rVahidov@jmsb.concordia.ca
2
Institute of Information Systems and
Management (IISM)
Universität Karlsruhe (TH)
neumann@iism.uni-karlsruhe.de
Abstract
In this paper we propose a situated decision
support system for efficient negotiating of service
level agreements (SLAs) in Grids. Situated decision
support systems effectively combine human
judgment with autonomous decision making and
action by agents. The intuition of using this approach
for SLA negotiations lies in the monitoring
and controlling of the fleet of local agents
negotiating single services from multiple service
providers by the use of a “manager” agent and
human decision maker. We show through numerical
experiments that the approach performs
well under a set of simplifying assumptions.
1. Introduction
Quite recently, Grid computing has been increasingly
gaining traction in a number of application
areas. Grid computing denotes a computing
model that distributes processing across an administratively
and locally dispersed infrastructure.
The ability to connect numerous heterogeneous
computing devices, allows the creation of
virtual computer architectures from otherwise
idle resources.
According to market surveys is the average productivity
of IT infrastructures extremely low:
mainframes are up to 40 % of the time idle;
UNIX servers even less than 10 % of the time
used; individual PCs reach a utilization rate of at
most 5 % [2]. This daunting ineffectiveness of IT
infrastructures and associated costs entail that
enterprises seek to use Grid technologies to outsource
IT services. Rather than investing and
maintaining proprietary infrastructures enterprises
tap to the Grid and consume IT services
on-demand. Service providers like SUN and
Amazon accommodate this need by offering
CPU hours and storage over web-services.
Outsourcing of IT services is, however, not a
panacea as proposed by critics Nicholas Carr [2].
Any outsourced IT service incurs high risks, as
the deployment is outside the control of the enterprises.
Service providers address this risk by
referring to formal contracts, so-called Service
Level Agreements (SLAs). SLAs define the
common understanding about services (e.g.,
priorities, responsibilities, guarantees and penalties
once SLAs are violated) between service
providers and consumers. Essentially, SLAs are
the crucial instrument for service consumers to
formulate guarantees on the service quality delivered
by the service provider. Nonetheless,
SLAs are also an important tool for service providers
to advertise free service capacities as well
as to manage their internal resources.
It is thus not astounding that SLA lifecycle management
is currently a hot topic [6, 10]. One
major success has been the specification of the
de-facto standard WS-Agreement as a description
language for SLAs by the Open Grid Forum
(OGF). WS-Agreement allows the formulation
of well defined and comprehensive contracts,
which minimizes the risks of disputes resulting
from unclear formulations among service provider
and consumer. Clearly, a description language
is merely prerequisite for well defined
SLAs. What is equally important is a procedure
with which SLAs are created. This becomes even
more important, if SLAs are dynamically
adapted dependent on resource availability and
demand. If the SLAs are well-negotiated, the
likelihood that balanced SLAs are signed is
much higher. In this spirit, the Grid Resource
Allocation Agreement Protocol Working Group
(GRAAP-WG)
(http://forge.gridforum.org/projects/graap-wg) of
OGF set on their agenda for future work to define
so-called negotiation profiles or protocols
for SLA negotiation as part of the WS-
Agreement specification [22].
SLA negotiations can become very complex as
service consumers need to negotiate with multiple
service providers to reserve different resources.
The resource reservations that are
205

needed are typically described via a workflow.
Service consumers either want to have all resources
reserved along the workflow or none at
all. This is reasoned by the fact that failure to
reserve one single constituent of the workflow
results in unsuccessful termination of the service
[15]. Due to this complexity the use of software
agents have been proposed for managing SLA
negotiations [4, 7, 19].
In this work our interest is in one-to many SLA
negotiations involving one service provider and
multiple potential customers. In such settings the
complexity involved in tracking and management
of multiple on-going negotiations can be
alleviated by means of intelligent support (e.g.
software agents). However, in our opinion, a
balanced approach combining autonomous action
with overall human control and decisionmaking
is a more adequate approach, as it allows
for timely intervention and improves the predictability
of business outcomes. In this respect the
recently introduced model of “situated decision
support” seems to be a promising framework to
apply to SLA negotiations [26, 28]. Situated
decision support advocates abandoning the traditional
“toolbox” type of decision aids in favor of
a more connected and active mode. The major
components of situated decision support system
include sensors, effectors, manager, and active
user interface. In this setup, decision support
expands to include problem sensing and action
generation and monitoring of the implementation
of decisions in addition to the conventional intelligence/design/choice
phases. It also allows
flexibility for effectively combining autonomous
action by the system with judgmental input from
the human decision makers.
Since the fields of decision and negotiation support
are closely related, in this work we show the
potential applicability of the situated decision
support model to managing multiple concurrent
SLA negotiations, where a managing entity
monitors and controls the local negotiation
agents. The contribution of this paper is threefold:
• Firstly, we provide motivation for the adoption
of situated decision support model for
SLA negotiations;
• Secondly, we briefly describe a situated decision
support framework for SLA negotiations.
• Thirdly, we illustrate the approach through
numerical simulation experiments for the
scenario involving two issues.
The remainder of this paper is structured as follows.
Section 2 provides the background on
SLAs, SLA negotiation and on agent strategies
in automated negotiations. Section 3 motivates
and describes the situated decision support
framework and shows how it can be used for
SLA negotiation. Section 4 provides an evaluation
on the basis of numerical experiments showing
that situated decision support is advantageous
for SLA negotiations. Section concludes
with a summary and an outlook on future research.
2. Background
This section provides brief overview of service
level agreements (definition, content of typical
SLAs), SLA negotiations (common protocols)
and agent-based strategies. It further discusses
why situated DSS model is deemed promising
for use in SLA negotiations.
Service Level Agreements
A SLA is defined a contract between a service
provider and consumer that specifies the rights
and obligations of the provider and the penalties
that will be applied if those obligations are not
satisfied.
Since the late 80s SLAs have been used by telecom
operators as contracts with their corporate
customers. With the recent trend of outsourcing
and application service providing, IT departments
have adopted the idea of using service
level agreements with their internal or external
customers.
SLAs for Grid computing are not very different
than those for other services (e.g. computation,
or storage services). But while the elementary
issues of an SLA can be the same, negotiating
SLAs for Grid services is aggravated by the fact
that Grid services are typically composed of
several basic services. This implies that it is
essential needs to manage concurrent SLA negotiations
with multiple service providers [24].
Typical attributes of SLAs for Grid are compiled
in Table 1.
Attributes
Operating System
Attribute Levels
Linux
Solaris
Windows
Service Availability 99,999 %
99,99 %
99,9 %
Price
< 1,000 €/month
< 5,000 €/month
< 10,000 €/month
> 10,000 €/month
206

CPU type
CPU utilization
Storage (RAM)
Bandwidth
Deployment
Co-allocation
Single CPU
Dual CPU
4 CPU
> 4CPU
1,000 CPU hours
5,000 CPU hours
10,000 CPU hours
1 GB
2 GB
4 GB
> 4 GB
1 Mb
10 Mb
100 Mb
1 Tb
Webservice
Source code
others
Service is provided by a
single provider
Service is provided by a
multiple provider
Table 1: Common Attributes and attribute
levels for Grid services
Service Level Agreements Negotiation
It is commonly accepted that the success of SLA
negotiations relates to the specification and implementation
of a protocol that creates legallybinding
agreements. This protocol needs to describe
a set of domain-independent messages,
together with an abstract schema, such that it can
be exchanged by all negotiators [15]. Several
bargaining protocols that meet those requirements
have been proposed by [4, 7, 9, 20, 25].
The bargaining protocols share a very similar
structure but differ in some details. For example
the protocol proposed by [25] involves multiple
rounds among service consumers (e.g., coallocators)
and providers (e.g., schedulers) until
they reach an agreement. The providers post
their offers on request to the consumers, who can
either select among the available offers or enter
re-negotiation by relaxing some constraints.
Another noteworthy effort has been proposed by
[17] who combine the negotiation protocol with
a template-based contract generation process. At
the end of the negotiation protocol a fullyfledged
SLA is generated as WS-Agreement
instance.
Agent-based Negotiation
SLA negotiations can involve a number of participants
both on the consumer as well as the
provider side trying to reach an agreement in
multiple bi-lateral interactions. A given provider,
for example, could have several on-going negotiations
at the same time and has to set the objectives,
constraints and strategies in those negotiation
instances in such a way that maximizes the
achievement of business objectives, while avoiding
over-commitment, i.e. promising more than
the provider can deliver. Clearly, with the growing
complexity of the composite services and
increasing customer base, automated means of
supporting multiple negotiations could help the
providers to better handle SLA negotiation processes.
In this respect, employing intelligent
agents seems to be an adequate approach. Below
we briefly describe some agent-based approaches
to automated negotiations, in particular for the
case of multi-bilateral settings.
Most of the related work in the area of agentbased
negotiation has been devoted to the design
of strategies for agents. The fundamental work
has been undertaken without considerations to
SLA negotiations. Nonetheless, as those approaches
are domain-independent, they are
highly relevant for SLA negotiations.
Faratin et al. have proposed a “smart” strategy
for autonomous negotiating agents [8]. Agents
following this strategy would try to make tradeoffs
in a manner that the newly generated offer is
similar to the opponent’s last offer, before trying
a concession. Another work in this direction
seeks to map business policies and contexts to
negotiation goals, strategies, plans, and decisionaction
rules [16].
While fully automated negotiations may not
always be a feasible choice, agents could also act
as intelligent assistants, helping the users by
providing advice, critiquing user’s own candidate
offers as well as the offers by an opponent,
and generating candidate offers for a user to
consider [3, 14].
In multi-bilateral negotiations a negotiator may
be having several concurrent negotiation processes
taking place at the same time and involving
multiple opponents A fuzzy set-theoretic approach
to analysis of alternatives in multibilateral
negotiations has been proposed in [29].
The authors have considered scenarios involving
RFQ sent by one seller to multiple buyers
(agents) with the purpose of deciding which
potential buyers to negotiate with. They used
fuzzy-relational approach to obtain a partial
rank-order of the prospective buyers.
A setup where multiple agents negotiate
autonomously and one agent is designated as a
coordinator has been proposed in [18] and [21].
207

In [18] a buyer agent runs multiple concurrent
negotiation threads that interact with several
sellers. The coordinator agent provides each
thread with reservation values and negotiation
strategy. The threads report back to the coordinator,
which then advises them on the possible
changes to reservation values or strategies. In
[21] a similar approach has been used including
coordinating agents and multiple “sub-buyer”
agents. In [5] a protocol for handling many-tomany
concurrent negotiations for internet services
has been proposed.
In most of the above models of one-to-many
negotiations it is assumed that the party on the
one side can only have one agreement as an
outcome of negotiations with multiple opponents
(i.e. XOR case). In this work we interested in
managing multiple SLA negotiations where each
negotiation may fail or succeed regardless of
others (OR case). In such settings the system
could learn from the agreements made recently
and take into account other relevant information
(e.g. market situation) to direct concurrent negotiations.
3. Situated DSS for SLA Negotiation
When negotiation mechanism is used as one of
the means to conduct daily basic business activities,
much human effort may be required. Employing
automated software components to conduct
or assist human negotiators in conducting
regular SLA negotiations may be a promising
solution to achieve significant cost and time
savings. However, with automated negotiations
there is a potential danger that the overall process
and the important business outcomes may
become somewhat unpredictable. Moreover,
often the course of negotiations depends on other
factors, lying outside the domain of the expertise
or sensory capabilities of the agents. For example,
customer behavior may be heavily affected
by the latest important economic, technologyrelated
and other types of events. This calls for a
type of solution allowing certain degree of automation,
but allowing the control by the user of
the overall process.
In light of the above considerations we propose a
type of solution that would effectively combine
human judgment capabilities with autonomous
actions by agents. The key motivation here is to
relieve human users from the necessity of being
involved in each and every SLA negotiation
process. Rather, the system should allow the
human decision maker to effectively and efficiently
manage the fleet of negotiating agents to
meet and maintain higher-level business targets.
The field of decision support systems (DSS) is
very much related to the area of negotiations and
negotiation support systems [13]. In the recent
DSS literature there has been important research
streams directed towards building “active” and
agent-based systems that would transform the
original “toolbox” model of DSS into an active
participant in the decision-making process,
which could perform some decision-related tasks
in an autonomous fashion [1, 11, 23, 27].
One such model introduced recently is known as
“situated decision support system”, or “decision
station” [26, 28]. Situated DSS looks to combine
the benefits of agent technologies and those of
decision support systems to facilitate active
problem sensing, decision implementation and
monitoring in addition to pure decision support.
In essence, situated DSS expands the traditional
model from purely “problem-solving/decision
making” frame to situation assessment and action
generation.
Situated DSS is made up from different active
(agent) components in addition to the traditional
“toolbox” of data, models, and knowledge. The
components include: sensors (for information
search and retrieval), effectors (for affecting
current state of affairs), manager (for deciding
how to handle a particular situation), and active
user interfaces (for intelligent interaction with
the user).
The composition and interactions of different
components of the situated DSS model provides
an appealing blueprint for facilitating one-tomany
SLA negotiations with limited autonomous
action and overall control of the process by a
human decision maker. Various negotiationrelated
tasks can be mapped to specific agents in
the model. For example, tasks including monitoring
service levels, sensing potential problems,
generating alerts, and predicting market trends
for SLA could be delegated to the “sensor”
agents. Controlling the fleet of negotiating
agents by monitoring their progress and instructing
them on the adjustments to negotiation
strategies, reservation levels and preference
structures can be delegated to some extent to the
“manager” agent. Human decision maker could
set the limits of authority of this agent and intervene
if necessary. Each negotiation instance
could be delegated to a particular “effector”
agent that conducts a given negotiation and reports
on the progress.
The model for supporting multiple SLA negotiations
is shown in figure 1.The situated DSS
208

model is essentially hierarchical involving three
layers. At the bottom layer, which could be
called “operational” the agents perform negotiations.
They are given the preferences, aspiration
and reservation layers and strategies to follow
and try to negotiate effectively with a given
opponent. The basic cycle of generating an offer
by these agents could be described as: retrieve
preference structure (possibly updated by the
“manager” agent); compare past offer and
counter-offer; generate new offer according to
adopted strategy. Automated negotiations have
been studied extensively in the recent past and
thus we do not focus on the detailed design of
negotiating agents. One possibility is to employ
“smart” strategy by the negotiating agents proposed
in [8].
The second layer contains the manager agent,
which makes use of the traditional DSS components
(data, models and knowledge) to manage
the negotiating agents. The manager monitors
key indicators of the negotiation processes, such
as number of agreements reached, proportion of
failed negotiations, resource consumption and
others and decides whether to intervene or not.
For example, if memory resource becomes
scarce, then in the next round of SLA negotiations
the manager would instruct the agents to
stress more the importance of memory resource
in the utility calculation. One way to encode the
knowledge of the manager could be through “If-
Then” (possibly fuzzy) rules. The following
simple example from a manager policy in SWRL
[12] is the following rule which expresses that if
memory units sold is somewhat larger than
memory units estimated, the importance of
memory units is slightly increased:
Function(?MemoryUnitsSold, ?MemoryUnitsEstimated,
?MemoryImportance,
?Increment) ←
builtin:greaterThanOrEquals(?MemoryU
nitsSold, ? MemoryUnitsEstimated),
,builtin:add(?MemoryImportance,?Increm
ent,? MemoryImportance)
Such rule, when invoked would change the preference
structure for the negotiator agents, which
then would be willing to give up less memory
units as a trade-off compared to other issues.
This level could be termed “planning” layer.
Figure 1: Situated DSS for managing multiple
SLA negotiations
The manager agent would base its decisions
solely on the information available to it, i.e.
reports received from the sensors and negotiators.
However, in reality, there are many other
sources and types of information that may not be
accessible to the agent. Furthermore, often
judgmental input would be required to set the
limits of authority for the manager agent, set the
objectives and constraints (e.g. to maintain the
ratio of failed vs. successful negotiations at a
certain level) and attune its parameters, e.g. its
risk profile, or the thresholds for sensitivity for
reacting to undesirable developments. This is
accomplished by the user through active user
interface. Active user interface facilitates effective
interaction with the user, while learning user
preferences. This layer could be called the
“judgmental” layer. Components of situated DSS
could send various alerts to the user when human
intervention may be desirable.
4. Numerical Experiments
To illustrate the situated DSS approach with the
example and obtain preliminary empirical results
we have conducted simulation experiments described
in this section.
4.1 Data Generation
We have used the scenario where service provider
negotiates with multiple prospective customers
the terms of the SLA. For the simplicity,
we considered only two issues: price of service
and number of CPU hours. Simplifying assumptions
were made, e.g. the negotiation ended the
same day as it began. We did not actually simulate
every negotiation process, as automated
209

negotiations have been studied in the past, and
the focus of the work is on the overall performance
of situated DSS that includes negotiating
agents as components. We generated various
values for reservation levels for the incoming
customers and their preference structures. We
furthermore assumed that if agreement between a
customer and an agent was possible it would be a
Nash solution. Thus, we assumed that the parties
(consumers and negotiating agents) followed
reasonable strategies to achieve mutually beneficial
outcomes. The latter assumption has been
adopted, since in [8] strategies were proposed for
autonomous negotiations that lead to desirable
outcomes for the parties involved.
In simulated settings we had included 20 days of
SLA negotiations with twenty potential service
consumers generated each day. The consumer
reservation levels for price were generated from
normal distribution around the “market” price,
which was set to $100 per 1000 CPU hours. The
requested number of CPU hours was also distributed
normally. The number of available CPU
hours was set to two millions and this “capacity”
did not change throughout the period. Each service
consumer was assigned a separate negotiator
and the reservation levels were provided by
the “manager” (agent). In addition, we have
incorporated the preferences for the number of
CPU hours vs. unit price by introducing discount
options, which were also controlled by the manager.
In particular, if agents committed more
(less) CPU hours than targeted for a given period,
the manager would decrease (increase) the
discount values for additional CPU hours. This
was accomplished through “If-Then” type of
rules, similar to that described in the previous
section.
On a single day a number of agreements were
made ranging from 0 to 20. Then the outcomes
were analyzed and corrections were made by the
manager to price limits and discounts using simple
rules, e.g.: “if the number of successful negotiations
is smaller than expected by a given margin
then adjust the reservation level for price
downward”. The manager could make such adjustments
only within the limits specified by the
human user. In the extreme case no such adjustments
could be allowed.
Figure 2: Effect of price level adjustments on
number of agreements reached
4.2 Results
Figure 2 shows the number of service level
agreements reached throughout the 20-day period
as it is affected by the manager’s adjustments.
Vertical bars represent increases and
decreases to price levels set by the manager
agent and communicated to the negotiator
agents. For example, in period 8 the manager
drops the reservation levels for negotiating
agents to increase the number of deals made. The
number of agreements reduces to zero by the end
of the period as the capacity limit for CPU hours
is reached. Figure 3 shows the effect of adjustments
to the discount levels on the number of
CPU hours.
Figure 3: Effect of adjustments to discount
levels on the number of CPU hours
As mentioned earlier, the freedom given to the
manager in setting the levels is controlled by the
human decision maker. In setting these levels the
human decision maker should use his or her
knowledge of the market situation and make a
judgment. The settings could also translate into
210

the various levels of risk the decision maker is
willing to take. For example if these are set high
and tight there is a risk that not all CPU hours
would be sold by the end of the period. If the
spread between the upper and lower levels is
large, then manager would start with higher
levels and then gradually decrease them to be
aligned with the market, thus possibly losing
some early opportunities.
The above figures were based on a scenario
when the market price remains stable. In case the
market price is initially much lower than the
price levels set by the manager and then gradually
increasing, the manager would have to make
several adjustments to its reservation level before
it could get the needed number of agreements.
This situation is shown in figure 4.
Figure 4: Adjustments when market price
increases from low to high
Also, in case of a sudden change in market price
the manager would be able to adjust, though it
would take some time and result in the loss of
some possibilities. In such cases making early
interventions as a result of rightly exercised
judgment by the human decision maker would be
beneficial. Such a combination of human judgment
with the automated capabilities could be
beneficial to the successful operation of the business.
Finally, figure 5 compares profits achieved by
the fixed pricing policy (price was fixed equal to
the market price) vs. negotiation based policies
based on the profits averaged over 20 runs.
Figure 5: Comparison of average profits
achieved by fixed pricing vs. situated DSSdriven
negotiation-based approaches
The negotiation-based scenarios divide into two:
the one where manager’s actions were restricted
(SDSS1), and the one where there are little restrictions
(SDSS2). The latter case promises
highest profits, though also least control that
might jeopardize higher-level policies. As one
can see the more adaptive situated DSS-based
solutions perform better than the fixed price
based one.
Table 2 shows the average profits made by using
the fixed pricing and situated DSS approaches
under various price settings. The column headings
reading “price” refer to constant price set in
case of the fixed pricing and to the initial reservation
price in cases of situated DSS.
Method Price = Price = Price =
90 100 110
Fixed 171.8 190.9 163.0
SDSS1 206.9 213.4 191.1
SDSS2 210.6 213.0 207.4
Table 2: Mean Profits earned by situated DSS
vs. fixed pricing (in $1000)
5. Conclusion
This work proposed applying situated decision
support approach to managing automated SLA
negotiations. The framework is based on the
model for situated decision support that effectively
combines human judgment and autonomous
decision making and action by agent components.
The key idea behind the approach lies in
the managing the fleet of negotiating agents by
the use of a “manager” agent and human decision
maker. We have illustrated the approach
through simulation experiments performed under
a set of simplifying assumptions.
Possible future work could be directed towards
extending the situated DSS to workflow cases,
211

where the service consumers engages into concurrent
negotiations with many service providers
such that SLAs will be made with respect to all
constituents of the workflow. In addition, future
work will comprise the implementation of a
user-friendly prototype involving multiple negotiated
issues and empirical testing involving
human subjects.
References
[1] Angehrn, A.A. and S. Dutta, Case-Based
Decision Support. Communications of the
ACM, 1998. 41(5): p. 77-86.
[2] Carr, N., The End of Corporate Computing.
MIT Sloan Management Review, 2005. 46(3):
p. 66-73.
[3] Chen, E., R. Vahidov, and G.E. Kersten,
Agent-Supported Negotiaitons in the e-
Marketplace. Int. J. Electronic Business, 2005.
3(1): p. 28-49.
[4] Czajkowski, K., et al. SNAP: A Protocol for
Negotiating Service Level Agreements and Coordinating
Resource Management in Distributed
Systems. in 8th Workshop on Job Scheduling
Strategies for Parallel Processing. 2002.
Edinburgh, Scotland.
[5] Dang, J. and M.N. Huhns, Concurrent Multiple-Issue
Negotiation for Internet-Based Services.
Internet Computing, 2006. 10(6): p. 42-
49.
[6] Deora, V. and O. Rana, SLA Enforcement in
Grid using Runtime Violation Prediction and
Reaction. Working Paper, 2007.
[7] Eymann, T., et al. On the Design of a Two-
Tiered Grid Market Structure. in Business Applications
of P2P and Grid Computing, MKWI
2006. 2006.
[8] Faratin, P., C. Sierra, and N.R. Jennings,
Using Similarity Criteria to Make Issue Trade-
Offs in Automated Negotiations. Artificial Intelligence,
2002. 142: p. 205-237.
[9] Gimpel, H., et al. PANDA: Specifying Policies
for Automated Negotiations of Service
Contracts. in 1st International Conference on
Service Oriented Computing (ICSOC 03).
2003. Trento, Italy.
[10] Hasselmeyer, P., et al. Towards Autonomous
Brokered SLA Negotiation. in eChallenges
2006. 2006. Barcelona.
[11] Hess, T.J., L.P. Rees, and T.R. Rakes, Using
Autonomous Software Agents to Create Next
Generation of Decision Support Systems. Decision
Sciences, 2000. 31(1): p. 1-31.
[12] Horrocks, I., et al., Swrl: A semantic web
rule language combining owl and ruleml.
Technical report, 2004.
[13] Jarke, M., M.T. Jelassi, and M.F. Shakun,
MEDIATOR: Towards a Negotiation Support
System. European Journal of Operational Research,
1987. 31(3): p. 314-334.
[14] Kersten, G.E. and G. Lo, Aspire: Integration
of Negotiation Support System and Software
Agents for E-Business Negotiation. International
Journal of Internet and Enterprise
Management (IJIEM), 2003. 1(3): p. (in print).
[15] Kuo, D., M. Parkin, and J. Brooke. Negotiating
Contracts on the Grid. in eChallenges.
2006. Barcelona, Spain.
[16] Li, H., S.Y.W. Su, and H. Lam, On Automated
e-Business Negotiations: Goal, Policy,
Strategy, and Plans of Decision and Action.
Journal of Organizational Computing and Electronic
Commerce, 2006. 13: p. 1-29.
[17] Ludwig, H., et al. Template-Based Automated
Service Provisioning - Supporting the
Agreement-Driven Service Life-Cycle. in International
Conference on Service Oriented Computing.
2005.
[18] Nguyen, T.D. and N.R. Jennings. Coordinating
multiple concurrent negotiations. in 3rd
Int. Conf. on Autonomous Agents and Multi-
Agent Systems. 2004. New York, NY.
[19] Ouelhadj, D., et al., A Multi-agent infrastructure
and a Service Level agreement Negotiation
Protocol for Robust Scheduling in Grid
Computing, in Advances in Grid Computing -
EGC 2005, T.P.A.R. Peter M.A. Sloot Alfons
G. Hoekstra, Marian Bubak, Editor. 2005,
Springer: Heidelberg. p. 651-660.
[20] Parkin, M., D. Kuo, and J. Brooke. A
Framework & Negotiation Protocol for Service
Contracts. in Proceedings of the IEEE International
Conference on Services Computing.
2006.
[21] Rahwan, I., R. Kowalczyk, and H.H. Pham,
Intelligent Agents for Automated One-to-Many
e-Commerce Negotiation. Australian Computer
Science Communications, 2002. 24(197-204).
[22] Rana, O., OGF-19 Report. Technical Report,
2006.
[23] Rao, H.R., R. Sridhar, and S. Narain, An
Active Intelligent Decision Support System.
Decision Support Systems, 1994. 12(1): p. 79-
91.
[24] Sahai, A., et al. Specifying and Monitoring
Guarantees in Commercial Grids through SLA.
in Conference of Cluster Computing and Grids
(CCGrid). 2003.
[25] Siddiqui, M., A. Villazon, and T. Fahringer.
Grid Capacity Planning with Negotiationbased
Advance Reservation for Optimized QoS.
212

in Proceedings of the ACM/IEEE Conference
on Supercomputing. 2006. Tampa, USA.
[26] Vahidov, R. Decision station: a Notion for a
Situated DSS. in 35th Hawaii International
Conference on System Sciences. 2002.
[27] Vahidov, R. and B. Fazlollahi, Pluralistic
Multi-Agent Decision Support System: A
Framework and an Empirical Test. Information
& Management, 2004. 41(7): p. 883-398.
[28] Vahidov, R. and G.E. Kersten, Decision
Station: Situating Decision Support Systems.
Decision Support Systems, 2004. 38(2): p. 283-
303.
[29] Van de Walle, B., S. Heitsch, and P. Faratin.
Coping with One-to-Many Multi-criteria Negotiations
in Electronic Markets. in 12th International
Workshop on Database and Expert Systems
Applications. 2001.
213

Using k-pricing for Penalty Calculation in Grid Market
Michael Becker, Institute of Information Systems and Management, University of Karlsruhe 1
Nikolay Borrisov, Institute of Information Systems and Management, University of Karlsruhe 2
Vikas Deora, School of Computer Science, Cardiff University 3
Omer F. Rana, School of Computer Science, Cardiff University 4
Dirk Neumann, Institute of Information Systems and Management, University of Karlsruhe 5
Abstract
To distribute risk in Grid, the design of service level
agreements (SLAs) plays an important role, since
these contracts determine the price for a service at
an agreed quality level as well as the penalties in
case of SLA violation. This paper proposes a price
function over the quality of service (QoS) on the
basis of the agreements negotiated upon price and
quality objective. This function defines fair prices for
every possible quality of a service, which are in line
with the business of the customer and incentivize the
provider to supply welfare-maximizing quality.
Therewith, penalties can be calculated for every
possible quality level as the difference between the
agreed price and the output of the price function for
the effectively met quality. A price function according
to the k-pricing scheme is presented for a single
service scenario and for a scenario with multiple
interdependent services.
1. Introduction
In order to drive sustainable business growth,
organizations are nowadays constantly improving
their information technology systems to produce
business results sooner and make them available to a
wider audience within the organization. To meet
increasing demands on IT services without bearing
enormous costs, IT services are sourced out more and
more. To be able to fulfill high requirements even on
peak demand, services are purchased on demand.
Unfortunately, all organizations purchasing and
selling computational services in a Service Oriented
Computing (SOC) environment such as Grid have to
_____________________________
1
michael.becker1981@web.de
2 borrisov@iism.uni-karlsruhe.de
3 v.deora@cs.cardiff.ac.uk
4 o.f.rana@cs.cardiff.ac.uk
5
neumann@iism.uni-karlsruhe.de
bear risks. Each outsourced service used by a
company could provide high risks since they cannot
be controlled adequate anymore. But also service
providers bear risks, since their reputation will shrink
and a penalty fee will have to be paid, if it fails to
fulfil the Service Level Agreement (SLA).
A SLA defines the responsibilities of the
service provider and the consumer of the relevant
service by describing the service provided and the
quality standards of the service, measurement
criteria, reporting criteria as well as penalties in cause
of SLA violation. A SLA can be seen as contract that
clearly outlines the rights and obligations of the
parties in order to reduce the scope for disagreement
to arise in the course of the parties` business
relationship. The source of the largest number of
disputes is likely to be the gap between the actual
performance by a service and the performance
expected by the service consumer. Thus, wellnegotiated
SLAs that are comprehensive and well
drafted are fundamental to minimizing the risk of
disputes. Thereby SLAs should be negotiated in a
way that both, the service provider and consumer will
profit by successful executed SLAs. “The best SLAs
are setup to allow both you and your service provider
to share in the success and failure of an agreement”
[3].
Besides the establishment of transparency, the
creation of SLAs aims to arrange risk sharing in
order to provide a fair market, which equates
financially weak and strong enterprises as well as
service providers and consumers. To establish a fair
agreement it is important that agreed service level
objective, price and penalties are accomplishable and
in line with the business of the contractual partners.
To incentivize providers to fulfill the agreement,
penalty design plays an important role. Thereby,
penalties have to be designed according to the
business value of the service for the customer and in
a way that incentivizes the provider to fulfill the
agreement satisfactory.
214

The main contribution of this paper is to define
balanced penalties that induce the right incentives for
service providers to meet the SLAs. This paper
introduces a price function on the basis of the results
of the negotiation (price and service quality). This
function provides fair prices for every possible
quality of a service, which are in line with the
business of the customer and incentivize the provider
to supply welfare-maximizing quality. On this basis,
penalty fees can easily be calculated for as the
difference between the agreed price (for the agreed
service level objective) and the output of the price
function (for the effectively met quality).
Section 2 discusses related work. Section 3
defines the price function for penalty calculation
referring to a k-pricing scheme for a single service
scenario and for a scenario with multiple
interdependent services. Section 5 concludes with a
summary and an outlook on future research.
2. Related work
In the field of match making and pricing of
services at a certain quality level, considerable
research provides several pricing schemes in the Grid
context (e.g. [2], [8] for auction and commodity
market models and [6] for an overview of state-ofthe-art
bargaining models).
A review of current research of penalty
calculation in the context of Grid Computing showed
that there is no satisfactory comprehensive research
in this area. In their paper, Yeo and Buyya [9]
presented a proportional share allocation technique
called “LibraSLA” that takes into account the utility
of accepting new jobs into the cluster based on their
SLA. Thereby a linear penalty function that reduces
the profit of a job over time after the lapse of its
deadline is introduced. Equally, a linear penalty
function that reduces profit after a jobs run time has
been presented by [1]. Multiple interdependent
services are not considered in related work.
Since linear penalty functions are not sufficient
to determine penalties according to the business
valuations of service provider and consumer in
practice, this work focuses specifically on presenting
a framework for penalty calculation according to the
business value of the agreed services for the
customer, which also incentivizes the provider to
perform at the best. In doing so, the k-pricing
scheme, well-known as a method for pricing in
double-auctions [4] and already applied for pricing in
Grid context [5] is used to come up with a fair and
incentive compatible price function for all possible
qualities of multiple, interdependent services. In the
following we employ the intuition of k-pricing to the
calculation of penalties.
3. Penalty calculation
3.1. SLA violation
This section explains what SLA violation means
in the context of Grid and how SLA violation could
be measured. For the purpose of this paper, following
general definition of SLA violation is sufficient:
“A service level agreement is violated, as soon as it
is impossible to meet the agreed service level
objective(s).”
The determination of the degree of the breach is
important for penalty calculation to enable penalty
design according to the occurring loss of the
customer. For the purpose of this paper, the quality of
service q is arbitrary normalized on a percentage rate
basis, i.e. q ∈ 0,100 , in order to provide
comparability of the degree of SLA breaches. For
example, availability measured in percent could be
considered as normalized service quality.
3.2. Requirements on penalty scheme
The value of the penalty fee of a SLA has deep
impact on the SLA negotiation and the decision of
accepting or rejecting an offer. Along with the price
for a service and the reputation and reliability of the
prospective associate partner the penalty in case of
violation constitute the basis of decision-making
whether to accept an offer or not. Hence following
objectives for penalty calculation can be identified:
Requirement 1: Penalties should be fair. The
penalty agreed in a SLA should be proportionable to
the loss triggered by violation. Penalty calculation
has to be specific to customer`s (and provider`s)
business and should truly establish risk sharing [3].
Requirement 2: Penalties should incentivize
contractual partners to perform best for each
other. The penalty agreed in the SLA should
incentivize the provider to fulfill the agreement
satisfactory to the customer concerning the delivered
Quality of service (QoS). This effect can be achieved
by establishing a penalty proportional to the degree
of SLA violation and by implementing a reward,
which will be paid, if the agreed QoS is
215

outperformed. Although the amount of this penalty
and reward respectively has to be greater than the
cost the provider may save by not performing at its
best. Furthermore, the business value of the provided
service level for the customer has to be taken into
account, since the consumer should not be worse off,
if the provider increases service quality (suppose the
business value of a service will diminish as from a
certain value).
According to this, the business value of the
service for the consumer as well as the production
costs expressed by the reservation price for the
provider have to be taken into account when
calculating the penalty and reward respectively in
SLAs.
3.3. Required information and assumptions
This section describes the needed information
for proper penalty calculation. Thereby a single
service scenario, which is about one single SLA
between one service consumer and one provider that
includes exactly one service, as well as a multiple
service scenario is considered. The multiple service
scenario is about multiple, interdependent SLAs
between one service consumer and one service
provider, whereas every service is represented by
exactly one SLA.
Information 1: Business value of service
qualities for the consumer. The business value BV
of a service for the consumer can be expressed as a
function of the quality q of a service. Thereby, the
business value represents the amount of currency
units CU the service is worth for the consumer at
each possible quality of a service. For example, the
consumer is willing to pay $1.5 for every additional
percentage point of availability of a service. Then its
business value of this service can be expressed as a
straight line BV q = 1.5q [$] in the domain
q ∈ [0,100]. The cardinal definition based on the
measurement in CU is essential for penalty
calculation, since a same base for the functions BV
and the provider`s valuation function R (introduced
below) is indispensible for penalty calculation.
Furthermore, the strength of preferences (slope of
function BV) has magnificent influence on the
penalty calculation. The quality is defined as the
provided quality relative to the determined maximum
quality (e.g. availability expressed in percent). Thus,
the business value of a single service can be
expressed as follows:
for single services and
BV: q → IR + ∀q ∈ [0,100]
BV: q 1 , … , q n → IR +
∀q i ∈ [0,100]
for the multiple service scenario respectively,
whereas q i represents the quality of service i.
Thereby, the business value represents the amount of
currency units CU each possible combination of
qualities of the n services is worth for the consumer.
This graph mirrors the consumer`s preferences about
the quality of service and could vary for every
consumer as well as for every service considered.
To simplify the design of incentive constraints,
it is assumed that the function BV(q i ) is completely
differentiable in every direction in the domain
q i ∈ [0,100] for all possible i = 1, … , n. Since the
business value of a service for the consumer has to be
determined in advance of the negotiation and match
making process, it is assumed that consumers have
already disclosed their valuations to the market
directory when it comes to penalty calculation.
Furthermore it is assumed that consumers truly
disclose business values and do not distort their
preferences in order to manipulate penalty
calculation. This assumption will be discussed in
section 5.
Information 2: Reservation price of the
service provider. Equally, the reservation price R
(measured in CU), which represents the value of the
service for the provider can be expressed as a
function of the quality q of service
R: q → IR + ∀q ∈ [0,100]
in the single service scenario and as following graph
R: q 1 , … , q n → IR +
∀q i ∈ [0,100]
in the multiple service scenario. Thereby the same
assumptions as for the Business Value BV(q i ) apply.
Opportunity costs occurring during service execution
because of arising business alternatives are not taken
into account, since they are very hard to measure and
the service provider will have no incentive to
disclose his true estimation.
216

Information 3: Quality of service according
to SLA. The quality of service q ∗ defined in the
service level objectives (SLO) of the SLA is
necessary to determine whether a penalty has to be
paid by the provider or a reward has to be paid by the
service consumer. For multiple services several SLO
have been negotiated:
BV [CU]
q i ∗ ∈ 0,100 ∀ i = 1, … , n
Information 4: Price of service according to
SLA. In the negotiation process both parties have
agreed upon one certain price p ∗ for the services
provided at the quality q ∗ i . Thereby it is assumed,
that both parties have agreed upon one certain all
round price p ∗ for all services i provided at the
agreed qualities q ∗ i . This assumption is maintainable,
since all considered services, each represented by one
SLA, usually belong to one contract between the
service provider and consumer. All these services
included in one contract should be negotiated
together, since interdependencies between the
services could have significant influence on the
valuation for both contractual partners. Therefore,
these interdependencies already have been taken into
account in the negotiation process. The result of the
negotiation is a bundle of service level objectives
(q ∗ 1 , q ∗ ∗
2 , … , q n−1 , q ∗ n ) and one agreed all round price
p ∗ ∈ IR + in currency units [CU].
p ∗ ∈ IR + ,
measured in currency units [CU]
This price p ∗ will be basis for penalty and reward
calculation respectively via a price function
described subsequently.
Figure 1. exemplary shows the initial
information in the single service scenario: business
value for the consumer BV (q), reservation price of
the provider R (q), agreed price p ∗ and quality of
service q ∗ according to SLA.
The exemplary function BV may be typical for
IT services: It increases slowly at low qualities,
because very low service quality (e.g. availability)
usually will have no useful effect to the consumer.
From a certain quality level on, BV may increase
progressive. At high quality levels marginal utility
could diminish caused by saturation.
3.4. Incentive and fairness constraints
In the following several constraints for
designing a price function P(q i ) are described, which
secure to meet the objective of establishing fairness
(fairness constraints) as well as incentivizing the
provider to act satisfactory to its customer (incentive
constraints). This function P(q i ) represents the total
amount of currency units (CU) the consumer will
have to pay for the services. The penalty and the
reward respectively can be calculated as the
difference between the agreed price p ∗ and the output
P(q i ) of effectively met qualities q i . P(q i ) can be
defined as follows:
P: q 1 , … , q n → IR +
∀ q i ∈ [0,100]
whereas following constraint has to be fulfilled:
P q 1 ∗ , … , q n
∗
= p ∗
This constraint guarantees the agreed price p ∗ for met
service level objectives q i ∗ , i.e. no penalty or reward
has to be paid, if the agreed qualities are met exactly.
BV(q)
R(q)
p*
q*
0 50 100
q [%]
To implement a SLA pricing design which
maximizes overall welfare, several constraints have
to be added. In this context, the welfare which has to
be distributed among the contractual partners is
defined by the difference W q = BV q – R(q) for
every possible service quality q. To clarify the
explanations of the following constraints, only one
service is regarded and a case differentiation is
introduced in order to reduce complexity:
Figure 1. Initial situation for single service
217

a. ∀q ∈ q
∂R(q)
∂q
∂R q
∂q
≤
∂BV q
∂q
≤ ∂P(q)
∂p
, R q ≤ BV(q)}:
≤ ∂BV(q)
∂q
R q ≤ P q ≤ BV q
(IC)
FC
Case a. represents the situation that the given
quality of service is rated higher by the consumer
BV q than by the provider R q . Furthermore the
slope of function BV(q) is greater than the slope of
reservation price R(q). In this case welfare W(q) is
positive and increases with rising service quality q.
To incentivize the service provider to deliver higher
quality in order to maximize overall welfare, the
price P(q) has to increase faster than the reservation
price R(q). Although, the price P(q) has to increase
slower than the business value BV q to secure that
the consumer will not be worse off, if the provider
delivers higher quality. These requirements are
expressed by the incentive constraint (IC). The
fairness constraint (FC) guarantees that no party will
have to bear a loss as long as overall welfare is
positive.
∂R q
b. ∀q ∈ {q|
∂q
∂R(q)
∂q
≤
∂BV q
∂q
≤ ∂P(q)
∂p
, R q ≥ BV q }:
≤ ∂BV(q)
∂q
BV q ≤ P q ≤ R q
(IC)
FC
On the other hand case b. represents the
situation that reservation prices of the provider are
higher than the monetary utility for the consumer
R q > BV q . Furthermore the slope of function
BV(q) is greater than the slope of the reservation
price function R(q). This means that overall welfare
W(q) is negative but increasing with rising service
quality q. The incentive constraint (IC) implements
the incentive for the service provider to supply higher
quality in order to maximize overall welfare without
making the consumer being worse off, i.e. the price
P(q) has to increase faster than its reservation price
R(q) and slower than BV q . The fairness constraint
(FC) guarantees that no party will make a profit as
long as overall welfare is negative.
c. ∀q ∈ q
∂R q
∂q
∂BV(q)
∂q
>
∂BV q
∂q
≤ ∂P(q)
∂q
, R q ≤ BV q :
≤ ∂R(q)
∂q
(IC)
R q ≤ P q ≤ BV q
FC
Case c. represents the situation of positive
overall welfare W(q), which is decreasing with
rising service quality q. To incentivize the service
provider to deliver lower quality in order to
maximize overall welfare, the price P(q) has to
increase slower than R(q), but faster than the
business value BV q to secure that the consumer
will not be worse off if the provider decreases
quality. Again, the fairness constraint secures that no
party will have to bear a loss as long as overall
welfare is positive.
d. ∀q ∈ q
∂R q
∂q
∂BV(q)
∂q
>
∂BV q
∂q
≤ ∂P(q)
∂q
, R q ≥ BV q :
≤ ∂R(q)
∂q
BV q ≤ P q ≤ R q
FC
(IC)
Finally, in case d. overall welfare W(q) is
negative and decreasing with rising service quality q.
To incentivize the service provider to deliver lower
quality in order to maximize overall welfare, the
price P(q) has to increase slower than the reservation
price R(q) of the provider but faster as the business
value to the consumer again. The fairness constraint
guarantees that no contractual partner will make a
profit as long as overall welfare is negative.
Note that following assumption forms the basis
of these constraints: If the marginal reservation price
of the service provider equals the marginal revenue
of the service provider, it will act according the
preferences of the service consumer.
In conclusion, these constraints can be extended
for the multiple service scenario as follows:
s.t.
P: q 1 , … , q n → IR +
∀q i ∈ [0,100]
P q 1 ∗ , … , q n
∗
= p ∗
and the incentive constraint IC :
min ∂R q i
∂q i
, ∂BV q i
∂q i
≤ ∂P q i
∂q i
≤ max{ ∂R(q i)
, ∂BV(q i)
} ∀q
∂q i ∂q i ∈ [0,100]
i
218

and the fairness constraint (FC):
min R q i , BV q i ≤ P q i
≤ max{R q i , BV(q i )}
∀i = 1, … , n
Every function P(q i ) which fulfills these
constraints incentivizes the service provider to act in
a way that maximizes overall welfare and guarantees
that no party will bear a loss, if overall welfare is
positive (and no party will make a profit, if overall
welfare is negative respectively), i.e. the objective
“service provider and consumer win and lose
together” is implemented.
Even for not completely differentiable valuation
functions BV(q i ) and R(q i ), a price function P(q i ),
which fulfils the requirements can be designed by
introducing steps according the considerations made
above.
The introduced constraints do not determine the
distribution of welfare to service provider and
customer. This profit distribution is depending on the
design of the price function P q i according to these
constraints. An exemplary price function which
guarantees fair distribution of the overall welfare is
introduced in the following section.
3.5. Using k-pricing for penalty calculation
for single services
This section aims to come-up with a price
function that fulfils the incentive constraints
introduced above for a single service scenario, i.e.
only one SLA between one service provider and one
consumer is considered. The resulting penalty of a
SLA can easily be computed as the difference of the
agreed price p ∗ for met service level objective and
the output of the price function for effectively met
service quality q:
Penalty(q) = p ∗ − P(q)
Thereby a negative penalty is called a reward, since it
has to be paid from the consumer to the provider for
outperforming services, i.e. if the effectively met
service quality is higher than the agreed service level
objective.
Sattherthwaite and Williams [4] presented a k-
pricing scheme to distribute the welfare of trades in
double auction markets. The underlying idea of the k-
pricing scheme is to determine prices for a buyer and
a seller on the basis of the difference between their
valuations. For instance, suppose that a buyer wants
to purchase a computation service for $5 and a seller
wants to sell for at least $4. The difference between
these bids is W = 5 − 4 = 1, where W is the surplus
of this transaction that can be distributed among the
participants. For a single service exchange, the k-
pricing scheme can be formalized as follows [5].
According to chapter 0, let BV q = bv be the
valuation of the consumer and R q = r be the
valuation of the provider of the service at a given
quality level q. The price for the service of quality q
can be calculated by P q = kbv + 1 − k r with
k ∈ [0,1]. In regard to the above-mentioned
valuations for the service with the quality q, the
resulting price is P q = 0.5 ∗ 4 + 1 − 0.5 ∗ 5 =
4.5 [$] for k = 0.5.
Sattherthwaite and Williams assumed that
BV q ≥ R(q), i.e. the consumer has a valuation for
the commodity which is at least as high as the seller`s
reservation price. Otherwise no trade would occur.
This assumption is not applying in the context of
penalty calculation, since (unwished) situations could
occur in which the business value for the consumer is
lower than the reservation price of the provider. E.g.,
suppose the provider has served 100% availability
until a certain point of time and suddenly it is unable
to provide the service due to an electrical power
outage until the agreed time is past. Then the met
quality of service could be unexpected low and the
situation mentioned above could occur.
The k-pricing scheme can also be applied to
penalty calculation in SLAs in order to provide a
price function that guarantees fair distribution of the
welfare according to the constraints introduced in the
previous chapter for every possible valuation
functions BV q and R q :
P q ≔ k ∗ BV q + 1 − k ∗ R q
∀q ∈ [0,100]
whereas following equation determines parameter k:
k ∶= p∗ − R(q ∗ )
BV q ∗ − R(q ∗ )
This equation guarantees the fulfilment of the
constraint P q ∗ = p ∗ and secures the distribution of
positive and negative overall welfare according to the
welfare distribution agreed in the SLA negotiation
process (for the price p ∗ of the service at the quality
219

level q ∗ ). Figure 2. shows the price function P(q) for
the exemplary initial situation showed in Figure1.
BV [CU]
P(q)
BV(q)
R(q)
p*
q*
0 50 100
q [%]
Figure 2. Price function calculation using k-
pricing (k=0.8)
3.6. Using k-pricing for penalty calculation
for multiple services
Provided that only one overall price has been
negotiated (see section 3.3) to meet exactly one
service level objective for every service, k-pricing
could be used to calculate fair, incentive compatible
penalties for multiple services as follows:
with
P q 1 , … , q n ≔
k ∗ BV q 1 , … , q n + 1 − k ∗ R q 1 , … , q n
k ∶=
∀q i ∈ 0,100 ; i = 1, … , n
p ∗ − R(q 1 ∗ , … , q n ∗ )
BV q 1 ∗ , … , q n
∗
− R(q 1 ∗ , … , q n ∗ )
Again, this equation guarantees the fulfilment
of the constraint P(q 1 ∗ , … , q n ∗ ) = p ∗ and secures the
distribution of positive and negative overall welfare
according to the welfare distribution agreed in
negotiation process. The incentive constraints
determined in section 3.4 are met by this price
function.
3.7. Application in Grid Markets
Since the mathematical formulation of valuation
functions, especially for n-dimensional valuations in
the multiple service scenario, would be too complex
in practice, a simplified framework with tiered
valuation functions is introduced in this section.
To ease the formulation of valuation functions,
the quality range of all services s i is divided into a
certain number j i of intervals. The relevant service
qualities have to be valuated by the provider and the
consumer with one value for each interval in the
single service scenario and with one value for each
possible combination of quality domains in the
multiple service scenario (since services are
interdependent). Thereby, the number of intervals j i
does not necessarily have to be equal for all services
and the length of intervals does not necessarily be the
same within one service. These parameters have to be
set according to the usual structure of the valuation
functions for every service. The parameter setting
stands for a trade-off between convenience and
accuracy, i.e. a high number of short valuation
intervals will provide better incentives, but will also
raise complexity. It must be pointed out, that
complexity increases very fast with the number of
relevant services. With two services and five quality
intervals for each service, 5 2 = 25 combinations
have to be valuated by both contractual partners.
With three services and the same classification,
already 5 3 = 75 combinations have to be valuated.
This classification of quality levels leads to step
functions according to the business value and the
reservation price of the services, which could be
discontinuous at the interval bounds. In the intervals
the slope of the functions is zero. According to this,
the valuation functions generally can be defined as
follows:
Information 1 (simplified): Business value
BV(q i ) of service qualities for the consumer for n
services:
BV:
n
i=1
Q i
→ IR + ∀i = 1, … , n
whereas BV is a graph spanned over the Cartesian
product
with
n
i=1
Q i
= Q 1 × … × Q n ∶=
q 1 , … , q n |q i ∈ Q i ; i = 1, … , n
Q i : = (q 1,1 , q 1,2 , … , q 1,j1 −1, q 1,j1 ).
Thereby Q i represents the set of quality intervals
q i,1 , q i,2 , … , q i,j1 −1, q i,j1 of service i, whereas j i is the
number of quality intervals of service i. In other
220

words, all possible combinations of quality intervals
of the different services have to be valuated by the
consumer by allocation of one positive, monetary
value.
Information 2 (simplified): Reservation price
R(q i ) of the service consumer for n services.
Equally, the reservation price of the provider can be
expressed as a graph:
R:
n
i=1
Q i
→ IR + ∀i = 1, … , n
whereas R is spanned over the Cartesian product
n
i=1 Q i again.
With information about the results of the
negotiation process, as a bundle of service level
objectives (q ∗ 1 , q ∗ ∗
2 , … , q n−1 , q ∗ n ) and one agreed all
round price p ∗ for all the services, the price function
applying k-pricing can be written as follows:
with
P q 1 , … , q n ≔
k ∗ BV q 1 , … , q n + 1 − k ∗ R q 1 , … , q n
k ∶=
∀q i ∈ 0,100 ; i = 1, … , n
p ∗ − R(q 1 ∗ , … , q n ∗ )
BV q 1 ∗ , … , q n
∗
− R(q 1 ∗ , … , q n ∗ )
This equation guarantees the fulfilment of the
constraint P(q 1 ∗ , … , q n ∗ ) = p ∗ and secures the
distribution of positive and negative overall welfare
according to the welfare distribution agreed in
negotiation process. Of course, fulfilling the
incentive compatibility conditions presented in
section 3.4 is not feasible with tiered functions. Since
these step functions could be considered as
approximation and pass into real valuation functions
for sufficiently small intervals, incentive
compatibility can be achieved in some degree
anyway.
Figure 3. visualises k-pricing in case of the
simplified scenario by using step functions for
service valuation for a single service.
BV [CU]
P(q)
BV(q)
R(q)
p*
q*
0 50 100
q[%]
Figure 3. Simplified step functions BV(q),
R(q) and P(q) using k-pricing (k=0.8)
In the following a numerical example with two
interdependent services is drawn to clarify the
(simplified) k-pricing scheme in multiple services
scenario. For the sake of clarity a graphic account has
been left out.
In this example, two interdependent services are
considered. These services are supposed to be
complements for the consumer, i.e. these services
should be consumed together. In economics, a good 1
is a complement to good 2, if the demand for good 1
goes down when the price of good 2 goes up [7]. A
famous example of complement goods would be
hamburgers and hamburger buns. In a Grid market,
CPU and storage could be considered as complement
goods, since there may be only very limited use for
an application to purchase computing power without
having the appropriate memory capacity. The same
services could be of substitute character to the
service provider, for example if the provision of these
services is independent from each other. Then the
reservation price for both services will simply be the
sum of the reservation prices of the two single
services. Following tables show exemplary
valuations of two services, which are of complement
character for the consumer and of substitute character
for the provider respectively.
Table 1. Exemplary business value BV(q 1 , q 2 )
of complements for the consumer
q 2,1 q 2,2 q 2,3 q 2,4
∈ 0,25 ∈ [25,50) ∈ [50,75) ∈ [75,100]
q 1,1 ∈ 0,25 0 5 10 20
q 1,2 ∈ [25,50) 5 50 60 70
q 1,3 ∈ [50,75) 10 60 100 110
q 1,4 ∈ [75,100] 20 70 110 150
221

Table 2. Exemplary provider`s reservation
price R(q 1 , q 2 ) of substitutes
q 2,1 q 2,2 q 2,3 q 2,4
∈ 0,25 ∈ [25,50) ∈ [50,75) ∈ [75,100]
q 1,1 ∈ 0,25 10 20 30 40
q 1,2 ∈ [25,50) 20 40 50 60
q 1,3 ∈ [50,75) 30 50 60 70
q 1,4 ∈ [75,100] 40 60 70 80
Table 4. Penalties and rewards for different
service levels
q 2,1 q 2,2 q ∗ 2 = q 2,3 q 2,4
∈ 0,25 ∈ [25,50)] ∈ [50,75) ∈ [75,100]
q 1,1 ∈ 0,25 72.5 63.75 55 45
q 1,2 ∈ [25,50) 63.75 37.5 27.5 17.5
∗
q 1,3 ∈ [50,75) 55 27.5 10 0
q 1,4 ∈ [75,100] 45 17.5 0 -17.5
Suppose the service level objective is to meet a
quality q 1 ∈ [75,100] for service s 1 and quality
q 2 ∈ [50,75) for service s 2 for an all round price
p ∗ = 80 [CU], which has been determined for met
service level objectives in the negotiation process. So
factor k can be calculated as
k =
p ∗ − R q 1 ∗ , q 2
∗
BV q 1 ∗ , q 2
∗
− R q 1 ∗ , q 2
∗
= 0.25
=
80 − 70
110 − 70
With the equation presented above the price can be
calculated for every possible combination of services.
P q 1,a , q 2,b =
k ∗ BV q 1,a , q 2,b + 1 − k ∗ R q 1,a , q 2,b
∀ a = 1, … ,4 and ∀ b = 1, … ,4
Table 3. shows the results of a price calculation.
Table 3. Prices P(q 1 , q 2 ) calculated via k-
pricing (k = 0, 5)
q 2,1 q 2,2 q ∗ 2 = q 2,3 q 2,4
∈ 0,25 ∈ [25,50) ∈ [50,75) ∈ [75,100]
q 1,1 ∈ 0,25 7.5 16.25 25 35
q 1,2 ∈ [25,50) 16.25 42.5 52.5 62.5
∗
q 1,3 ∈ [50,75) 25 52.5 70 80
q 1,4 ∈ [75,100] 35 62.5 80 = p ∗ 97.5
The calculated prices are always between the
valuations of the relevant service partners of the
contractual partners. For the agreed service level
objective q 1 ∗ ∈ [75,100] and q 2 ∗ ∈ [50,75) the
calculated price P(q 1 ∗ , q 2 ∗ ) equals the agreed price p ∗ .
From these prices again the penalty and reward
respectively can be calculated as
Penalty(q 1 , q 2 ) = p ∗ − P(q 1 , q 2 )
whereas q i is the actually met quality of service i.
For the agreed service level objective
q 1 ∗ ∈ [75,100] and q 2 ∗ ∈ [50,75) the penalty is zero,
since the calculated price P(q 1 ∗ , q 2 ∗ ) equals the agreed
price p ∗ . The penalty for q 1 ∈ [50,75) and q 2 ∈
[75,100] is also zero, because of the symmetry of the
valuations in this example If the SLO is
outperformed, i.e. q 1 ∈ [75,100] and q 2 ∈
[75,100], the consumer will pay a reward (negative
penalty) to the service provider.
Table 5. Benefit of service provider
P(q 1 , q 2 ) − R(q 1 , q 2 )
q 2,1 q 2,2 q ∗ 2 = q 2,3 q 2,4
∈ 0,25 ∈ [25,50) ∈ [50,75) ∈ [75,100]
q 1,1 ∈ 0,25 -2.5 -3.75 -5 -5
q 1,2 ∈ [25,50) -3.75 2.5 2.5 2.5
∗
q 1,3 ∈ [50,75) -5 2.5 10 10
q 1,4 ∈ [75,100] -5 2.5 10 17.5
Table 5. shows the benefit of the service
provider at each quality level combination calculated
as the difference P(q 1, q 2 ) − R(q 1, q 2 ). Table 6.
presents the associated benefits for the consumer,
calculated equivalent as the difference between its
valuation and the determined price BV(q 1, q 2 ) −
P(q 1, q 2 ).
Table 6. Benefit of service consumer
BV(q 1 , q 2 ) − P(q 1 , q 2 )
q 2,1 q 2,2 q ∗ 2 = q 2,3 q 2,4
∈ 0,25 ∈ [25,50) ∈ [50,75) ∈ [75,100]
q 1,1 ∈ 0,25 -7.5 -11.25 -15 -15
q 1,2 ∈ [25,50) -11.25 7.5 7.5 7.5
∗
q 1,3 ∈ [50,75) -15 7.5 30 30
q 1,4 ∈ [75,100] -15 7.5 30 52.5
As can be seen from these tables, both parties
will bear a loss, if at least one service is delivered at
the lowest quality. As soon as both services are
provided in a quality greater or equal than 25%, both
parties will make a (positive) profit, which is
222

increasing with rising quality of both services. For
every possible change of service qualities making the
provider being better off, the service consumer will
also be better off. Hence, the provider is incentivized
to provide both services in best possible quality to
maximize its profit. Due to the fixed ratio of overall
welfare distribution provided by the k-pricing
scheme, this will maximize the benefit of the
consumer too. In this case the SLO is outperformed
by the service provider and an overall welfare even
greater than the welfare in the originally agreed
situation will be distributed fairly.
4. Conclusion
Penalty calculation is crucial for true risk
sharing in SLAs, since it determines the distribution
of the associated risk of the SLA among the
contractual partners. The k-pricing scheme provides
fair and economical reasonable prices for every
quality level of a single service. Thereby the overall
welfare is distributed at a fixed ratio according to the
result of a negotiation process for all possible met
service qualities. Thus, the party with the bigger
share of positive welfare will also bear a higher share
of the loss, if the SLA is violated, i.e. the party which
gets a bigger share of the positive welfare also bears
a higher risk. Due to these characteristics and the
fact, that k-pricing fulfils the incentive constraints
presented in section 3.4 (proof trivial) and provides
fair welfare distribution, k-pricing is the ideal scheme
for penalty calculation in automated markets.
However, the problem of using k-pricing for penalty
calculation is that it provides incentives for market
participants to distort their true valuations. Suppose a
service consumer is interested in a service at one
specific level of quality. Then he is incentivized to
understate its valuation for all other service levels in
order to maximize penalties in case of violated SLA
and to minimize rewards in case of outperformed
SLO. This incentive problem cannot be solved in
context with k-pricing, since the price function is
always directly associated with the valuations and
thus directly influence able. Thus, the incentive to
report true valuations has to be provided by other
components in a Grid Market, maybe the Market
Mechanism, which matches supply and demand. Just
as well the creditability of the contractual partners
regarding the disclosed valuations could indirectly be
gauged by observing valuations for similar services
in the past. But surely the incentive to distort
valuations will remain as a challenge to be solved.
References
[1] Irwin D.E., Grit L.E., Chase J.S. (2004): Balancing
Risk and Reward in a Market-based Task Service. In
Proceedings of the 13th International Symposium on High
Performance Distributed Computing (HPDC13), Honolulu,
Hawaii, June 2004, pp. 160-169.
[2] Joita L., Rana O.F., Gray W., Miles J.(2004): A Double
Auction Economic Model for Grid Services. Proceedings
of Euro-Par 2004 Parallel Processing - 10th International
Euro-Par Conference, Pisa, Italy, 31 August - 3 September
2004, ISBN 3-540-22924-8 Springer Berlin Heidelberg
New York,
http://www.springerlink.com/index/XLADBT1AW3HPXC
8L .
[3] Quick A. (2007): Service Level Agreements (SLA)
Boot Camp. http://helpdesk.wyopub.com/2005/07/servicelevel-agreement-sla-boot-camp.html
, recalled May 14
2007.
[4] Sattherthwaite, M. A., Williams, S. R., (1993). The
Bayesian Theory of the k-Double Auction. In: Friedman,
D., Rust J. (Eds.), The Double Auction Market -
Institutions, Theories, and Evidence. Addison-Wesley, Ch.
4, pp. 99-123.
[5] Schnizler B., Neumann D., D. Veit, C. Weinhardt
(2006): Trading Grid Services – A Multi-attribute
Combinatorial Approach. European Journal of Operations
Research. Forthcoming, Download from Science Direct.
[6] Sim K. M. (2006): A Survey of Bargaining Models for
Grid Resource Allocation. ACM SIGECOM: E-commerce
Exchange, Vol. 5, No. 5, January, 2006.
[7] Varian H. R. (2003): Intermediate microeconomics: a
modern approach. 6 th ed, Norton, New York, 2003.
[8] Wolski R., Brevik J., Plank J., Bryan T. (2003): Grid
Resource Allocation and Control Using Computational
Economies. In F. Berman, A. Hey, and G. Fox (eds.): Grid
Computing – Making the Global Infrastructure a Reality. J.
Wiley, New York, 2003.
[9] Yeo C.S., Buyya R. (2005): Service Level Agreement
based Allocation of Cluster Resources: Handling Penalty to
Enhance Utility. Proceedings of the 7th IEEE International
Conference on Cluster Computing (Cluster 2005, IEEE CS
Press, Los Alamitos, CA, USA), Sept. 27 - 30, 2005,
Boston, Massachusetts, USA.
223

20 th Bled eConference
eMergence:
Merging and Emerging Technologies, Processes, and Institutions
June 4 - 6, 2007; Bled, Slovenia
Rightsizing of Incentives for collaborative e-Science
Grid Applications with TES
Arun Anandasivam, Dirk Neumann, Christof Weinhardt
Institute for Information Systems and Management,
University of Karlsruhe, Germany
{name}@iism.uni-karlsruhe.de
Abstract
The particle physics community is sharing its resources over a Grid network.
Currently resources are used on availability. But there is no fair allocation
considering the importance or the private value of a job. Researchers are using
Grid resources regardless of whether others need them more urgently.
Furthermore, they do not even provide their own resources as they do not have an
incentive for sharing them. Researchers are demanding a Grid infrastructure
where an incentive mechanism is implemented. This will support a fair allocation
of resources. Incentives can be provided by using money. But one constraint is
that no real money should be involved for billing the resources in the science
community. In this paper, a mechanism called Token Exchange System (TES) is
proposed. Considering the requirements from the particle physics community this
mechanism combines the advantages of reputation and payment mechanism in
one coherent system to realize a fair allocation. It enables to build up a Grid
infrastructure with an incentive mechanism which is scalable, incentivecompatible
and does not require a face-to-face agreement like the current system.
Keywords: Token Exchange System, incentive mechanism, Grid network
1 Introduction
In the e-Science community, Grid is an important infrastructure for the
collaboration of scientific researchers to run calculations in a reasonable period of
time. Combining distributed computing resources together and sharing these
resources with others is demanded by research institutes. Especially researchers
from particle physics are interested in sharing resources. They have a great
resource demand for analyzing the data collected from the particle detector. The
large amount of data comprising 4 million Gigabytes a year that is generated by
the Large Hadron Collider experiment cannot be stored and processed by a single
research institute (Berlich et al. 2005). Currently the GSI institute in Darmstadt,
Germany, with its focus on particle and ionic physics produces about 1 Gigabyte
data per second with its detectors. The various institutes take an interest in
cooperative data analysis for saving precious time. Any institute can contribute his
224

esources to the e-Science community and can in turn use the resources of others.
The technical infrastructure allowing such coordinated resource sharing is
provided by the Grid environment so as to realize an e-Collaboration among the
researchers.
This resource sharing approach, however, is problematic due to free-riding as
known from Peer-to-Peer coupled with the prevailing competition in the research
field (Adar, Huberman 2000). It turns out that even in Academia researchers are
quite not willing to share resources or knowledge with other institutes without
reciprocity. Furthermore, in the particle physics the resources are not utilized
efficiently. Users tend to send too many jobs on others’ computing nodes. To
avoid overconsumption, the sharing of resources has to be compensated to enable
a fair transaction, which is not trivial in this context. This is not a “tragedy of the
commons” problem (Hardin 1968, Ostrom 1990), since the resources belong to
institutes or research projects, who are the owner. The need for more resources
from other institutes is due to analysis of the large amount of data.
In industry, compensation for shared resources is typically made by transferring
money. Money is used to pay the counterpart a compensation for having access to
the resources like applications (Application Service Providing), information or
services. In principle, this payment system is also applicable to the scientific
communities. However, the community faces major problems by implementing
this system: researchers deny the usage of money for sharing resources for two
main reasons. Firstly, researchers are not allowed to spend money without
presenting valid reasons for their expenses to the sponsor. They always take care
of the investment they make. An investment can lead to revenue. In some
countries it is prohibited by law for institutes to make profit (e.g. Germany).
Secondly, it is regarded unethical that those institutes with a deeper pocket are in a
position to do better research than others. Thus, it is not surprising that researchers
tend to prefer reputation mechanisms rather than money as incentive mechanisms.
In reputation mechanisms the researchers do not rate the value of resources, but
the behavior of the user. A disadvantage of reputation mechanism is the lack of
incentive for motivating researchers to offer scarce resources. The participants do
only the bare minimum to retain an acceptable reputation level that allows
participating in the exchange process without restrictions. This means that there is
no clear incentive to share scarce resources (i.e. high computational power).
The main contribution of this paper is to give an overview of the requirements of a
mechanism in the scientific community, compare former research work regarding
the requirements and present a tailored mechanism for the e-Science Grid
community. The incentive mechanism we propose is called Token Exchange
System (TES). TES promises to achieve sharing resources in e-Science without
money while retaining the beneficial coordination effect of money. TES is based
on a reputation mechanism and a payment mechanism merged into one system.
We use tokens instead of money for compensation. These tokens have a certain
value. In contrast to real money or e-Cash, the tokens can be issued and
distributed by the participants themselves. The value of the tokens is derived from
the reputation of a participant and the number of issued tokens. Several
restrictions are realized through a function calculating the token value to restrain
flooding the whole system with their own tokens to manipulate the system. This
approach is novel, as it combines existing payment and reputation mechanism into
one coherent system. This way, transaction between institutes can be
accomplished without burdening their financial endowment.
2
225

System analysis
This paper is structured as follows: in section 2 requirements and the related work
will be outlined. The model of TES is described in section 3 and the simulation
results are shown in section 4. Section 5 concludes the paper.
2 System analysis
In a Grid network agents can have different behavioral characteristics. They can
be altruistic, selfish or malicious (Yu, Singh 2003). These characteristics are not
fixed per se. The agents can change the way they behave. For instance, Liebau et
al (2005) define a normal agent who behaves selfishly, but to a certain extent
altruistic. This type of agent only shares a few files for a certain period of time.
One goal in a network with different agents is to maximize the overall welfare to
enhance trust in the network. The good agents expect to be treated fairly, if
resources are allocated in the Grid. Otherwise, they will not participate in the
network any longer (in the long run). The idea behind TES is to provide a new
incentive which cannot be fulfilled by current payment and reputation
mechanisms. This will make the agent behave more cooperative and lead to an
enhanced overall outcome. The incentive is set by combining both mechanisms
into one coherent system. The payment mechanism is used to evaluate a
commodity (goods or services). The reputation mechanism is necessary to
evaluate the behavior of a user in a network. In TES the reputation has an impact
on the budget of the users.
2.1 Requirements
The Grid usage for particle physics is one example from a variety of possible
scenarios. Although TES is not limited to the application in particle physics Grid,
we use it as a sample scenario. Different scenarios are defined by different
requirements. The requirements describe the upcoming issues for exchanging
resources in Grid. The most important requirement is the absence of real money in
the Grid (requirement 1). In scientific environment the Grid computing resources
are financed by research projects. These projects are led and controlled by
national governments. One restriction by the national governments is not to earn
money with the government-funded resources. Implementation of a non-monetary
system is, therefore, essential. Instead of money, tokens are used which have other
properties.
Another property is to provide a compensation payment (requirement 2). Grid
resources have to be consumed immediately. A reservation without consumption
is a loss for the provider and other consumers who are waiting to complete their
jobs. Without a payment mechanism (neither real money nor tokens), the provider
will not receive compensation. As real money is not applicable, tokens can be
used for paying compensation to the provider.
Tokens as a payment mechanism can give agents an incentive to share their
scarce resources with others (requirement 3). The particle physics Grid has
several institutes in its projects. They have to share their resources with other
project members, which is not always done consistently. An economic incentive
can elicit the researcher or his institute to contribute high computational power or
machines with specific configuration for certain jobs. These incentives have to be
set by the economic model based on the token payment system. The model must
226
3

e designed in such a way that it enforces the participants to behave in a fair
manner.
Enforcements are implemented on both at the technical and at economical level.
The implementation at both levels has to be complementary. The goal of
enforcements is to achieve fairness and robustness for the participants using the
accounting system (requirement 4). An exchange in the Grid is fair, if the
requester gets the desired service and the provider receives adequate payment for
the service. In case even if one of the rules is not obeyed, the accounting system
has to be robust enough to identify the failure and punish the offending party
(Dasgupta 2000; Papaioannou, Stamoulis 2005).
A Grid network must be capable of accommodating the expanding number of
users. Scalability is, therefore, necessary for technical and volume throughput
expansion (requirement 5). A scalable economic model can be implemented into a
technically well-defined system to guarantee a compound dynamic and flexible
system as emphasized by Forster (Foster et al. 2006).
2.2 Related work
In TES, every agent has his own tokens. For scalability reasons, the underlying
technical infrastructure is supposed to be a decentralized network as known from
P2P. There has been a lot of research work done on implementing payment
mechanisms for P2P and Grid systems (Barmouta, Buyya 2003; Dutta et al. 2003;
Irwin et al. 2005). One example is Karma (Vishnumurthy, Chandrakumar, Sirer
2003). Karma uses a uniform currency called Karma for payment. Every user in
the network can earn and pay with Karma. The problem is the inflation and
deflation of the currency. Furthermore, a reputation mechanism is not considered
in Karma.
A potential combination of payment and reputation was used in PeerTrust system
(Xiong, Liu 2004). They introduced a Transaction Context Factor in their
reputation mechanism to give a big transaction more weight than a small
transaction. The aim is to avoid malicious attacks, where small transactions are
done truthfully and fairly for building up a reputation and then to be dishonest on
a big transaction to earn more profit while not substantially decreasing their own
reputation. Hence, the amount of a transaction has an impact on the reputation.
But the reputation has no impact on the user’s budget.
Liebau et al. (2004) propose a mechanism called Token-based Accounting System
(TbAS) for P2P systems. Tokens serve as a receipt for a transaction between two
peers. The peer who receives a file hands over a token to the file provider. The
provider exchanges the received token into one of his own. The exchange of the
token is done via a public aggregation function, which is not defined. Moreover, a
reputation mechanism is not integrated in the accounting system. In TbAS the
token itself does not have a value. They assume that one token is fixed to a trading
parameter like one token for one Megabyte or for one minute of service. Thus, the
“price” will not vary for one Megabyte, i.e. depending on quality of service or
reputation of provider.
The solution to avoid double-spending in TbAS is done by using a serial number
for each token. However, the double-spending check is done after the transaction.
The problem here is that once the tokens have been exchanged and the reputation
evaluated, there is no possibility for the provider to revoke the process or to get
compensation. In TES the check is always done during the transaction before the
4
227

The model
services are exchanged. There is no way that once spent tokens can be re-offered
and distributed successfully.
Moreton and Twigg (2003) presented a token-based protocol, where the token has
a value and it has an impact on the exchange of resources and payment. The
reputation is represented by the token-value. However, they did not examine how
their calculation of the token value can set incentives for behaving truthfully. The
calculation function is a one way track. Once the peer behaves badly and gets a
bad evaluation, he cannot “clean” his reputation again. On one side it sounds as a
good incentive to behave well. On the other side, if it is an attack where several
peers rate him incorrectly, he has no chance to show his true face later on.
3 The model
In TbAS, each agent in the network is assigned to a group of trusted peers (Liebau
et al. 2004). The tokens are issued by the trusted peers who are assumed to be
trustworthy. These trusted peers manage the accounting information of the agent.
An agent has a certain number of tokens. To consume resources, the agent has to
send a request for the emission of the tokens for payment. The trusted peers check
the amount of available tokens with the amount of requested tokens to avoid
unauthorized token distribution and emit the tokens to the agent. With the
received tokens the agent can consume resources from other agents. For every
utilized service he has to hand over tokens to the service provider. The service
provider redeems the tokens at his trusted peers. This will raise his “budget”. The
foreign tokens from the consumer agent will be changed into the tokens of the
service provider. The exchange rate is a publicly available function. The provider
can use his earned tokens to acquire resources for himself.
3.1 Exchanging the tokens in TES
The function for calculating the exchange rate was not defined by Liebau et al.
From the economic perspective, this is essential for an incentive-based resource
exchange. Tokens can be changed at a predefined ratio, for instance 1:1. In our
approach, the exchange rate is not based on a public function. Every price offer
from the provider side is based on the number of offered tokens and is open to
bargain. For instance, a provider can ask 5 tokens for 1 CPU hour. The requester
can offer only 4 tokens as payment.
The aim of TES is to set an incentive for an unselfish and a non-malicious
behavior. Tokens are deployed to be used as a payment system. The participants
in a network can pay another participant with the tokens. One characteristic of
these tokens is individuality. Every participant has his own tokens, which can be
created and distributed by himself. The agent is the token issuer. For instance,
user A wants to receive a Grid service from user B. A offers his tokens to B. If B
accepts the offer he will be “paid” in A’s Tokens. B as the owner of a certain
number of tokens from A can decide whether he wants to redeem these tokens at
A or to offer them to another participant C. After several transactions a
participants possesses tokens from different users. The incentive of possessing
others’ tokens is to increase the liquidity. Every token from agent A has a token
value v A which can change over time. The token value is influenced by the
reputation. A bad behavior of A has an impact on his token value. If B has tokens
228
5

from A, the value of these tokens will decrease and therefore his budget will
decrease, too. In future, B will think twice about accepting tokens belonging to A.
In the long run the tokens of users with a bad token value (and thus a bad
reputation) will not be accepted widely. The bad users can still earn money by
providing services. However, their token value is further an indicator of their
reputation. Thus, a service provider with a bad reputation has an issue with
canvassing customers. Both ways he has to increase his token value by behaving
well. Three constraints have to be considered in providing this token system.
Axiom 1: Since everybody has his own tokens, these tokens have to be
limited in number (Moreton, Twigg 2003). Otherwise the users can
mint their tokens arbitrarily and have infinite money (like minting
Axiom 2:
your own Cent-Coins).
The reputation has an impact on the token value. An incentive for
emitting tokens is necessary. Otherwise the agents will not
distribute their tokens and their token value will not have an effect
on others’ budget.
Axiom 3: The number of distributed tokens has to be bounded (Axiom 1).
However, a hard constraint will make it less flexible for the user.
Any agent should be able to obtain a credit for an emergency. In
this case, the agent has to pay some kind of interest.
3.2 Calculating the token value
The determination of the token value has to consider these axioms. To fulfill these
axioms the token value function v x of an agent X is calculated as follows
6
v ( m
x
x
, ncr
x
mx
0 x
x
k
l
2
) = max{0, ub − ncr − j * ( e * m − ) } (1)
with v x ∈ [0, …, ub 0 - ncr x ], m x ∈ IN 0 . ub 0 is the upper bound value which has to
be determined for a system. The influence of the reputation value on the token
value is expressed by the non-credibility rate ncr x with 0 as the best reputation
value (similar to Papaioannou, Stamoulis 2005). The higher the value of ncr x the
worse is the reputation. The calculation of the reputation value is only an example.
Other reputation mechanisms can be included by calculating the ncr x value. For
instance, by including a reputation mechanism with a valuation rep x ∈ [0, 1] the
function does not need an upper bound, as it is limited by the reputation value:
v ( m
x
x
, rep
x
m
l
2
) = max{0, rep − j * ( e * m − k)
}
(2)
In this paper we will consider the first function (1). Every distributed token has
the current value v x which is adjusted dynamically based on the two variables m x
and ncr x . The number of distributed tokens m x is necessary to define the liquidity
of an agent X. By spending his own tokens (increasing m x ), the agent has less
tokens for further transaction. He can earn foreign tokens or wait until others
redeem his tokens to increase his budget. But the distribution of his tokens will
raise his token value until a predefined maximum m* is reached.
Furthermore, with the three parameters j, k and l the value function can be
adjusted regarding the initial token value for m x = 0, the maximum token value m*
and the gradient of the decrease of the token value (Figure 1). The gradient is
229
x
x
x

Evaluation
crucial to determine how hard an agent is punished when distributing more than
m* tokens. The option to provide more tokens than the maximum reflects
axiom 3. The agent can provide more tokens, but he has to take his reputation loss
into account (as payment of interest).
The token value cannot be increased infinitely by distributing unlimited number of
tokens (axiom 1). The limitation of the distributed tokens is not a hard constraint,
but it reduces the reputation and the acceptance of the tokens and the transaction
with a negatively rated agent will decrease:
lim
m →∞
x
v ( m
x
x
, ncr
x
) =
lim
m →∞
x
max{0,
l
2
ub0 − ncr − j * ( e * m − k)
} = 0
x
m
x
x
(2)
The fulfillment of axiom 2 is presented under the terms of a simple utility function
of u x (v x ) = v x . For m x,1 , m x,2 ∈ [0, …, m*] with m x,1 < m x,2 the token value is as
follows: v x (m x,1 , ncr x ) < v x (m x,2 , ncr x ). The agent X is always better off by
increasing his token value until m x = m*.
Figure 1 illustrates how the axioms are fulfilled by the token value function.
Figure 1: Token value function
4 Evaluation
4.1 Simulation setup
In this simulation we want to show the effect of reputation on the budget of the
agents. There are two types of agents we name Pastors and Mavericks. An agent’s
behavior can be good (providing full service or transferring payment) or bad
(providing bad service or not transferring payment). As an agent will not always
behave well or badly, a probability parameter decides how well or badly an agent
might act. Thus, the Pastors might behave well with a probability of 90% and a
Maverick might behave well with a probability of 10%. All agents of one type are
symmetric. Different agents do not have different utility functions.
The parameters are defined as follows: ub 0 = 1000; ncr x = 0; j = 0.05; k = 70;
l = 110. The reputation mechanism is based on the credibility mechanism
(Papaioannou, Stamoulis 2005). However, the punishment is not implemented
since any utility function and learning algorithms are not considered yet. The
7
230

eputation mechanism is based on an asymmetric approach, where a bad rating
increases the ncr-value by 10 and a good rating decrease the ncr-value by 2 with 0
as the best value and with no upper bound value.
For the transaction in each case two randomly selected agents will be allocated
together (25 allocations for 50 agents in one round). After the allocation the
transaction begins and is divided into 4 phases. In phase 1 it has to be decided, if
the agent behaves well or badly based on the probability. The second phase is
about how the buyer and the seller are thinking about the reputation of the
counterpart and if they want to settle a transaction with the partner. Pastors will
not deal with agents who have a token value below a threshold of v 0 (·) = 500. In
Phase 3 the offered price and the payment will be determined. The process of
finding the right tokens (token selection process) is divided into three subsections.
The succeeding subsection is only executed, if the preceding subsection was not
successful:
1. At first, the buyer tries to maximize his own reputation. Thus, he
scrutinizes, whether he can distribute his tokens without decreasing his
reputation.
2. If he cannot distribute his own tokens (m x = m*), he will check, if he
possesses enough tokens from the seller.
3. In case he does not have any tokens from the seller, he tries to offer tokens
from other agents.
After this token selection process there can be an offer (in case there are enough
tokens for payment) or an offer is not possible due to lack of enough tokens and
the token value respectively. The price of an offer will be calculated by summing
up the current token value of each token. Now the provider can decide whether he
accepts the offer or not. The price varies randomly between 0 and 10,000. In
Phase 4 the transfer of the tokens and the evaluation of the provider and consumer
are settled. A well behaving agent will be rated well whereas a bad agent will be
rated poorly. The underlying assumption is that the agent will rate each other
truthfully and there are no opportunistic or false ratings.
4.2 Simulation scenarios and results
We analyzed two scenarios for a better understanding of the effects of reputation
on the budget. In the first scenario the Mavericks did not take their individual
token value into account whereas it was considered in the second scenario. If their
token value falls below the threshold value v 0 (·), the Mavericks will change their
strategy and behave well until their token value is over the threshold value again.
The results are compared to the combination of a payment and a reputation
mechanism. The identical reputation mechanism was implemented in both
systems. In the payment system the agents have an initial budget of 41488
currency units. For comparison, this value is derived from the area below the
curve of TES (assuming the agents have the best reputation ncr x = 0):
m*
46
∫ =
vx
0
( m
x
) ∂m
x
=
46
∫
0
1000 − 0.05 * ( e
mx
110
* m
x
2
− 70) ∂m
x
= 41488
(3)
The simulation was run over 20 generations with each 5000 rounds for the two
scenarios. Nevertheless, in the figures only the relevant first 1000 rounds are
visible. In the simple scenario the Mavericks do not care about their own
8
231

Evaluation
reputation. The budget of the Pastors and Mavericks was used as a metric to
compare both systems. As illustrated in Figure 2 the Mavericks earn a lot of
tokens at the beginning where the Pastors trust the Mavericks. After 100 rounds
there are no more transactions between the two groups, because the reputation of
the Mavericks is below a predefined threshold value. The Pastors do not trust the
Mavericks. But Pastors still trust other Pastors. The Mavericks do not have the
chance to trust Pastors, because there is no transaction due to their token value. In
the monetary system, liquidity of the Mavericks is not decreasing whereas in TES
the Mavericks can still “lose money”. This is the major incentive not to behave
selfishly in TES. In the first scenario, the Mavericks only have a small loss,
because their tokens are distributed among all agents. We assumed the worst case,
where the Mavericks wait to behave selfishly until 90% of their tokens are
emitted. It has to be considered that not all tokens of the Mavericks are distributed
to Pastors. The Mavericks do not identify who the Pastors are. Thus, Mavericks
can earn the tokens of Mavericks as well. After 450 rounds the reputation of the
Mavericks cannot get worse anymore. Thus, the budget of each group in TES
remains almost stable. Because the Mavericks behave well and the Pastors behave
poorly with a probability of 10%, the reputation affects the budget marginally.
(a)
Figure 2: Simulation results for payment mechanism (a) and TES (b) in
scenario one (average budget of an agent)
If the Mavericks are concerned about their reputation as in the second scenario,
the difference between both systems is even more significant (Figure 3). Even
though the Mavericks can gain nearly all the money in the system with time, in
TES their budget is worse than the budget of the Pastors. In the first case they
behave well to gain more money from others, if their reputation is below the
threshold value. In TES, the Pastors can recover from the selfishness of the
Mavericks and regain their budget. The reason is a constraint of the acceptance of
the tokens. A Pastor who already has a good token value does not accept tokens
with a low token value. Thus, in a transaction between two agents of each group
the Maverick will accept the low valued token for payment and pay with high
valued tokens, because they need resources. They have no alternative. In the long
run the budget of the Mavericks will get even worse, since they only receive the
worse tokens. The Mavericks are forced to behave well over a long period of time.
In this simple simulation setup, we could present that the Mavericks in TES
perform worse in both scenarios than in a common payment system.
(b)
232
9

(a)
Figure 3: Simulation results for payment mechanism (a) and TES (b) in
scenario two (average budget of an agent)
(b)
5 Conclusion
In our approach, the tokens represent the reputation value and the liquidity of an
agent. A payment is done by exchanging tokens. As everybody has his individual
tokens, his reputation has a direct effect on the token value and thus the liquidity.
With a better performance and a good evaluation of others on his service and
payment the agent can increase his reputation.
TES enables a resource exchange in scientific Grids. Tokens are used for
payments and do not represent real money (requirement 1). The payment via
tokens allows the agents to pay compensation (requirement 2), if the resource of
the provider was not utilized or misused by the consumer. Furthermore, TES
provides an incentive to share scarce resources as the provider can earn more
tokens, if he offers the resources that are in greater demand (requirement 3). From
the economic perspective, the price for the resource (here the number of tokens
with a high value offered) will rise. Fairness and robustness have to be considered
in the reputation mechanism. An appropriate mechanism has to be chosen, which
affords a fair evaluation of the agent’s reputation and an attack-resistant reputation
system (requirement 4). Moreover, the scalability is given by the economically
decentralized token system. A P2P-System like TbAS is a complementary
technically decentralized system (requirement 5).
The simulation proved that TES has a better performance than a common payment
system combined with a reputation mechanism. Up to now in the simulation the
payment in TES does not offer a combination of tokens from different agents
within one transaction. This will lead to more transactions in a round. A
sensitivity analysis with the initial settings of the parameters is required to adapt
the function to different scenarios. Additionally, untruthful feedback of lying
agents can worsen the system.
Another aspect is to introduce an expiration date for the tokens. Firstly, this will
prevent the hoarding of tokens by one user as the tokens will get invalid.
Secondly, if agents join and leave the network the tokens should be taken out of
the system to avoid inflation. An expiration date will automate this process.
All in all, TES is an approach to set an incentive for good behavior and
cooperation the scientific Grid network. Our token based system is tailored to the
needs of the scientific community and is thus a promising starting point to
overcome incentive problems in e-Science.
10
233

Conclusion
References
Adar, E.; Huberman, B. (2000): Free riding on Gnutella. First Monday, Vol.5
No.10.
Barmouta, A.; Buyya, R.(2003): Gridbank: A grid accounting services
architecture (gasa) for distributed systems sharing and integration. In 17th
IEEE International Parallel & Distributed Processing Symposium,
Workshop on Internet Computing and E-Commerce, Nice, France.
Berlich, R.; Kunze, M.; Schwarz, K. (2005): Grid computing in Europe: from
research to deployment. In Proceedings of the 2005 Australasian workshop
on Grid computing and e-research, Vol. 44.
Dasgupta, P. (2000): Trust as a commodity. In Gambetta, D. (ed.): Trust: Making
and Breaking Cooperative Relations, electronicedition.
Dutta, D.; Goel, A.; Govindan, R.; Zhang, H. (2003): The design of a distributed
rating scheme for peer-to-peer systems. In Proceedings of the first workshop
on economics of Peer-to-Peer Systems, Berkeley (USA).
Foster, I.; Kishimoto, H.; Savva, A.; Berry, D.; Grimshaw, A.; Horn, B.; Maciel,
F.; Siebenlist, F.; Subramaniam, R.; Treadwell, J.; Von Reich, J.(2006): The
Open Grid Services Architecture, Version 1.5.
http://www.ggf.org/documents/GFD.80.pdf (accessed on 19Oct 2006).
Hardin, G. (1968): The Tragedy of the Commons. In Science Magazine, Vol. 162
No. 3859, pp. 1243-1248.
Irwin, D.; Chase, J.; Grit, L.; Yumerefendi, A. (2005): Self-recharging virtual
currency. In Proceeding of the 2005 ACM SIGCOMM, pp. 93-98.
Liebau, N.; Darlagiannis, V.; Mauthe, A.; Steinmetz, R. (2005): Token-based Accounting
for P2P-Systems. In: Proceeding of Kommunikation in Verteilten Systemen KiVS,
pp.16-28.
Moreton, T.; Twigg, A. (2003): Trading in Trust, Tokens and Stamps. In
Proceedings of the first workshop on economics of Peer-to-Peer Systems,
Berkeley (USA).
Ostrom, E. (1990): Governing the Commons: The Evolution of Institutions for
Collective Action. Cambridge University Press.
Papaioannou, T.G.; Stamoulis, G. (2005): An incentives' mechanism promoting truthful
feedback. In peer-to-peer systems, Cluster Computing and the Grid, CCGrid 2005.
IEEE International Symposium on, Vol.1, pp. 275-283.
Vishnumurthy, V.; Chandrakumar, S.; Sirer, E. (2003): KARMA: A Secure
Economic Framework for Peer-to-Peer Resource Sharing. In 1st Workshop
on Economics of Peer-to-Peer Systems.
Xiong, L.; Liu, L. (2004): PeerTrust: supporting reputation-based trust for peer-to-peer
electronic communities. In IEEE Transactions on Knowledge and Data
Engineering, Vol. 16 No. 7, pp. 843-857.
Yu, B.; Singh, M.(2003): Incentive Mechanisms for Agent-Based Peer-to-Peer Systems.
In Agents and Peer-to-Peer Computing, Second International Workshop, pp. 77-
88.
234
11

WI – Schwerpunktaufsatz
Distributed Ascending Proxy Auction
A Cryptographic Approach
The Authors
Daniel Rolli
Michael Conrad
Dirk Neumann
Christoph Sorge
Daniel Rolli
Dirk Neumann
Institut für Informationswirtschaft
und -management
Universität Karlsruhe (TH)
Englerstr. 14
76131 Karlsruhe
{rolli | neumann}@iw.uka.de
Michael Conrad
Christoph Sorge
Institut für Telematik
Universität Karlsruhe (TH)
Zirkel 2
76131 Karlsruhe
{conrad | sorge}@tm.uka.de
The Internet revolution has given rise to a
widespread use of online auctions as a price
discovery mechanism for selling different
items (e.g. eBay). When referring to auctions,
one traditionally has central or brokered
auctions in mind with a designated
entity that collects all the bids, determines
the resulting allocation, and corresponding
prices. Central auction platforms are typically
of high reliability, where reliability refers
to (1) correctness with respect to compliance
to the auction rules, (2) continuous
availability, and (3) security to a certain extent.
Since the designated broker offers
auctions to several groups repeatedly and
enforces agreements, it can build up reputation
that eventually may form trust in the
platform.
The central approach, however, also has
several disadvantages. Since a designated
central entity conducts all brokering services,
it is vulnerable against Denial-of-Service
(DoS) attacks and may suffer under
overload. Hence, scalability may be problematic,
as the broker can only handle a
certain number of participants at a time.
Furthermore, the central broker typically
charges fees from the participants in order
to recover the investments in infrastructure
and its maintenance.
Peer-to-Peer (P2P) technologies have
emerged that remove or at least alleviate
these technical disadvantages of client-server
architectures. In general, P2P networks
distinguish themselves by providing scalable
and robust applications. Combining
P2P technologies with auctions is expected
to bring forth auction systems that are scalable
and robust on a technical level and retain
the excellent allocation properties of
auctions on an economic level.
In so-called decentralized auctions, peers
can be brokers and bidders at a time. When
redundant broker peers supersede the central
broker, decentralized auctions are less
vulnerable to DoS attacks. Scalability properties
can be widely improved in comparison
to the central approach, since every
bidder contributes his share of additional
infrastructure. Since peers collectively
make for the total infrastructure, holding a
decentralized auction can be free of (monetary)
charge. Regarding the massive infrastructure
of major auction providers like
eBay, scalability may be less of an issue.
But a need for decentralized auctions is
further driven by the demand of P2P based
user communities, who strive for disintermediation.
Executive Summary
1 Introduction
The paper presents a distributed ascending proxy auction protocol. The protocol substitutes
trust in a central (auctioneer) entity by ensuring correct conduct and privacy preservation. It
finally enables the robust application of decentralized auctions in open environments like
Peer-to-Peer systems.
&
&
&
Initial registration of bidders is not required; they can enter and bid spontaneously.
Winner and second-price determination are robust to bidder and auctioneer failure or
disconnection, as well as malicious peers.
The privacy of the winning bidder’s highest bid is preserved – with a minor exception
theoretically possible.
Keywords: Internet Economics, Distributed Auctions, Privacy, Trust, Security Protocols, Peerto-Peer
Systems
235
WIRTSCHAFTSINFORMATIK 48 (2006) 1, S. 7–15

8 Daniel Rolli, Michael Conrad, Dirk Neumann, Christoph Sorge
In recent times, the benefits of decentralized
auction platforms have been recognized
in research [McAf00; HaSt05a;
HaSt05b] as an alternative to central auction
servers. Nonetheless, there are several
obstacles to overcome before decentralized
auctions will become an important pillar of
e-business solutions: Due to the lack of a
centralized platform, trust becomes an issue.
Since any peer can set up an auction,
there is no persistent instance, which assures
reliability of the auction process. To
introduce a threshold level of trust into decentralized
auctions, appropriate security
mechanisms become indispensable.
In the following, we present the distributed
ascending proxy auction as an example
for an iterative decentralized auction. It
is coupled with proxy agents that increase
the price on behalf of the corresponding
participants to beat the preceding highest
price, as long as the bid that is given to the
proxy agent is at least an increment higher
than the current price. This allows bidders
to have their bid value revealed step by step
– and no earlier than necessary – without
the need of constantly being online. This
auction corresponds to the original eBay
auction, while the latter is centralized. In
the present paper, this decentralized auction
is extended to a secure, decentralized
mechanism, which can be easily conducted
by peers over the internet.
The remainder of this paper is structured
as follows. The requirements of an ascending
proxy auction are pointed out in section
2. In section 3, related approaches in
literature are reviewed with respect to these
requirements. Subsequently, the protocol
for the auction is presented in section 4.
Section 5 evaluates the protocol in terms of
its security and robustness. Lastly, in section
6, the paper concludes with a summary
and future work.
2 Requirements
For a secure asynchronous ascending
proxy auction, the following main requirements
have been identified. Since the auction
is envisioned to be decentralized, its
protocol is no longer managed by a central
auctioneer but by multiple collaborating
auctioneer nodes. Principally, it is possible
that those auctioneer nodes are recruited
from the bidders.
In the context of ascending proxy auctions
there arise four main requirements
pertaining to the role of those multiple auctioneers
(A1–A4).
&
&
&
&
Distributed Result Determination (A1):
In absence of a central entity, which can
keep track of the bidding process, proxy
bidding is difficult to realize. Since we
want to avoid another trusted third
party, we rely on several auctioneers and
require that they jointly determine the
price (which equals in our case the second
highest bid) and the winner.
Preservation of Secret Highest Bid (A2):
The idea of a proxy auction is that bidders
may submit their valuation truthfully.
A bidder can be reluctant to do so,
as he wants to keep his true valuation secret
and may not trust the auction process.
Thus, ensuring the preservation of
the winning bidder’s highest bid within
the distributed protocol is an imperative.
Resistance to Bidder Exclusion (A3): The
ascending proxy auction is intended as
an open auction allowing all peers in the
network access. Neither a single auctioneer
nor “smaller” coalitions of auctioneers
are able to exclude bidders by
illegitimately dropping bids.
Robustness to Paralysis Attacks (A4):
For a smooth, error-free conduct of the
auction, it is required that neither a single
auctioneer nor “smaller” coalitions
of auctioneers are able to block the protocol.
From the auction format of an ascending
proxy auction five additional requirements
arise referring to the bidding process
(B1–B5).
&
&
&
&
Spontaneous Entry and Exit (B1): To account
for the dynamic environment of
the Internet, any bidder must be able to
submit bids without prior registration at
any time. This means, a bidder can spontaneously
join and leave the auction.
Asynchronous Bid Submission (B2): Bidders
can independently submit bids at
any time during the bidding phase. Only
when submitting a bid, a bidder must be
online. Under no circumstances are all
bidders required to be online at a time.
This requirement focuses on the monitoring
costs, which should be as low as
possible.
Iterative Nature (B3): Bids can be submitted
iteratively with the standing
highest price revealed as information
feedback.
Single Individual Independent Key-Pair
(B4): For participation in a distributed
ascending proxy auction, each bidder
needs only a single individual key-pair
for encryption, which are not dependent
on any other peer’s key-pair. These keys
can be (re-)used for an arbitrary number
236
of auctions. Generating keys for any
new auction creates high transaction
costs. This requirement is closely related
to B1 and emphasized, because many
existing secure auctions are particularly
costly in this respect.
& Non-repudiation (B5): As with any
other auction, the winning bidder may
not deny the submission of the winning
bid.
An auction protocol that achieves those requirements
is secure in a sense that there is
no single point of failure, assured bid submission,
protected bid acceptance, and a
certain degree of robustness against paralysis
attacks.
3 Related Work
In recent years, several cryptographic protocols
have been proposed mainly for secure
sealed-bid auctions. Among the first
was the auction protocol developed by
Franklin and Reiter [FrRe96] that was later
criticized due to its inadequate privacy preservation
after the auction is concluded.
Privacy preservation means that neither
bids nor bidder identities are revealed (A2).
Since privacy preservation is a crucial requirement
for sealed-bid auctions, and in
particular for second-price sealed bid auctions
– the so-called Vickrey auction – various
new protocols have been suggested to
remedy this shortcoming [Bran02].
Among those newly suggested approaches
only few can cope with auction
formats that both comply with the privacy
preservation requirement and embody
other pricing rules than pay-your-bid. The
auction protocols that do satisfy privacy
preservation and, nonetheless, support
more sophisticated pricing schemes, can be
classified into two classes.
& Distributed computation among multiple
auctioneers: Typically, the agents submit
shares of the bids to different auctioneers.
The auctioneers jointly compute
the auction price by using the techniques
of secure multiparty function evaluation.
The gist of those techniques is
that the auctioneers can compute the
auction price without knowing any bid
&
[KHAN00; Kiku01].
Partially trusted third party: The second
class of protocols requires the introduction
of a third party, which controls the
auctioneer. Crucial for those protocols
is that the third party needs not be fully
trusted but can draw on a weaker form
of trust [Cach99; BaSt01].
WIRTSCHAFTSINFORMATIK 48 (2006) 1, S. 7–15

Distributed Ascending Proxy Auction 9
Both classes inherently impose several
challenges and problems. In fact, the first
class requires a threshold number of obedient
auctioneers, i.e. auctioneers that correctly
follow the protocol. Otherwise, collusion
among malicious auctioneers might
compromise the auction process. The second
class of protocols requires some sort
of third party that is trustworthy to a certain
degree. Brandt introduced a third way
to secure sealed-bid auctions. In fact, his algorithm
requires all bidders in conjunction
with the seller to resolve the auction. Because
this approach preserves privacy completely,
as only the winner and the corresponding
bid is resolved, it is deemed
superior to the other approaches for supporting
sealed-bid auctions [Bran03].
Iterative auctions (B3), however, have
hardly been considered explicitly. In traditional
iterative auction formats, the privacy
requirements are not as demanding as in
sealed formats. As bids are by nature open,
there is no further need to secure them.
For preventing malicious behavior (A3 &
A4), it is often recommended to use the secure
sealed-bid auction iteratively. The lessons
learned from sealed-bid auctions
would suggest to refer to the approach of
Brandt [Bran03] for iterative auctions, as it
allows anonymity and correctness (B5)
along the bidding process.
This recommendation is, however, premature
and even faulty. Staging round-wise
bidder-resolved auctions does not support
a continuous bidding process (B2). This
shortcoming could be excused, but, even
worse, Brandt’s algorithm requires all participating
bidders staying in the auction
until it is resolved (B1
B4). Once a bidder leaves the auction during
one round, the auction cannot be resolved.
For a dynamic network such as the
Internet or any P2P network this requirement
is way too strong. Another argument
that challenges Brandt’s approach is concerned
with the active involvement of the
seller in the bidding process. Since the seller
receives the auction outcome first and,
subsequently, determines who has submitted
the winning bid, he can leave the
auction at will, before it is settled. In a oneshot
auction with no outside option, this is
rather unproblematic, as the seller can always
impose a reservation price (e.g., the
seller can also submit a bid specifying the
reservation price) and all bids above this
price contribute to positive utility. In reality,
this explanation is not adequate, as it
excludes alternative ways to sell the item.
For instance, the seller could use the auction
to get an idea about the bidders’ valuations,
which can be exploited in subsequent
re-negotiations with parts of the bidders.
From a strategic point of view, bidders will
presumably take this behavior into account
and may bid less aggressively.
Thus, Brandt’s approach cannot be used
for supporting iterative auctions in more
realistic settings – despite its appealing theoretical
conduct. Other existing alternative
cryptographic approaches hold even severer
problems. Existing protocols using
multiple auctioneers typically require all
auctioneers to jointly determine the winning
bid. As such, all auctioneers are
employed at a time to resolve the auction,
violating spontaneous exit and entry (B1)
and asynchronous bidding (B2) [Kiku01;
AbSu02]. Protocols using partially trusted
third parties can principally represent continuous
bidding processes with spontaneous
exit and entry of bidders (as long as
the third party is online). However, third
party protocols typically have problems
with non-repudiation and privacy preservation
[BaSt01; LiAN02].
Table 1 wraps up the current approaches
proposed in literature. Note that most protocols
are originally intended for sealed
auctions. Nonetheless, the authors claim
that these sealed protocols can be applied
for iterative auctions as well. This paper examines
the appropriateness of those protocols
for iterative auctions only, not questioning
their merits for sealed auctions.
As Table 1 illustrates, in most protocols,
possible bids (prices) have to be determined
before the auction starts. The computation
complexity usually depends on
the number k of possible bids. With Asynchrony,
we refer to the possibility of asynchronous
bid submission; additionally, a
protocol with more than one auctioneer
should not require all auctioneers to be online
at the same time. Verifiability means
that it is possible to verify the correctness
of the auction outcome.
This paper attempts to provide a reasonable
protocol for supporting secure iterative
auctions. Drawing on asymmetric encryption
the protocol can satisfy correctness
in a way that winning bidder and
corresponding price are accurately and
transparently determined. Accordingly,
bids can never be repudiated by any agent
and the possibility for successful collusion
among malicious auctioneers is extremely
low. Furthermore, bidders can join and
leave the auction spontaneously – there is
no need for connecting to the auction other
than submitting a bid.
4 Protocol Description
For the distributed ascending proxy auction,
in essence we use multiple auctioneers
in different groups to prevent at any time
one single auctioneer from obtaining complete
control over the auction process.
As in most iterative auctions, bidders
send their bids at arbitrary points in time
to the auction mechanism. Any bidder can
address an arbitrary auctioneer node in the
collection of auctioneers. Since one of our
main goals is the protection of the highest
bid amount, we produce for any bid a bid
chain BC, i.e. a sequence of strictly monotonously
increasing bid steps B i where i
denotes the order in the bid chain. The idea
of the protocol is that only one bid step at
a time can be revealed in the order the bid-
Table 1
Related Protocols (adapted from [Bran05])
Protocol Price Auctioneers Number of rounds Overall Computation Verifiability Non-Repudiation Asynchrony Spontaneous
entry and exit
[Bran03] ðM þ 1Þ st z þ 1 Oð1Þ OðkÞ Yes Yes No No
[LiAN02] ðM þ 1Þ st 2 Oð1Þ OðkÞ Yes No No Yes
[AbSu02] ðM þ 1Þ st 2 Oðlog kÞ OðkÞ Yes Yes No No
[BaSt01] 1 st 1 Oð1Þ Oðzðlog kÞ z 1 Þ No No No Yes
[Kiku01] ðM þ 1Þ st M > k Oðlog kÞ Oðz log kÞ Yes Yes No No
WIRTSCHAFTSINFORMATIK 48 (2006) 1, S. 7–15
237

10 Daniel Rolli, Michael Conrad, Dirk Neumann, Christoph Sorge
der intends. To ensure this sequential revelation,
the bidder encrypts the steps of the
bid chain in cascades. Avoiding that one
auctioneer reveals subsequent bid steps
every step is encrypted with the private
keys of different auctioneer groups. We use
two public keys of different auctioneer
groups to encrypt one bid step in order to
increase privacy preservation. The additional
auctioneer group of the second key
also needs to be compromised for illegitimately
revealing one bid step. The number
of involved groups per bid step can be arbitrarily
adjusted starting from one. Whenever
an auctioneer group receives a bid
chain during the auction process, one or
several randomly chosen members of the
group decrypt the next encrypted bid step,
if their private group key is applicable and
the preceding bid step in the chain is not
higher than the standing highest bid
(SHB). In case their key is not applicable,
the group passes the bid chain on to the
corresponding auctioneer group. A newly
decrypted bid step higher than the SHB is
broadcast to all other auctioneer groups as
the new SHB. Each auctioneer group holding
a partially decrypted bid chain only
passes it on to the next group in the chain,
if the bid step they decrypted does not beat
the SHB. As a consequence, only the bid
chain containing the SHB is typically kept
partially encrypted and held by one auctioneer
group, while all other bid chains
are fully decrypted and discarded.
At the end of the auction, the winner
with the highest bid chain pays the price of
the highest bid step in the second-highest
and fully revealed bid chain.
The description of the crucial details in
the auction protocol is structured along
three distinct parts: The initial setup describes
the selection of the auctioneer
groups and their order of appearance in the
decryption process, the bid placing specifies
the way bid chains are structured and
submitted, and finally the bid processing
describes how the auctioneers deal with
bid chains for outcome determination.
4.1 Initial Auction Setup
At the outset of the auction, seller S has to
create a document D describing the auction.
This document includes the description
of the items for sale and the fully specified
auction procedure with all relevant
parameters such as starting and ending rule
or minimum price. Document D is used to
determine the auctioneers and their groups
that share the same public/private key pair.
To ensure authenticity of document D,
seller S has to digitally sign D using the private
key of his public/private key pair.
Starting from a desirably large number
A total of auctioneers responsible for the
protocol, the number of groups G number
and their group size G size can be determined.
The number of groups should grow
faster, with an increasing A total , than the
size of each group. As a design choice for
our particular auction protocol, we use the
following concave function for G size and
the convex counterpart for G number :
G size ¼ log x ðA p
totalÞþ ffiffiffiffiffiffiffiffiffiffi
A total
2
G number ¼
A
total
G size
In the formula for G size , we use the average
of the base x logarithm and the square root
of the total number of auctioneers A total to
determine the size of any auctioneer group
and take the integer part of the result. x is
common knowledge throughout the auction
and in our examples set to 2. For the
number of groups G number we divide the
total number of auctioneers A total by the
group size G size and take the integer part.
This approach provides many auctioneer
groups relatively small in size. The particular
choice of this group generating function
is discussed in section 5.1.
After having determined G size and
G number the auctioneers themselves and
their membership in groups have to be
identified. This identification must be deterministic
and traceable, but hard to anticipate.
In particular, the seller himself
must not be capable of influencing the selection
process. Otherwise, he could select
collaborating or malicious nodes in the set
of auctioneers.
Hence, we employ the description document
D for the selection. One approach is
the application of a cryptographic hash
function. Such a function, which can be efficiently
computed, maps arbitrary input
onto a result of fixed length. No inverse
can be efficiently calculated, nor can two
different inputs leading to the same result
be efficiently determined. Typical hash
functions generate a 128-bit or 160-bit output
value.
Starting with the hash value of D, a chain
of hash values can be calculated for the
auctioneers.
H r ðDÞ ¼HðHðHð...HðDÞÞÞÞ
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
r þ 1 times
Given an addressing mechanism which is
capable of mapping such hash values to
238
distinct nodes, a chain of auctioneer nodes
can be derived from the chain of hash values.
Such mechanisms are commonly
available in structured P2P networks like
CAN [RFHK01] or Chord [SMKK01].
The configuration of auctioneer addresses
in groups is computed by using the following
equation, where A ij is the j-th auctioneer
in the i-th group.
Address ðA ij Þ¼H i Gsize þ jðDÞ
where 0 i < G number
and 0 j < G size
Having defined groups, the members of
each group have to obtain a common
group public/private key pair. For simplicity,
we have one auctioneer per group
generate a public/private key pair and distribute
it to all other auctioneers in the
group using the individual public key of
each group member.
After a successful conduct of this auction
setup, which includes determination of the
number G number and size G size of groups,
selection of group members and groupbased
key generation, we receive the following
entities:
G i ¼ group i of auctioneers,
where 0 i < G number
K i ¼ public key of group G i
A ij ¼ auctioneer j of group i
with public key K i ;
where 0 j < G size
4.2 Bid Placement
Having set up the auction, potential buyers
can join by placing bid chains. According
to requirement A2 (refer to section 2), a
bidder wants to hide his highest bid
amount as long as possible, which may reflect
his true valuation.
If a bidder has a valuation v, he typically
offers successively increasing bids in the
form of several bid chains up to v. Our
procedure makes it possible, but not necessary,
for the bidder to observe the auction
during its whole runtime in order to receive
information feedback.
The approach at hand gives the bidder
the possibility to place with any bid chain a
series of bid steps and ensures that these
steps are only revealed if necessary for
moving closer to or surpassing competing
bidders. When a bidder joins an auction,
he can retrace G number and G size by using
the same algorithm as the seller, which is
described in the previous section.
WIRTSCHAFTSINFORMATIK 48 (2006) 1, S. 7–15

Distributed Ascending Proxy Auction 11
According to his bidding strategy, a bidder
typically submits a bid chain BC with a
price p ¼ v, where p is the maximum value
that is submitted with BC. For breaking
down BC, he formulates n strictly monotonously
increasing bid steps starting from
a value p start < p (i.e. the SHB of the auction)
ranging to p with certain increments.
Each bid step in the chain is digitally
signed by the bidder to ensure non-repudiation.
The bidder can reproduce A ij , the auctioneer
addresses with group-association,
applying the same calculations as the seller.
From this he arbitrarily chooses a hopping
sequence (HS) of auctioneer groups and receives
the respective public keys. Each single
bid step is then encrypted using two
different keys of the HS.
HS i ¼ j
where
0 j < G number ^ HS 2k 6¼ HS 2k þ 1 j k 2 N 0
The idea is that the first bid step, which is
decrypted, contains all subsequent encrypted
bid steps in cascades. So, the construction
of any bid chain starts from the
highest to the lowest bid step in reverse order
to the decryption process.
For asymmetric cryptographic operations
described in the following paragraphs,
S A ðxÞ denotes x, signed by A; K i is
the public key of the i-th auctioneer group.
The bidder starts by signing the highest
bid step plus its index number, which results
in S L ðB n , nÞ, and encrypts this with
the public key K HS2n of the last auctioneer
group in the HS. As shown in the following
formula (1), this encrypted statement is
denoted as I 1 . With the second last public
key K HS2n 1
in the HS the bidder generates
I 2 – formula (2) – by encrypting the following:
& The index HS 2n of the auctioneer group,
which is responsible for decrypting I 1
With this information the subsequent
auctioneer group in the hopping sequence
can be addressed (c.f. 4.1) in the
decryption process.
& The encrypted statement I 1 for achieving
encryption in cascades.
& The signed hash value S L ðHðB n Þ, nÞ of
the previously encrypted bid step and its
index number. As described in 4.3 this
hash value is used by auctioneer group
HS 2n to verify the correctness of their
decryption operation.
With the third public key K HS2n 2
of the HS,
the bidder incorporates the index of the responsible
auctioneer group HS 2n 1 , the encrypted
statement I 2 and the signed bid
step plus index number S L ðB n 1 , n 1Þ.
The encryption analogously continues
until all bid steps are encrypted. Those
statements with uneven indices contain a
signed bid step with its index number
S L ðB i , iÞ, whereas the ones with an even index
contain a signed hash value of the bid
step and its index number S L ðHðB i Þ, iÞ.
To the very last statement, which is
firstly generated to the abovementioned
rule for even indices, the bidder adds HS 1 ,
which determines the auctioneer group
capable of decrypting the statement, and
signs this.
The following formulas illustrate the encryption
of the bid chain from the highest
to the lowest bid step. In the formulas, HS i
stands for the i-th key of the hopping sequence,
E i denotes the encryption with key
K i , S L the signing by bidder L, B j the j-th
bid step of the bid chain, HðB j Þ the hash
value of bid step B j , and I k for the k-th encrypted
statement of the bid chain BC,
where k ¼ 2n i þ 1.
E HS2n ðS L ðB n , nÞÞ ¼ I 1 ð1Þ
E HS2n 1
ðHS 2n , I 1 , S L ðHðB n Þ, nÞÞ ¼ I 2 ð2Þ
E HS2n 2
ðHS 2n 1 , I 2 , S L ðB n 1 , n 1ÞÞ ¼ I 3 ð3Þ
E HS2n 3
ðHS 2n 2 , I 3 , S L ðHðB n 1 Þ, n 1ÞÞ ¼ I 4 ð4Þ
...:
E HS4 ðHS 5 , I 2n 4 , S L ðB 2 ,2ÞÞ ¼ I 2n 3 ð5Þ
E HS3 ðHS 4 , I 2n 3 , S L ðHðB 2 Þ,2ÞÞ ¼ I 2n 2 ð6Þ
E HS2 ðHS 3 , I 2n 2 , S L ðB 1 ,1ÞÞ ¼ I 2n 1 ð7Þ
S L ðHS 1 , E HS1 ðHS 2 , I 2n 1 , S L ðHðB 1 Þ,1ÞÞÞ ¼ I 2n ð8Þ
Having constructed the encrypted bid chain,
the bidder transmits I 2n to any auctioneer.
If each bid chain consists of a fixed number
of bid steps and the value p is reached
after a random number of steps, which is
significantly larger than 1, other bidders or
auctioneers are not able to predict the value
of p from the starting value and/or other
parameters like the increments or number
of bid steps. Accordingly, the number of
bid steps has to be chosen sufficiently large.
4.3 Bid Processing
The decision tree in Figure 1 illustrates the
exact steps each auctioneer group has to
follow according to our ascending proxy
auction protocol.
So, according to Figure 1, an auctioneer
receiving a newly submitted bid chain BC
follows the decision tree in the following
239
way: he firstly sends BC in its initial form
of I 2n to HS 1 – even if it happens that he
belongs to this group. The auctioneer
group HS 1 then checks, if HS 1 in I 2n addresses
them. If it does address them and
there is no decrypted lower bid step, a subset
(randomly chosen) of all group members
of HS 1 directly starts the first decryption
step using key K HS1 ; otherwise they
would pass BC on to the actual HS 1 . After
the first successful decryption step, they
check, if the revealed S L ðHðB 1 Þ,1Þ contains
the index number 1. If it does not,
they discard the bid chain. If it does, however,
contain the right index number 1, the
auctioneer group sends a bid confirmation
for BC back to the bidder L. The group
also conveys I 2n 1 ¼ E HS2 ð...Þ and
S L ðHðB 1 Þ,1Þ to the auctioneer group with
index HS 2 to continue decryption. The
members of HS 2 check the index number 1
in S L ðHðB 1 Þ,1Þ that they receive and presume
that no bid step of BC has been decrypted,
yet. Hence, a subset (randomly
chosen) of all group members of HS 2 decrypts
I 2n 1 to HS 3 , I 2n 2 , and S L ðB 1 ,1Þ.
The decrypting members then propagate
the signed bid step S L ðB 1 ,1Þ to all other
group members. This enables all group
members firstly to check if the received index
number 1 in S L ðHðB 1 Þ,1Þ equals the
index number in the newly revealed
S L ðB 1 ,1Þ. If not, there is a decryption or
integrity error. Secondly, they can verify
valid decryption by hashing B 1 and comparing
the result to HðB 1 Þ of S L ðHðB 1 Þ,1Þ.
If those do not match, another auctioneer
subset of the respective group has to repeat
this whole process until validity is established
or an integrity error must be presumed.
Having successfully checked validity of
the decrypted bid step B 1 , and B 1 beating
the SHB, the auctioneer group broadcasts
S L ðB 1 ,1Þ that includes B 1 as the SHB to all
auctioneers.
WIRTSCHAFTSINFORMATIK 48 (2006) 1, S. 7–15

12 Daniel Rolli, Michael Conrad, Dirk Neumann, Christoph Sorge
5.2 Security Analysis
a
YES
b) (no decrypted bid step B i-1 )
b
or (B i-1 < SHB)?
YES NO
2) decrypt bid step B i 2
0
1) pass on bid chain BC 1
e) B i >B i-1 ?
f) newly decrypted B i >SHB?
3) publish B i as new SHB
Figure 1
Decision Tree
If bid step B 1 does not beat the leading
bid, the group transfers I 2n 2 , and S L ðB 1 ,1Þ
to the next auctioneer group HS 3 in the
hopping sequence. After successfully validating
the index numbers and B 1 as well as
processing of the statement I 2n 2 , the next
auctioneer group HS 3 passes on the newly
decrypted I 2n 3 and S L ðHðB 2 Þ,2Þ plus
S L ðB 1 ,1Þ to the next auctioneer group HS 4 .
After validating B 1 < SHB, HS 4 decrypts
I 2n 3 and firstly checks, if the index
number of their newly decrypted S L ðB 2 ,2Þ
is exactly one higher than the index number
they received in S L ðB 1 ,1Þ from HS 2 via
HS 3 . If so, they check if B 2 is higher than
B 1 . If not, they discard BC. IfB 2 > B 1 applies,
they also check if B 2 is higher than
SHB and, in case this applies, they publish
S L ðB 2 ,2Þ as the new SHB. If B 2 > SHB
does not apply, they pass on I 2n 4 and
S L ðB 2 ,2Þ to HS 5 .
This process is continued until the auction
terminates. Obviously, all bid chains –
except for the winning bidder’s – end up
fully decrypted.
5 Evaluation
YES
YES
YES
This section analyzes the protocol taking
into account the requirements presented in
section 2. After a discussion of the group
generating functions, the security and
threat analysis is performed, and the complexity
analysis closes the evaluation.
5.1 Group Generating Function
In section 4.1 we presented a pair of group
generating functions to define the number
and size of auctioneer groups. The proposed
auction protocol does not depend on
a special function, as long as the applied
3
1
YES
f
e
NO
1
c
NO
d
NO
4
NO
0
NO
1
a) bid chain intended for this group?
1) pass on bid chain BC 1
0) throw integrity error
c) more encryption on B i ?
d) i of B i =(jofreceivedB j )+1?
0) throw integrity error
4) discard BC 1
1) pass on bid chain BC 1
group generating function is common
knowledge. The intuition, however, behind
the chosen pair of functions is as follows:
The size of auctioneer groups increases robustness
and security as bigger groups can
tolerate more (in absolute terms) disabled
or even corrupted members. The number of
auctioneer groups, however, is associated
with a higher level of security, as more
available auctioneer groups entail a higher
number of encryption keys leaving any bidder
with more options to generate a hopping
sequence. Hence, we prefer the group
number to increase faster than the size of
groups. We regard log-based and squareroot-based
functions for G size as stylized
functions to motivate the trade-off between
robustness and communication overhead.
In essence, the log as well as the square root
both have certain disadvantages.
– Using the log-based function entails a rapid
growth of G number but a slow growth
of G size . For large auction setups with
many auctioneers, many different groups
of very small size are generated. This can
be exploited by an attacker, who may relatively
easily overtake a small group
empowering him to partially block the
auction.
– In contrast, the square-root-based function
generates almost identical results for
G number and G size . Compared to the logbased
function, it attains a slower growth
of group numbers, but faster growth of
group sizes. Hence, given a large number
of auctioneers, groups with more group
members are generated. As a side effect
this creates higher communication and
management costs within the groups,
which may become a bottleneck.
The proposed ad-hoc function combines
both functions to a compromise, where the
number of groups grows modestly faster
than the group size.
240
In the security analysis, we refer to security
risks within the auction protocol only.
Those risks related to standard cryptographic
algorithms like SHA1 or RSA as
well as common threats (e.g. data flooding
or virus infection) are beyond our scope.
Two major risks associated with the distributed
ascending proxy auction are
1. premature revelation of the highest bid
step(s) and
2. exclusion of a targeted bidder from the
auction.
For each threat we provide an attack tree
[Schn99], exhibiting the crucial attack possibilities
and giving a short explanation of
how such attacks are countered in the protocol.
Figure 3 shows the attack tree of the premature
revelation attack (refer to requirement
A2).
As Figure 2 shows, there are two approaches
to illegitimately revealing all bid
steps up to the highest one. Firstly, the attacker
infiltrates the auctioneer groups to
reveal the highest bid step. To do so, he
firstly needs to infiltrate the auctioneer
group holding the bid chain and forward it
to the next auctioneer group in the hopping
sequence. The attacker then has to infiltrate
this auctioneer group to decrypt the next
bid step. If this bid step is not the highest in
the bid chain, the attacker has to repeat this
process until the highest bid step is revealed.
Secondly, the attacker attempts to retrieve
the remaining bid chain from the
auctioneer group currently holding it.
Having succeeded, the attacker tries to decrypt
the remaining bid steps by using illegitimately
obtained private keys of the involved
auctioneer groups.
Accordingly, both attacks require the
identification of the auctioneer group holding
the bid chain and all subsequent groups
in the hopping sequence. The addresses of
the involved auctioneer groups are encoded
within the bid chain and can only sequentially
be obtained. The attacker can do
nothing but guess the current auctioneer
group with the bid chain and learn each
following auctioneer group stepwise.
Note that one possibility of revealing the
overall highest bid step is immanent to the
proposed protocol. If the highest bid step
of the second-highest bid chain is higher
than or equal to the second-highest bid
step of the highest bid chain, the overall
highest bid step is revealed at least to one
auctioneer group.
The second attack tree addresses excluding
potential bidders from an auction.
WIRTSCHAFTSINFORMATIK 48 (2006) 1, S. 7–15

Distributed Ascending Proxy Auction 13
There are three independent possible attacks,
which are shown in Figure 3.
The first attack possibility is to suppress
the submission of the bid chain. Hence, the
attacker can lock out the bidder by flooding
its peer and thus disconnecting it from
the network. Such an attack can be easily
detected.
The second possible attack is forging the
auction description. In this case, the attacker
could provide wrong information about
the auction, so that the bidder cannot join
the auction and place a valid bid chain.
However, this attack can easily be detected
by any bidder. At first the signature of the
auction description document generated by
the seller becomes invalid. Furthermore the
bidder does not retrieve a positive bid confirmation
after placing an invalid bid.
The third, more complex attack possibility
is the dedicated prevention of processing
bid steps of a specific bidder. To reach
this, the attacker has to overtake at least
one auctioneer group in every HS of the
bidder as whole. This attack requires the
knowledge of any HS of a bidder, and the
bidder can even react by adapting the HS
of a repeated bid chain. Accordingly, the
attacker must either employ a lot of effort
for maliciously obtaining and stepwise decrypting
all bid chains of one bidder, or
simply guess to take over several auctioneer
groups randomly, hoping to compromise
the right one by chance.
Most of the attacks base on the presumption
that the attacker succeeds in guessing
the HS of a bidder. Figure 4 shows the
probability of an attacker unfolding one
next bid step of the standing highest bid
chain depending on the number of infiltrated
auctioneer groups in relation to the
total group number. To reveal the next bid
step an attacker has to compromise at minimum
three auctioneers in different groups.
The first group is needed to maliciously
transfer the encrypted bid step to the next
group, without a higher SHB present; the
other two for actual decryption, since we
assume that each bid step is encrypted with
two different keys of the HS.
We analyze auctions starting from 16 up
to 1024 different auctioneers. Using our
group generating functions described in
section 4.1 results in 4 auctioneer groups
for 16 auctioneers up to 48 auctioneer
groups for 1024 auctioneers. We assume
the attacker is able to compromise between
3 and 15 auctioneers.
Figure 4 and its findings imply that the
probability of unfolding the next bid step
and the exclusion of a certain bidder by an
attacker decrease significantly with an ascending
number of auctioneer groups.
Even in case the attacker is able to compromise
almost one third (15 of 48) of all auctioneer
groups (from a total of 1024 auctioneers)
he reaches a probability lower
than 3 percent of unfolding even the next
bid step of a specific bid chain. All values
are calculated using a hypergeometric distribution:
M N M
x n x
PðX ¼ xÞ ¼
Forward bid chain
to responsible
auctioneer group
Infiltrate auctioneer
group holding
remaining bid chain
Know hopping
sequence of
dedicated bidder
Figure 2
Overtake first
auctioneer group
in bid chain
Know hopping
sequence of
dedicated bidder
Figure 3
Reveal remaining
bidding step(s) by
malicious auctioneers
N
n
AND
Decrypt next bidding
step of bid chain
Infiltrate auctioneer
group decrypting
next bidding step
Reveal highest
bidding step(s)
Attack Tree – Highest Bid Step Revelation
Prevent submission of
bid chain
Lock out bidder
from network
Attack Tree – Bidder Exclusion
241
Exclude a
dedicated bidder
Forge auction
description
Infiltrate auctioneer
group holding
remaining bid chain
Know hopping
sequence of
dedicated bidder
Infiltrate next
auctioneer group(s)
in bid chain
Know hopping
sequence of
dedicated bidder
Reveal remaining
bidding step(s) by
attacker
AND
Prevent processing of
next bidding step
Variable N stands for the total number of
auctioneer groups; n is the number of compromised
groups. The variables M and x
by assumption are 3, because an attacker
needs all of the three involved auctioneer
groups to reveal the next bidding step. The
value PðX ¼ xÞ gives the probability of unfolding
the next bid step of a specific bid
chain.
5.3 Complexity Analysis
Decrypt remaining
bidding step(s)
Know private keys
of auctioneer groups
used in bid chain
Know hopping
sequence of
dedicated bidder
Overtake current
auctioneer group
in bid chain
Know hopping
sequence of
dedicated bidder
This section covers a basic complexity analysis
of the presented auction protocol concerning
calculation and communication
complexity.
For the analysis, let z be the number of
bidders, A total the number of auctioneers, n
the (average) number of bid steps with new
information in a bid chain, e the number of
groups whose keys are used to encrypt each
bid step (two in our example), and b the
average number of bid chains per bidder.
Each bidder has to generate b n bid
steps. As each bid step is encrypted e times,
b n e encrypt operations have to be performed
by a bidder, leading to a (computation)
complexity in Oðb n eÞ for a bid-
WIRTSCHAFTSINFORMATIK 48 (2006) 1, S. 7–15

14 Daniel Rolli, Michael Conrad, Dirk Neumann, Christoph Sorge
Revealing Probability
1
0.8
0.6
0.4
0.2
Figure 4
Decrypt Operations
Probability of revealing the next Bidding Step of the Winner’s Bid Chain
P (3 compromised auctioneers) (log+sqrt)
P (6 compromised auctioneers) (log+sqrt)
P (9 compromised auctioneers) (log+sqrt)
P (12 compromised auctioneers) (log+sqrt)
P (15 compromised auctioneers) (log+sqrt)
0
4 6 9 14 21 34 48
Number of Auctioneer Groups
400
350
300
250
200
150
100
50
Figure 5
Probability of Maliciously Revealing Next Bid Step
Decrypt Operations per Auctioneer
(10 bidders with 3 bid chains each per auction, 100 steps per bid chain)
Bid Chains per Auctioneer (one group per bidding step) (log)
Bid Chains per Auctioneer (two groups per bidding step) (log)
Bid Chains per Auctioneer (one group per bidding step) (sqrt)
Bid Chains per Auctioneer (two groups per bidding step) (sqrt)
Bid Chains per Auctioneer (one group per bidding step) (log+sqrt)
Bid Chains per Auctioneer (two groups per bidding step) (log+sqrt)
0
16 32 64 128 256 512 1024
Total Number of Auctioneers
Decrypt Operations per Auctioneer
der. On the other hand, all bid chains but
one are completely decrypted, and those
members of an auctioneer group who have
not decrypted a bid step must check the
proper decryption. As the latter only requires
the use of a cryptographic hash
function, the computation time for this
check can be neglected in comparison to
the asymmetric decryption operations necessary.
All in all, this leads to a computation
complexity in Oðz b n eÞ for the
entirety of bidders and auctioneers
throughout the whole auction process. The
complexity for one auctioneer depends on
the
number of auctioneers and is in
O
z b n e
.
A total
Communication complexity turns out
similar, as each bid step is sent to all auctioneers
of a group to allow for decryption
and its verification. This results in
an overall communication complexity in
Oðz b n e G size Þ.
Note that b and e are typically small
constants and the suggested group generation
functions also keep G size small.
Figure 5 illustrates the number of decrypt
operations for a single auctioneer.
For this figure, we assume that on average
10 bidders are involved in an auction
(z ¼ 10). Furthermore, we set b ¼ 3 and
n ¼ 100. Encryption with two group keys
for each bid step (e ¼ 2) is compared to a
single key (e ¼ 1). These parameters result
in 3000 (10 3 100) bid steps with 3000 or
6000 decrypt operations that must be processed
by the auctioneers. As can easily be
seen, complexity of the protocol is independent
of the group generating functions:
in all three cases the auctioneers have to
overall decrypt the same number of bid
steps. Only the number of encryptions per
bid step impacts the complexity as two decryption
steps per bid step double the
costs. Generally, Figure 5 shows that the
computation complexity of each auctioneer
is significantly reduced with a growing
number of auctioneers. With 256 auctioneers,
e.g., each auctioneer has to perform
less than 50 decrypt operations during the
whole auction.
Table 2
Proposed Protocol
Protocol Price Auctioneers Number of rounds Overall Computation Verifiability Non-Repudiation Asynchrony Spontaneous
entry and exit
Proposed 2 nd M not applicable Oðz b n eÞ Yes Yes Yes Yes
z – number of Bidders, M – constant, k – possible bids, b – bid chains per bidder (1 for the non-iterative auction protocols in table 1), n – bid steps per bid chain,
e – number of encryption operations per bid step
242
WIRTSCHAFTSINFORMATIK 48 (2006) 1, S. 7–15

Distributed Ascending Proxy Auction 15
6 Conclusion
We present a secure mechanism for distributed
ascending proxy auctions. It fosters
privacy and overall correct conduct to a degree
that enables the trusted application of
an asynchronous, iterative, second-price
auction in systems as open as P2P networks.
The approach produces several desirable
properties for the auction process: It eliminates
the dependency on one single auctioneer,
and the winning bidder can hide his true
valuation, respectively his highest bid.
Using an encrypted bid chain for bidding
ensures that only a fraction of information
is revealed to each auctioneer. With any new
bid chain each bidder can freely decide,
which auctioneer groups to trust. Robustness
is achieved by forming groups of auctioneers,
where even one group member
suffices to decrypt a bid step. All requirements
stated in section 2 are fulfilled.
In contrast to previous distributed second-price
auction mechanisms, our approach
is suitable for iterative open-cry
auctions. It is never necessary for all participants
to be online at the same time, which
is what makes other approaches highly vulnerable,
as it allows the blocking of an auction
fairly by attacking any participating
node [cf. Bran03]. In our protocol, all a
bidder needs to do is to convey his bid
chains, whereas the auctioneer groups must
be accessible during the whole auction process.
However, one single obedient auction
group member being online at a time suffices
to conduct the auction with a sufficient
standard of security.
Table 2 summarizes all the properties of
our ascending proxy auction. Obviously,
the proposed protocol can remedy some of
the shortcomings of the protocols presented
in Table 1.
The overall computation complexity
does not depend on the number of auctioneers;
hence, the computation per auctioneer
decreases linearly in the number of auctioneers.
The possible bids need not be
specified in advance, nor does computation
complexity depend on them. The values b
and e, on which our protocol’s complexity
depends, are small; n can be chosen arbitrarily.
Note that the proposed auction protocol
can only support 2 nd -price auctions.
As a next step, a transfer of the presented
notions to multi-unit and possibly combinatorial
auctions is envisioned. For these
challenging tasks, distributed algorithmic
mechanism design and cryptography have
to be combined to further realize the potential
of decentralized auctions.
Acknowledgement
This research was partially funded by the
German Federal Ministry of Education
and Research (BMBF) in the SESAM-project
as part of the research program Internet
Economics. The authors are responsible
for the content of this publication.
References
[AbSu02] Abe, M.; Suzuki, K.: Mþ1-st price auction
using homomorphic encryption. 5th International
Conference on Public Key Cryptography,
2002.
[BaSt01] Baudron, O.; Stern, J.: Non-interactive
private auctions. 5th Annual Conference on Financial
Cryptography, 2001.
[Bran02] Brandt, F.: A verifiable, bidder-resolved
auction protocol. AAMAS Workshop on Deception,
Fraud and Trust in Agent Societies, 2002.
[Bran03] Brandt, F.: Fully private auctions in a
constant number of rounds. 7th International
Conference on Financial Cryptography (FC),
2003.
[Bran05] Brandt, F.: How To Obtain Full Privacy
in Auctions. Preliminary Draft, February 9,
2005.
[Cach99] Cachin, C.: Efficient private bidding and
auctions with an oblivious third party. 6th ACM
Conference on Computer and Communications
Security, 1999.
[FoIM05] Fontoura, M.; Ionescu, M.; Minsky, N.:
Decentralized Peer-to-Peer Auctions. In: Electronic
Commerce Research Journal 5 (2005) 1, S.
7–24.
Abstract
[FrRe96] Franklin, M. K.; Reiter, M. K.: The design
and implementation of a secure auction service.
In: IEEE Transactions on Software Engineering
22 (1996), S. 302–312.
[HaSt05a] Hausheer, D.; Stiller, B.: Decentralized
Auction-based Pricing with PeerMart. 9th IFIP/
IEEE International Symposium on Integrated
Network Management (IM 2005), Nice, France,
2005.
[HaSt05b] Hausheer, D.; Stiller, B.: PeerMart: The
Technology for a Distributed Auction-based
Market for Peer-to-Peer Services. 40th IEEE International
Conference on Communications
(ICC 2005), Seoul, Korea, 2005.
[Kiku01] Kikuchi, H.: (Mþ1)st-price auction protocol.
5th International Conference on Financial
Cryptography (FC), 2001.
[KHAN00] Kikuchi, H.; Hotta, S.; Abe, K.; Nakanishi,
S.: Resolving winner and winning bid
without revealing privacy of bids. International
Workshop on Next Generation Internet, 2000.
[LiAN02] Lipmaa, H.; Asokan, N.; Niemi, V.: Secure
Vickrey auctions without threshold trust.
6th International Conference on Financial Cryptography
(FC), 2002.
[McAf00] McAfee, A.: The Napsterization of B2B.
In: Harvard Business Review 78 (2000) 6, S. 18–
9.
[RFHK01] Ratnasamy, S.; Francis, P.; Handley, M.;
Karp, R.; Shenker, S.: A Scalable Content-Addressable
Network. ACM SIGCOMM, San Diego,
CA, 2001.
[Schn99] Schneier, B.: Attack Trees: Modeling Security
Threats. In: Dr. Dobb’s Journal. December
1999.
[SMKK01] Stoica, I.; Morris, R.; Karger, D.; Kaashoek,
M. F.; Balakrishnan, H.: Chord: A scalable
peer-to-peer lookup service for internet applications.
ACM SIGCOMM, San Diego, CA,
2001.
Distributed Ascending Proxy Auction – A Cryptographic Approach
In recent years, auctions have become a very popular price discovery mechanism in the
Internet. The common auction formats are typically centralized in nature. The peer-to-peer
paradigm demands gearing up auctions for decentralized infrastructures. In this context, this
paper proposes a distributed mechanism for ascending second-price auctions that relies on
standard cryptographic algorithms. In essence, the auction protocol has the capability of
preserving the privacy of the winning bidder’s true valuation.
The auction protocol makes use of a high number of auctioneers divided into several
groups. A bidder creates an encrypted chain of monotonously increasing bidding steps,
where each bidding step can be decrypted by a different auctioneer group. This considerably
reduces the attack and manipulation possibilities of malicious auctioneers. In addition,
this secure approach does not require bidders to be online unless they are submitting their
bid chain to the auctioneers.
Keywords: Internet Economics, Distributed Auctions, Privacy, Trust, Security Protocols, Peerto-Peer
Systems
243
WIRTSCHAFTSINFORMATIK 48 (2006) 1, S. 7–15

Technology Assessment and Comparison:
The Case of Auction and E‐negotiation Systems
Gregory E. Kersten 1 , Eva Chen 1 , Dirk Neumann 2 , Rustam Vahidov 1
and Christof Weinhardt 2
1 InterNeg Research Centre, Concordia University, Canada
2 IISM, Universität Karlsruhe, Karlsruhe, Germany
An e‐market system is a concrete implementation of a market institution; it embeds one or
more exchange mechanisms. The mechanisms are—from the economic point of view—
disembodied objects (models and procedures) which control access to and regulate execution
of transactions. E‐market systems are also information systems which are information and
communication technology artifacts. They interact with their users; have different features
and tools for searching, processing and displaying information. This work puts forward an argument
that the study of e‐markets must incorporate both the behavioural economic as well
as the information systems perspectives. To this end the paper proposes a conceptual framework
that integrates the two. This framework is used to formulate models, which incorporate
the essential features of exchange mechanisms, as well as their implementations as IS artefacts.
The focus of attention is on two classes of mechanisms, namely auctions and negotiations.
They both may serve the same purpose and their various types have been embedded in
many e‐market systems.
244

2
1. Introduction
E‐markets are an important component of e‐business that bring demand and supply for goods
and services into balance. They are the meeting places for buyers and sellers who use exchange
mechanisms to conduct transactions. Exchange mechanisms are market institutions
providing sets of rules, which specify the functioning of the market and permissible behaviour
of its participants. Mechanisms include catalogues, where requests and offers are posted, negotiations,
where the participants bargain over the conditions of an exchange, and auctions,
where one side automates the process during which participants from the other side compete
against each other.
The design of exchange mechanisms is the concern of economists. The economists’ interest in
market institutions or mechanisms emphasizes problems of social decision making of who
shall get what resources. In their influential articles Hayek (1945) and Hurwicz (1973), formulated
the resource allocation problem as an information aggregation problem. Because the
valuation of different kinds of resources by individuals and organizations is typically private
information, it is difficult for a planner to define the efficient allocation that secures the best
use of resources by the society. To solve this social decision making problem, the economic
planner needs to extract this dispersed knowledge of the individuals by setting the right incentives.
Economists became involved with the engineering of market institutions, which involves the
specification of rules to control the exchanges and trade execution, since the 1990’s (Roth
2002). The field of market design provides a mathematical framework with which the outcomes
of particular mechanisms can be computed as equilibrium, and the mechanisms could
be designed to induce socially optimal outcomes (Maskin and Sjöström 2002).
In recent years, the economists have adopted laboratory experiments to confront theory with
empirical observations in their “toolbox” (Roth 1995; Smith 2003). The test‐beds used for these
experiments abstract from the real world situations in order to isolate specific human decision‐making
behaviours (Friedman and Sunder 1994). Isolation is achieved by inducing the
pre‐specified characteristics in participants of the experiments via a specially designed incentive
system. This incentive‐based control entails that the participants’ innate characteristics
can be assumed away (Smith 1994) and the market institutions can be implemented in a
highly simplified form.
An exchange mechanism can be used only if it is implemented in some and kind of an information
system (IS). Behavioural economists make a deliberate effort to devoid the system of
any features, which may contribute to its ease or difficulty of use or its appeal. This allows
studying the relationships between different incentive structures and mechanism outcome,
the selection of a revenue‐maximizing mechanism and separate market participants’ deviations
from perfectly rational behaviour.
Viewed from a different perspective, e‐markets are ISs deployed on the networks, which incorporate
(among others) exchange mechanisms. As such, they are of interest to
IS researchers, who study behaviours of users, with a particular interest in explaining perceived
usefulness, fit with the task, intention to use, as well as other important IS adoption‐
245

3
related factors (e.g., Goodhue 1995; e.g., DeLone and McLean 2003; Iivari 2005). Beginning
with the technology acceptance model (TAM) proposed by Davis (1989) to the unified theory
of acceptance and use of technology (Venkatesh et al. 2003), to TAM/user satisfaction integrated
model (Wixom and Todd 2005), users’ attitude and behavioural intention has been the
predominant concern of IS researchers.
In IS research four views of technology have been proposed: tool, perception, model and project
(Orlikowski and Lacono 2001). In some of these views certain facets of a system are ignored
or various components (and underlying models) are made undistinguishable. Depending
on the chosen perspective the studies of the technology use and performance focus on its
selected characteristics or actions. Arguably, this occasionally vague or partial consideration
of technology can be contrasted with a key purpose of IS studies which is technology performance
leading to its individual and organizational success (DeLone and McLean 2003; Garrity et
al. 2005) for which there have been “nearly as many measures as there are studies” (DeLone
and McLean 1992, p. 61). Knowledge of many rejected, underperforming and misaligned systems
led IS researchers to seek answers to such questions as: Will this system be adopted? Do
users have positive attitude about using the system? Do they intend to use it?
Many studies of technology use, adoption and performance consider the overall perception of
the experience with hardware, software and rules of their use in performing a particular task
(DeSanctis et al. 1994; Goodhue et al. 1995; Zigurs et al. 1998). If—as it is typically the case—
several technologies are bundled together in order to undertake the task, then such an approach
may not give useful results for the developers of a particular technology and/or the
technology adopting organization. Similarly, practical value of the assessment of software performance
that produces similar results but uses different embedded mechanisms is questionable
if these mechanisms are not explicitly distinguished.
Over the last 30 years ISs and other information and communication technologies (ICTs) underwent
dramatic changes. While the old systems were single‐purposed, the new systems allow
to dynamically binding different low‐level services into a service requested by the users
(Turner et al. 2003). The old ISs were difficult to modify; the new ISs have components with
partial and provisional connections which can be easily changed in order to adapt the system
to the user’s preferences and the task at‐hand (Orlikowski and Lacono 2001). The old ISs were
difficult to use; the new ones share common graphical interface features and require little
training. These and other changes led to ICTs being increasingly ubiquitous and entering all
dimensions of work and leisure. Often it is no more a question of an IS being used but of the
selection of the system or the compositions of ITC components and services so that the usersystem
interaction yields better results than if another choice was made. We therefore argue
that we need to concentrate on seeking comprehensive answers to the following question as:
What systems features, components and services are particularly useful for the given user, task,
and context?
In order to study these issues we choose e‐market systems because of their exemplary ubiquity,
ease of use and customization. They are well‐known information systems which embed
different mechanism and services. Some of them may be complementary (e.g., enhancing
shoppers experience), others may be competitive giving grounds to the question formulated
in the preceding paragraph.
246

4
The purpose of this work is to propose a research framework and associated models that allow
studying the performance of ISs. For this purpose we define IS technology as a system which is
designed in order to provide services to its users so that they can achieve their goals by interacting
with the system. Note that the system may be a single software program or several programs;
in the later case the system may be designed at the run time. Note also that we take a
narrow view of the system as being designed with a given purpose so that we do not consider
servers, modems and PCs as an e‐market system if they are also used for other reasons.
Typically, an IS has several ICT components which comprise a complex technological solution
and the services they provide. These components include user and software interfaces, data
storage and access, configuration and management, and computational models. What components
are used and in what way they are configured determines the system’s services.
Therefore, to assess the system usefulness and performance we need to be able to assess its
components or artefacts of a technology separately rather than jointly as a black‐box. To obtain
a comprehensive assessment we also need to consider all four views of technology (tool,
perception, model and project). In the case of e‐market systems this leads us to draw from behavioural
economics and IS research so that:
1. Performance is measured at level of individual user, at the group level (all users involved
in a single transaction), role‐level (e.g., buyers and sellers), and at the marketlevel;
and
2. Both behavioural (beliefs and attitudes) and economic (costs, returns, profits) measures
of process inputs and outputs are included.
This work proposes a conceptual framework that integrates the behavioural economic and the
information systems approaches to studying socio‐technical processes. This framework is used
to formulate a model, which incorporates the essential features of exchange mechanisms, as
well as their implementations as IS artefacts. The focus of attention is on two classes of
mechanisms: auctions and negotiations because they both may serve the same purpose and
their various types have been embedded in very similar and in very distinct e‐market systems
(Neumann et al. 2003). In addition to mechanism and system we also include three additional
constructs in the proposed research model: task, individual, and environment. The interaction
of these constructs results in the behavioural and in economic outcomes. Outcomes are used
to describe the transaction process and its results.
The interactions of the individuals and systems and mechanisms is the process, which is not
explicitly modelled in IS studies that typically use either variance or process models (Seddon
et al. 1999; Iivari 2005; Kim et al. 2006). One exception is the DeLone‐MceLean IS success
model (DeLone and McLean 1992)) in which system use is one of the modelled constructs.
The framework proposed in this work, called TIMES proposes to integrate variance and process
models, that is, explicitly and in‐dept consider process as it is a category of concepts which
explain causal relationships between input (independent variables) and output (dependent
variables). “One significant way to improve the robustness of process explanations in variance
theories is to explicitly observe the process argument that is assumed to explain why an independent
variable causes a dependent variable. To do so requires opening the proverbial 'black
box' between inputs and outcomes, and to directly observe process.” (van de Ven 1992, p. 170).
247

5
The TIMES framework provides a basis for the construction of models that allow studying information
systems from several perspectives, including:
1. Analysis of the use of and assessment of the efficacy of components and exchange
mechanisms;
2. Assessment of the impact of IS services and economic mechanisms on human interactions,
their behaviours and behavioural and economic assessments;
3. Study of the behaviours of the individuals, groups of individuals participating in a single
transaction, and individuals participating in the separate roles (e.g., buyers and
sellers); and
4. Development of prescriptions for advocating exchange mechanisms depending on the
participants and their objectives, context and the exchanged object(s).
The remainder of the paper is structured as follows. In Section 2 two exchange mechanisms:
auctions and negotiations are discussed and their hybrid forms are introduced. Sections 3 and
4 review the two perspectives on e‐market design, namely those of economics and IS. Section 5
proposes an integrated framework for assessing e‐market systems that incorporates the two
perspectives. Section 6 utilizes the framework to describe a concrete research model, which is
tested through the comparisons of mechanism and system in a laboratory experiment. Section
7 concludes with a summary and an outlook on future work.
2. Auctions and negotiations
A market mechanism is a set of rules governing the transaction process; it defines how the
market participants reach their agreements. These rules can be understood as mechanisms,
because they make behaviour more orderly and thus more predictable. The space of potential
market mechanism is large, since there are many ways to define the rules which can be imposed
on the exchange process. While the commonly used catalogue‐based exchanges is one
example of an institution, of greater interest to the researchers are classes of mechanisms,
which permit richer dynamics and more complex behaviour on part of the market participants,
such as auctions and negotiations (Wolfstetter 2000).
2.1 Auctions
Auctions are defined by an explicit set of rules which determine resource allocation and prices
on the basis of the bids made by the market participants (McAfee and McMillan 1987). The
four following characteristics differentiate auctions from other exchange mechanisms:
5. Auction rules are explicit and known to bidder prior to the auction. Therefore, rules
cannot be modified during the auction.
6. The rules completely describe the mechanisms allowing for the determination one or
more winners solely based on the bids. Auctioneers or any other party have no discretion
in winner choice.
7. Rules typically include: (a) bidding rules stating how bids can be formulated and when
they can be submitted; (b) allocation rules describing who gets what on the basis of
248

6
submitted bids; and (c) pricing rules stating the corresponding prices the bidders have
to pay (Reiter 1977); and
8. Auction rules allow mapping of bids on a single dimension – the price. They focus on
prices to achieve either an efficient allocation or revenue maximization.
The adequateness of utilizing the price mechanism has been conjectured by Hayek (1945) and
has been demonstrated under certain assumptions by Hurwicz (1973). If those assumptions
are violated (e.g. there are externalities in preferences), then richer message spaces, e.g. inclusion
of delivery schedules and quality in addition to prices, may become necessary (Mount
and Reiter 1974).
Auctions can be either single‐sided, where one seller auctioneers off goods to a number of
bidders, or double‐sided, where competition is employed on both sides of a market. Auctions
have been widely used in practice to allocate products as varied as securities, offshore mineral
rights and emission certificates (Wolfstetter 1995; Lucking‐Reiley 2000).
Modern auction theory has adapted to these needs of richer message spaces by including
other than price attributes (Bichler 2000). Multi‐attribute auctions and combinatorial auctions
have been designed but have not been used because of their complexity in formulating
strategies, eliciting valuations for combinatorial goods, solving the winner determination
problem and communicating bids (Parkes 2001). Most auction formats still rely on a single
price dimension.
Auction mechanisms dominate the economic view. In fact, the field of market design is almost
exclusively focussed on the theoretical and applied auction theory. This focus has a major
ramification upon practical market design: the decision whether to use auctions or other exchange
mechanisms favours auctions.
2.2 Negotiations
Negotiation is a rich and ill‐defined family of processes, used for exchanging goods or services
among buyers and sellers, and resolving inter‐personal and inter‐organizational conflicts. Negotiation
is an iterative communication and decision making process between two or more
participants who cannot achieve their objectives through unilateral actions. In addition, negotiation
involves exchange of information comprising offers, counter‐offers and arguments
with the purpose of finding a consensus (Bichler et al. 2003).
This notion of negotiations includes negotiation as an exchange mechanism, mediation, voting
and other forms of conflict resolution. Restricting the attention to trade negotiations only,
in which the participating agents engage in order to reach an agreement about an allocation
(which goods are exchanged to whom at what compensation), simplifies negotiations to bilateral
and multi‐bilateral encounters in which two roles are distinguished: buyers and sellers.
The peculiarity of negotiation as an exchange mechanism is its degree of freedom in the level
of structuring the process and the scope of exchanged information. Negotiations differ in the
degree of their structuredness, possibility of modification, and participation rules (Bichler,
Kersten et al. 2003; Ströbel and Weinhardt 2003). The process of interactions may be not prescribed
a priori as it is the case with many face‐to‐face negotiations; and the rules may be
249

7
known only implicitly, for example, based on tradition. Negotiations typically allow for the
modification of the protocol, but this characteristic may be limited in the case of e‐
negotiations. Also, the participation may be limited or determined by one part allowing another
to participate or not.
The above indicates the degree of flexibility in the design of negotiation mechanisms and this
is the key characteristic that differentiates them from auctions. A particular instance of multibilateral
negotiation can be indistinguishable from an auction, with offers having the same
format as bids and forbidden exchange of arguments and other messages. The difference is
that every negotiation allows for a change of its protocol while this is not allowed in auctions;
instead one auction would need to be cancelled and a new one initiated.
Negotiations mechanisms have been studied in economics and game theory. For example, bilateral
bargaining models were used to investigate how the surplus of a transaction is allocated
between two parties (Nash 1954). Recently, in response to the evolution of e‐
marketplaces, two streams of bargaining models have been studied in the field of economics
in order to explain market behaviour. In the first stream of studies several bilateral bargaining
processes between buyers and sellers have been investigated in order to determine the properties
of the aggregate market, which results from the bilateral relationships (De Fraja and
Sakovics 2001; Satterthwaite and Shneyerov 2003). One extension of this mechanism is a double‐multi‐bilateral
negotiation in which one party (e.g. a buyer) negotiates with several sellers
and each seller negotiates with several buyers. Such negotiations are similar to double‐side
auctions, but buyers are engaged in bilateral negotiations with many sellers and sellers are
negotiating with many buyers (Satterthwaite and Shneyerov 2006). The second stream of empirical
research extends bilateral bargaining models to multilateral bargaining (Thomas and
Wilson 2002). The latter are similar to single‐sided auctions in that they incorporate explicit
competition among all participants representing one side.
2.3 Hybrid forms
Auctions and negotiations are considered here as two distinct classes of market institutions.
The popularity of on‐line auctions led some researchers to state that internet negotiations are
auctions (Kumar and S.I.F. 1999), thus indicating that there is no need for negotiation mechanisms,
while others stated that e‐negotiations should be replaced with on‐line auctions (Segev
and Beam 1999). These statements, contrasted with beliefs that negotiations differ from auctions
in essence, may have been behind efforts to design mechanisms, which have certain
properties derived from both classes of mechanisms.
Some of the combined or hybrid mechanisms are well known, albeit little studied. For instance,
the mechanism used in hiring normally involves first a mechanism closely resembling
sealed bid auction with candidates submitting their resumes and other supporting documents.
In the second stage, a few of the “auction winners” are invited for interviews involving
negotiations over salary and benefits. For e‐markets, Teich, Wallenius et al. (2001) propose
NegotiAuction, a mechanism generalizing the process of auction followed by negotiation
(Leskela et al. 2006). A reverse process, in which multilateral negotiations are followed—if the
parties agree to do so—by an auction is suggested by Shakun (2005). These exchange mechanisms
are said to have a hybrid format.
250

8
Other types of hybrid formats occur in practice. For example, e‐Bay auctions allow for inclusion
of a “Buy‐it now” option that is typical for a catalogue, while Alibaba (http://alibaba.com)
combines catalogues with free‐text communication between the buyers and sellers.
3. Experimental and field studies
We are interested in the assessment and comparison of e‐market systems which includes both
economic and IS perspectives. Economic studies focus on market mechanisms, their functioning
and efficiency. The focus of IS studies is on the systems in which the mechanisms are embedded
with—in most cases—little concern for the mechanisms themselves. This indicates
that the two areas of research are complementary as we attempt to show through the discussion
of the theoretical and experimental studies.
3.1 Auctions and negotiations in economics
Research by the economists into auction mechanisms can be found in several theoretic studies
on single‐ and multi‐attribute auctions. Interestingly, the empirical research provides no
clear answer as to the superiority of one mechanism with respect to another. These results are
one of the motivations for the model we present in Section 5, which could be used in more
comprehensive comparative studies of both auctions and negotiations.
3.1.1 Theoretical comparisons
Bulow and Klemperer (1996) have shown in one of the first formal comparative studies that
simple English auction with N + 1 participating bidders always yielded higher revenue than a
negotiation with N participants. In essence, Bulow and Klemperer did not analyze a specific
negotiation protocol, instead they referred to an “optimal” mechanism. Such mechanism denotes
an abstract selling scheme, which is designed so as to maximize revenue of the seller.
The implication is that a seller should “devote resources to expanding the market than to collecting
the information and making the calculations required to figure out the best mechanism.”
(op. cit., p. 180).
This result does, however, hold only under fairly strong assumptions, e.g. attribute preference
independence. For affiliated preferences for example, the Bulow‐Klemperer (1996) result also
holds when the seller’s choice of the negotiation mechanism is restricted. If other assumptions
are relaxed (regularity, symmetry, etc.) the result may not be valid any more.
Kirkegaard (2004) showed that a seller‐offer bargaining game is more advantageous than an
English auction when demand is discrete, i.e., it has finite space, and the buyers are patient. In
those cases, sellers prefer a bargaining protocol over an English auction. Furthermore, Kirkegaard
(op. cit.) showed that when demand is continuous, an English auction can be improved
by some kind of pre‐negotiation.
The models proposed by Bulow and Klemperer (1996) and Kirkegaard (2004) focus on revenue
as the sole comparison criterion. If other objectives (i.e. allocative efficiency) are considered,
then these models do not provide valuable insights. But even in terms of revenue, the decision
whether to use auctions or negotiations crucially depends on the assumptions.
251

9
Both papers only refer to single‐issue mechanisms, a comparison between auctions and negotiations
that are capable of submitting several attributes beyond the price has not yet been
undertaken. The reason for this can be in the fact that studies of auctions that consider more
attributes have only been published recently.
Che (1993) and Branco (1997) initiated studies on the buyer’s payoffs in the two‐attribute (i.e.,
price and quality) auctions. The private information of buyers determining the utility can be
represented in one dimension; this shortcut allows applying the auction design apparatus to
these problems. More recently, Beil and Wein (2003) analyzed the problem of designing the
multi‐attribute auction. They were in particular concerned with finding a scoring rule to
maximize buyer’s utility.
3.1.2 Laboratory and field experiments
Thomas and Wilson (2002; 2005) conducted two experimental studies, in which reverse auctions
were compared with multi‐bilateral negotiations. In their most recent laboratory experiments,
they noted that general superiority of auctions predicted by Bulow and Klemperer’s
(1996) was not supported by the empirical data. In multi‐bilateral negotiations, a buyer
solicits price offers from multiple sellers and then the buyer requests more favourable offers
form the sellers who need to compete against each other. In their first experiment Thomas
and Wilson (2002) compared multi‐bilateral negotiations with first‐price sealed‐bid auction.
In the second experiment they replaced the first‐price with the second‐price (Vickrey) sealedbid
auction (2005).
Thomas and Wilson (2002) observed that for the inexperienced buyers and sellers multibilateral
negotiations with two sellers led to significantly higher prices than first‐price sealed
bid auctions. In the experiment with four sellers both mechanisms were found outcomeequivalent.
In their second study, Thomas and Wilson (2005) observed that prices in second
price sealed bid auctions exceed the prices generated in multi‐bilateral negotiations, suggesting
that this auction mechanisms is inefficient in the given experimental setting.
These two studies compared price‐only auctions and negotiations. Therefore, it is surprising
that in some experimental settings negotiations were found to be more efficient than auctions.
Many commercial and non‐commercial transactions concern objects characterized by multiple
attributes. Few are now supported electronically with multi‐attribute auction systems.
Ongoing and future work in auctions and negotiation suggests an increase in their use in
multi‐attribute transactions. At present time, we are not aware of any comparative studies of
multi‐attribute auctions and negotiations. Therefore, we restrict our discussion to selected
experiments with multi‐attribute auctions followed by discussion of negotiation experiments.
The highly stylized information exchange in auctions makes it impossible for the participants
(e.g., sellers) to learn the preferences (needs, limitations) of the auction owners (e.g., buyers).
Therefore, much effort in multi‐attribute auctions experiments has been devoted to the role
and scope of preference revelation schemes. Experiments in which the bidders (sellers) were
given the utility (value) function of the buyer show that multi‐attribute auctions do not provide
substantial benefits over comparable single‐attribute auctions (Bichler 2000). In other
252

10
words, even with fully‐revealed utilities the additional complexity outweighs the theoretical
gains.
The impact of information availability on the mechanism efficiency was studied by Koppius
and van Heck (2002). The information availability specifies the type of information that is
given to whom, or when and how it becomes available to whom during the auction. They considered
two types of multi‐attribute English auctions: (1) with unrestricted information availability,
in which suppliers are provided with the standing highest bid and the corresponding
bidder as well as score or bid ranking of the most currently loosing bids; and (2) restricted information
availability, in which the bidders are only informed about the standing highest bid
and bidder. The experiments indicated that auctions with unrestricted information availability
yield higher efficiency than auctions with restricted availability.
Strecker (2004) analyzed the impact of preference revelation schemes on the efficiency of
multi‐attribute English and Vickrey auctions. He concluded that English auctions with revealed
preference structure of the buyer are more efficient than Vickrey auctions, and English
auctions with hidden preferences. Chen‐Ritzo, Harrison at al. (2005) introduced a multiattribute
English auction, where only partial information about the buyer’s utility function
was revealed. They showed that this variant performs better in terms of efficiency than a single
attribute (price‐only) auction. This outperforming of the multi‐attribute over the single
attribute auctions holds even though the bids in the multi‐attribute auction were far away
from those predicted by the solution predicted by theory. Notably, complexity in the auction
mechanism consumes a portion of the efficiency gains over price‐only auctions. This observation
however, contradicts with the findings reported by Bichler (2000).
An empirical analysis of auctions and negotiations in the construction industry revealed the
use of the exchange mechanism depends on the knowledge and complexity of the context and
task (Bajari et al. (2002). Negotiations have advantages if the specifications of the product to
be traded are not well‐defined a priori, which is often the case in this industry. Negotiations,
unlike auctions allow for the discussion and clarification of the specifications. Not surprisingly,
their empirical analysis also reveals that auctions perform poorly in terms of efficiency
when changes in the product design need to be made after the transaction took place.
3.2 Lessons learned form economic research
The apparatus of economics has shed some light into the selection decision whether to employ
negotiations or auctions. However, while the theoretical results are unequivocal, the results
from experiments are inconclusive. The strength of economic approach to study market
mechanisms lies in its formal approach to experiment design and conduct. The power of economic
approach to markets is in its capability of abstracting the functions and mechanisms,
removing interfering events and processes and focusing on transactions. This strength coupled
with the focus on the mechanism design and efficiency may also be seen as a weakness,
in particular if the mechanisms implementation and functioning is an issue. In particular, we
consider here four issues of importance for our discussion.
1. Economics is interested in mechanisms and their economic properties rather than in
their users. The concerns are to determine if a mechanism functions according to its
design, what is its efficiency and what outcomes it produces given an assumed envi‐
253

11
ronment (e.g. preferences), which leaves out many relevant peculiarities of the users
(e.g. beliefs, emotions).
2. One of the most persistent issues within behavioural economics is the representation
of human behaviour. Traditionally economics begins with “the notion of homo
economicus acting in a world with full information, independent decision making,
polypolistic competition, transitivity, and fixed preferences” (Beckert 2003). It relaxes
some of the assumptions in order to study the violations of perfect rationality, but
does not consider the real decision‐makers “messiness” (e.g., deviation from
(bounded) rationality and adherence to various rationalities (Simon 1986)) making its
results difficult to implement in e‐markets.
3. The impact of the system in which an exchange mechanism is embedded is often ignored.
ICT imposes additional rules on human behaviour and it changes the behaviour
of the participating humans. The devotion of economics to analyze highly abstract
mechanisms “with little or no concerns for practical application” (Palfrey 2001) diminishes
the value of economic theory for assessing real‐world negotiation and auction
systems.
4. Economics largely focuses on price‐only mechanisms. Only a few papers address negotiations
or auctions that use multiple attributes in allocation. Neglecting multiattribute
problems diminishes the applicability of economics in procurement and
other settings, where the resources for sale are not specified from the outset of procurement
process. Although advancements have been made in the area of multiattribute
auctions but a comprehensive comparison with multi‐attribute negotiations
is still missing.
The few shortcomings mentioned here do not take away significant contributions of economics
to market research. The discipline of information systems with its concern for peoples’
perceptions, attitudes and feelings, may help introducing ICT to behavioural economics. IS
may allow for a more comprehensive e‐market framework which can be used for the assessment
of auctions and negotiation systems, in addition to the assessment of the mechanisms
embedded in them.
3.3 Information systems research
Information systems research has been concerned with the success of systems; their roles in
organizations and their impact on individual and organizational performance. Its purpose is
to “further knowledge that aids in the productive application of information technology to
human organizations and their management” (ISR 2002). Arguably, one of the reasons of IS
researchers focus on success and productivity is because of the “IT productivity debate” of the
1980s and 1990s and the fact that past empirical research generally did not significant productivity
improvements associated with IT investments (see for review of 150 articles Brynjolfsson
and Yang 1996). Economic studies at both micro and macro levels led to a “sobering conclusion:
our understanding of how information technology affects productivity either at the level
of the firm or for the economy as a whole is extremely limited” (Wilson 1995).
Glass et al. (2004) review of over 1,500 papers from IS, computer science and software engi‐
254

12
neering shows that the IS papers are predominantly concerned with evaluative and descriptive
research (76%) with validation coming from field studies (27%), laboratory experiments (16%)
and case studies (13%). Discussion on the IT evaluation measure has been, as DeLone and
McLean (1992) put it “the quest for the dependent variable” and it spanned the same period as
the discussion on the IT productivity debate. The results of this discussion are three main
models presented below and their numerous variants.
3.3.1 Technology acceptance model
TAM is one of the models most often used to explain the willingness of potential users to actually
use an information system (Davis 1989). In this, as well as in other IS models technology
is viewed as hardware, software, services and their combinations.
This model was extensively tested empirically (see, Legris et al. 2003 for a review) and several
extensions have been proposed (e.g. Szajna 1996; Al‐Khaldi 1999). According to the basic
model, the actual use of a system is determined by the behavioural intention to use a system.
This intention depends on the attitude towards the system, which is described by two subjective
factors: the perceived usefulness and the perceived ease of use of the system. The model
and the research instruments constructed within its framework have been found useful in
studying potential and actual information system adoption and use in organizations (e.g. Szajna
1996; Agarwal et al. 1997; Jackson et al. 1997).
More recently Vankatesh et al. (2000) proposed TAM2, which adds social norms to TAM. The
extended model had been tested in field studies and the authors found that social norms and
perceptions significantly influence user acceptance of technology. It should be noted, however,
according to meta‐analyses and field studies, most versions explain about 40% of the
system’s use (Agarwal et al. 1997; Venkatesh et al. 2000; Legris et al. 2003)
3.3.2 Task‐technology fit model
Technology may be accepted and used, and yet fail to bring forth expected changes. Goodhue
and Thompson (1995) proposed the technology‐to‐performance model and tested its part
known as task‐technology fit (ttf) model. The dependent variable is the individual performance
and the purpose of the model is to measure the impact of the technology on such.
The model considers tasks, which are activities requiring application of knowledge; they involve
synthesis, assessment, problem solving and decision making. It asserts that individual
performance depends on the fit between the task a user needs to undertake and the technology
used for this task (Goodhue 1995).
The fit between task and technology is “the matching of the functional capability of available
information technology with the activity demands of the task at hand” (Dishaw and Strong
1996). The better the fit is the higher user performance and, thus value of the technology. The
model also includes the moderating role of user’s characteristics on the user assessment of the
task‐technology fit.
The TTF model, like many other IS research models, considers technology as a complete computer
system (hardware, software and data) as well as organizational IS policies and services
(Goodhue and Thompson 1995, p. 216). While the individual characteristics of the user (e.g.,
255

13
education, training and experience) are distinguished, the technology is black‐boxed and no
specific features, tools and mechanisms are included. Recent modifications include the “exploration
of the black box” (Gebauer and Ginsburg 2006) in terms of system characteristics
rather than its structure. System characteristics include data quality, ability to retrieve and
consolidate required data and reliability. They are important but their assessment provides
little insight regarding system adaptation and modification.
The model shares its subjective orientation with TAM and other models; it models user attitudes
and behaviours toward the system (Dishaw and Strong 1998). The measurement of fit,
utilization and performance are based on the user beliefs rather than any objectively measurable
variables pertaining to cognitive effort, costs, time or productivity.
3.3.3 DeLone‐McLean model of information system success
The TAM, TTF and other early models could have been used to study almost any technology.
Although developed within the IS field, they do not consider the specific characteristics of
software and make no distinction between software, hardware and services of the IT departments.
The end user computing satisfaction (EUCS) model (Doll and Torkzadeh 1988) and the
information systems success (ISS) model (DeLone and McLean 1992; DeLone and McLean
2003) explicitly consider the attributes of information and system which produces information.
The purpose of EUCS is to measure satisfaction of IS users through the measurement of five
variables describing information content, accuracy and format, and the system ease of use and
timelines in producing data. The model has been widely used and considered as one that is
principally connected to the IS field with the premise that this field differs from social and
cognitive psychology (Doll and Torkzadeh 1991; Pikkarainen et al. 2006).
The ISS model was designed to explain the key factors that are accountable for the success of
information system projects. The key feature of the model is the separation of the quality of
information provided by the system from the operational measures of quality of a system. The
former includes such measures as information accuracy, precision, timeliness, and relevance,
while the latter refers to such indicators as reliability, response time, and ease of use.
The model aims at predicting the individual and organizational impact of the system based on
the attitudes formed by its users. The key factor conveying the attitude is user satisfaction.
According to the model user satisfaction is influenced by the two essential quality‐related factors
of the system. Satisfaction has a significant impact on the actual system use, and the latter
reciprocally affects the level of user satisfaction. Both of these constructs (use and satisfaction)
affect the construct of individual impact, which in turns influences the organizational
impact.
3.4 Lessons learned from information systems research
The major focus of IS research has been on the attitudes and behaviours induced by the use of
instantiated systems in organizations. This perspective brings in the psychological, organizational
and social factors into the study of the impacts of information systems. The variables
that describe subjective perceptions and aim at the facilitation of adoption of information sys‐
256

14
tems and their impact on the individuals and organizations have been identified. These findings
help explain how people perceive systems and which variables help successful adoption
of systems by the users.
Behavioural models discussed in the preceding section and their various refinements are
largely concerned with technology use, adoption, efficacy and satisfaction of its users. They
are less concerned with the issues pertaining to the identification of the concrete IS artefacts,
features and contributions to the achieved results. They are also less concerned with the identification
of types of users, decision problems and organizations that are particularly
(un)suitable for a given technological solution.
In this respect, it is worthwhile mentioning recent important concerns raised by the prominent
researchers in the field. The first one is concerned with the diversity of themes in IS research
and venturing by the researchers into the areas remotely related to the information
systems (Benbasat and Weber 1996). These tendencies, according to the authors are undermining
the very identity of the IS field (Benbasat and Zmud 2003). The authors stress that the
research should focus on the phenomena intimately linked with the core subject of the field,
which is information systems. On the other hand, another important and possibly related issue
is that of relevance vs. rigor. It has been stressed that IS researchers tend to overemphasize
the rigor in conducting research, while largely sacrificing the relevance of the studies
(Benbasat and Zmud 1999). In our view, the incorporation of system features in theoretical
models could help tackle both of these issues. Firstly, the IS artefact would become an organic
part of these models, thus battling the “identity crisis”. Secondly, this would enable illuminating
better design decisions during system development, as the theoretical findings will provide
direction and support as to which salient features and functionalities to incorporate for a
given class of contexts.
To construct prescriptive rather than descriptive models of socio‐technical systems there are
several issues which need to be considered, which are focussed specifically on the ICT technologies
and their configurations.
1. Presumably, technology is the focus when one tries to identify its usefulness, ease of
use and its other user‐oriented attributes. However, in many IS models technology is
not specified or even broadly defined. Models such as TAM and TTF can be applied to,
for example, computing devices as well as to transportation apparatus. There is nothing
that allows for the identification of the special nature of information systems viz.
some mechanical systems.
2. IS researchers have focused on instantiated systems in the past and paid little attention
to abstract features and models (e.g. mechanisms). This effectively restricted the
possibilities of incorporating the characteristics of systems in theoretical models.
While the IS success model does include the informational and system qualities, it
does not provide sufficient insight into the drivers of IS quality, e.g. whether it derives
from the model embedded in the system, or the way the model has been implemented
in the system. It could be beneficial to separate the concrete features of implemented
systems (e.g. user interface) from more abstract characteristics of the underlying
mechanisms.
257

15
3. IS literature tends to underestimate the value of the objective measures (e.g. outcomes)
associated with the system use, paying more attention to the subjective factors.
While the subjective categories (beliefs, attitudes, etc.) are undoubtedly important,
we believe that they are intimately linked to the objective variables, especially
when studying new economic institutions, like e‐markets.
4. Mechanism and system design
E‐markets are economic entities as well as information systems. These markets rely (in their
deep structure) on some family of abstractly defined economic mechanisms. They are also information
systems composed of software and hardware, which expose their functionalities
through specific features of user interface. Thus, the theories that would usefully allow researchers
to assess, compare, and prescribe e‐market systems must include both economic
and IS perspectives. One important question arising in connection with the possibility of such
integration is how the issue of mechanism/system design is viewed in economics and IS.
4.1 Basic concepts and theory building
Market design in economics traditionally uses axiomatic models to obtain propositions and
theorems about the impact exchanges have on resource allocation. Unlike other areas of economics,
market design has breed out a well‐accepted methodology for designing and analyzing
auctions. In recent times, laboratory experiments and numerical simulations have been
added to complement the study of auctions (Roth 2002).
A framework for economic market design research has been proposed by Smith (1982; 2003),
who sought to reconcile theoretic research with laboratory experiments. The framework is so
general that it applies to any microeconomic system. In essence, Smith suggests that the
analysis of alternative exchange mechanisms always share a common view on the structure of
the microeconomic system (Hurwicz 1973; Reiter 1977). The framework sketches an economic
system comprising economic environment, individual preferences, behaviour, mechanism,
outcomes and performance (see Figure 4).
258

16
Economic environment;
context
Individuals: preferences,
costs, resources, knowledge
Individual’s choice
behavior
Institution: market
mechanism, rules of
contracts, fullfilment
Outcomes:
allocations and prices
Mechanism
performance
Figure 1: Microeconomic system framework (adapted from Smith 2003)
The economic environment describes all factors (e.g., legal regulations, types of products and
services, customs and laws), which influence demand and supply in which economic agents
(individuals) operate. These factors are out of direct control of the mechanism designer. Deviating
from Smith’s depiction, we distinguish two parts of the environment, the context and
the individual characteristics.
The context characterizes the situation, in which the resources are being allocated. For example,
the context may describe a procurement setting with multiple sellers competing against
each other to serve one buyer. Economic theory attempts to eliminate context by abstracting
the resource allocation as much as possible.
The individual (economic agent) is described by all characteristics relevant to the decision
making, that is, his or her concrete choices. The key characteristics are preferences, risk attitude,
costs, resources (endowment) and knowledge.
The individuals’ are affected by the context, in which they operate. Their actual preference
structure and other innate characteristics determine their individual choice behaviour which
impacts, in turn, the functioning of the exchange mechanism.
The market mechanism is the set of rules and formulae which specify the way offers (bids,
proposal) can be formulated and how they are translated into outcomes (i.e., allocation and
prices). In other words, it is fed inputs (offers) and yields as its output an allocation of resources
among the participating users. The mechanism can be seen as a dialogue between its
users leading to an allocation of resources.
Individual choice behaviour does not depend solely on the individual and the environment
she or he operates in. The behaviour is a middle layer between the motivations of individual
buyers and sellers embedded in their local environment and the feasible actions confined by
the market mechanism and the resulting outcomes. The participating users need to formulate
their needs in terms, which are acceptable by the mechanism. Therefore, market institutions
restrict the users’ response behaviour, without uniquely prescribing it (Hurwicz 1973).
259

17
The outcomes of the use of the mechanism by individuals are the allocation of the resources
among the individuals and the corresponding prices which they pay. Finally, performance is
determined by the outcomes, the environment and the mechanism users. The performance is
of the mechanism rather than its users; it is defined by the comparison of the outcomes
achieved when two or more mechanisms are employed.
Given the framework (Figure 4), the mechanism design problem is the identification of a set of
rules for which that the equilibrium outcome satisfies the desired performance expressed by
an evaluation function. The evaluation function values the outcome attained by the mechanism
in a specific environment. This formulation is rather general, as it regards not only the
preferences of the individual, but also all other possible arguments of the environment.
Mechanism design seeks to describe mechanisms, which maximize the evaluation function
subject to the following three types of constraints:
1. Incentive compatibility constraints require that the participants truthfully report their
information about their local environment. In such a case, the outcome is the same as
if a benevolent arbitrator would have chosen the outcome on the basis of the full information
about the environment.
2. Computational compatibility refers to the complexity of the outcome function. Outcome
functions can be very demanding concerning computational tractability
(Rothkopf et al. 1998) and these constraints assure the feasibility of the outcome function
achievement.
3. Participation constraints require that the participants voluntarily take part in the
mechanism’s use. The individuals participate if the benefit they draw out of participation
is higher than participation in an alternative mechanism (Ledyard 1993).
In laboratory experiments, both the environment and the institution are controlled by the experimenter.
Achieving control over institution is in principle straightforward, as the experimenter
not only explains the institutional rules but also enforces them. Achieving control
over the environment is less easy, as the humans who participate in the experiment have idiosyncratic
characteristics that are unknown to the experimenter (and can hardly be observed).
If the experimenter wants to examine theories that rely on certain assumptions on the characteristics
of the individuals, the experimenter has to align the unknown characteristics of the
individual with the assumed behaviour. This is particularly challenging as these two can fundamentally
differ from each other.
Experimental economics usually employs induced values. By coupling behaviour with a rewards
scheme the experimenter attempts to induce the values upon the individual, overruling
the real characteristics. Since it is very difficult to control the context, experimenters frequently
try to abstract from the context as much as possible testing theory, whereas the exploration
of new mechanisms under real world is often neglected.
4.2 Design research in IS
The problem of systems analysis and design has long been one of the core subjects of study in
the area of information systems. Considerable attempts have been made in the past in devel‐
260

18
oping methods, techniques, and tools for guiding the development of concrete systems. However,
the study of the IS design is a relatively recent and gaining popularity and consideration
of an important part of IS research (Gregg et al. 2001; Markus et al. 2002; Dufner 2003; Hevner
et al. 2004).
Nunamaker et al. (1991) argued that system development is an important research methodology
to study information systems. They encouraged researchers to develop prototypes as
means of better understanding the research domain, thus emphasizing the instrumental role
of artifacts (i.e., the means) in IS research. Burstein and Gregor (1999) proposed and defended
the view of system development as a type of action research. They stressed that it can play an
important role in theory building, experimentation, and observation. While these views put
forward a model for IS research whereby IS artifact design is the means of conducting research,
a stronger perspective treats the artifact itself as the end.
Herbert Simon’s seminal book “The Sciences of the Artificial” had introduced the issue of design
research in the computing disciplines (1996). One of the most significant insights by
Simon was the separation of the outer environment from the inner environment of an artifact.
The outer environment could be regarded as the requirements imposed on the artifacts’ function.
The inner environment is the internal organization of the artifact. Simon regarded design
as the interface between the outer and inner environments of an artifact. March and
Smith have shown that the design and natural science approaches in IS research should be
complementary, whereby the design phase consists of building and evaluating IS artifacts, and
the “natural science” phase with theorizing abut artifacts and justifying the theories (March
and Smith 1995). According to them there are four types of products of design research, including:
constructs (language), models, methods, and implementations (this included both
prototypes as well as working systems).
Walls et al. (1992) proposed the notion of a “design theory” for information systems, in which
they envisaged the use of “kernel theories” in the design of a class of artifacts. According to
them design theories should consist of type of user requirements (“meta‐requirements”); type
of system solution or class of artifacts (“meta‐design”); and the type of methodology used to
develop such artifacts. Thus, in this work they had explicitly emphasized the shift of focus
from instantiated systems to the classes of artifacts. They had used an example of “vigilant executive
information systems” to present an example of a design theory. More recently, Markus
et al. (2002) employed this concept of a design theory to devise the characteristics of a broad
class of systems that the authors called “systems that support emergent knowledge processes”.
In a recent publication Hevner et al. (2004) stressed the necessity of emphasizing rigor in
conducting design research and advocated a number of guidelines for design researchers.
In summary, the most important contributions of the design research to the study of information
systems include:
1. Approaching the design of IS artifacts as a legitimate and rigorous research initiative;
and
2. Recognition that the IS artifact in the context of design research refers to classes of systems,
rather than specific instantiations.
Together these principles help bridge the gap between the IS community on one hand and
261

19
those interested in the design of markets and economic exchange mechanisms on the other.
4.3 Convergence of IS and economics designs
Notable attempts have been made in the recent past by the researchers in both fields to entertain
the possibilities of adopting alternative perspectives to studying the core subjects of their
respective disciplines. While information systems are engineered artifacts, engineering approaches
seem to have wider applicability in somewhat less expected areas. Roth (2002) have
advocated an engineering view of economics and pointed at the emerging discipline of design
economics: “the part of economics intended to further the design and maintenance of markets
and other economic institutions” that would rely on the use of computational models and experimental
approaches. Subrahmanian and Talukdar (2004) have discussed how market design
can be formulated as an engineering problem.
Researchers in IS have started integrating economic perspective in their assessment of the utility
and impact of systems, in particular those that support e‐commerce applications (Zwass
1996; Bhargava and Sundaresan 2004; Zhu 2004). The recent discussion on the nature of the
core of the discipline highlights important alternative models for conceptualizing information
systems. Alter (2003) proposes and defends what he calls the IT‐reliant work system view,
where the emphasis is placed on the business and organizational processes as supported by
the information systems, rather than focusing on purely technological aspects. El‐Sawy (2003)
advocates the adoption of the “fusion” perspective, according to which the IS artifacts have
merged with the respective processes and entities in business environment to an extent that it
is no longer possible to separate them from each other.
As it has been stressed earlier, design research aims at producing generic system solutions to
practical problems, while preserving the level of rigor characteristic of other, more traditional
modes of research. The type of system solutions proposed by a design researcher is a class of
systems, or meta‐systems (Walls, Widmeyer et al. 1992; Iivari 2003). According to one proposed
representational framework, a researcher’s view of an abstract IS artifact can be organized
according to a number of perspectives (Vahidov 2006).
Following these lines of interdisciplinary study, in the next section we propose a unifying
theoretical model that allows studying the impact of exchange mechanisms together with
other contextual variables on the relevant dependent variables derived from the fields of information
systems and economics.
5. TIMES framework
The advance of electronic commerce and new forms of technology‐enabled exchange models
brings together the fields of IS and economics closer than ever before (Zwass 1996; Bhargava
and Sundaresan 2004; Zhu 2004). This de‐facto merging of the structures of exchanges among
participating economic entities and the types of evolving system solutions to facilitate such
exchanges necessitates the development of novel research models to study the resulting forms
of amalgamated mechanisms. Such models must necessarily integrate the existing theoretical
approaches from different and complementary research areas; this is the purpose of the TIMES
framework, it provides a common foundation for models that may be constructed with a nar‐
262

20
row and concrete focus. Examples of such models are given in Section 6.
5.1 Basic concepts and dependencies of the TIMES framework
The developments discussed in Sections 3 and 4 facilitate construction of a research model
that draws on the theoretical constructs derived both from economics as well as IS field. The
relevance of these constructs has been studied in behavioural economics and IS. An integrated
model puts forward relationships between the independent variables, including type of
mechanism and contextual factors on one hand, and the performance‐related criteria on the
other. The factors that have their origin in economics include mechanism, environment, and
individual characteristics which drive the choice behaviour of the market participants ultimately
resulting in certain outcomes.
The IS literature has been studying, among other phenomena, the effects of the system, individual,
and task characteristics on the performance and on the subjective (perceived) assessments
of the systems. In the integrated model proposed here the characteristics of: task, individual,
mechanism, environment, and system—jointly called TIMES—are included as independent
constructs. Below we will introduce basic concepts and dependencies first by pointing out
the differences with the microeconomic system framework and with the IS research models.
5.1.1 From the economic environment towards task, individuals and IS environment
Not only economic theories but also laboratory experiments too often define task complexity
away (e.g. in the context of market design the determination of the bidding strategy) by assuming
that people can determine an optimal strategy no matter how complex the task might
be. However, humans tend to adopt simplifying strategies if the task is complex. They also
may ignore complexity. For market design, this implies that the designed exchange mechanisms
may be too complex for humans to derive the optimal strategies.
Complex negotiation tasks that require substantial cognitive efforts tend to lead to suboptimal
solutions (Hyder et al. 2000). People’s cognitive limitations, their lack of interest in engaging
in highly complex transactions, and their involvement with many competitive activities often
lead to their selection of a quick and simple tool which does the work even if the results are
not optimal. Some may know that the use of a simple tool allows them doing more elsewhere;
others may be unable of learning the tool’s intricacies. This latter issue has been studied
within the IS domain.
Having in mind the importance of the task, it appears reasonable to separate the task from the
environment, where the task construct refers to the properties of the goals that need to be
accomplished through the use of an exchange mechanism. Rangaswamy and Starke (2000)
provide the characteristics of bargaining orientation, degree of conflict, time pressure, and
complexity for describing tasks.
The environment construct captures the environmental factors that may have impact on the
negotiation outcomes, including type of market, type of product, level of competitiveness
among buyers and sellers, and other important contextual considerations.
5.1.2 From institutions to mechanisms and systems
263

21
One lesson learned by the economics is that the institution matters. We use the term mechanism
here to reflect the economic notion of institutions. The mechanism reflects the convergence
point of institution construct from the economics perspective and meta‐system from
the IS perspective. Accordingly, the mechanism is an abstract artefact describing the protocol
and mode of exchange regardless of its implementation. The description of the mechanism
can be based upon the Montreal Taxonomy, which provides a comprehensive schema for classification
of such mechanisms (Ströbel and Weinhardt 2003). The characteristics include, but
are not limited to: flexibility, rounds, concurrency, number and nature of attributes of offers,
offer matching, offer evaluation, and others. Additionally, the availability of analytical support
(decision support capabilities) and mediation can be added for a useful description
(Rangaswamy and Starke 2000). Taking only the mechanism into consideration is, however,
not sufficient, as it abstracts away from the implementation and thus limits the consideration
of subjective factors explaining its acceptance.
Different from economics, the TIMES model regards the way and form a mechanism is presented
to its users, its embodiment into a system and the interaction process between the system
and the user to be crucial for studying both systems and their mechanism. We know that
the way problems are presented and the way people are prompted to make choices and solve
problems, affects their behaviours. A mechanism has to be implemented in some medium; it
has to communicate with the user using media. These are design issues and they are no less
important for the mechanism, the outcomes and the performance than the mechanism itself.
Thus, we include the construct system in our model; it reflects those characteristics of the instantiated
systems that are implementation‐specific, including user interface, various features,
and functionalities that process and manage information required by the market participants
(Lumuscio et al. 2003).
5.1.3 From individuals’ choice behaviour to individual
The individual construct refers to those aspects of individuals that tend to be relevant to market
processes and outcomes, including: individual characteristics, number of users, as well as
psychological issues such as attitude, beliefs etc. Here, the purpose is to explain the choices
made by the individuals in light of their characteristics, and their interactions with the task,
system, mechanism, and environment.
5.1.4 From individual to individual, group, role and mechanism
Experimental IS research focuses on the users and their behaviours. The focus of economic
research is on the design of market mechanisms and their functioning. In situations when the
use of a system is required and accepted, an important question is what system to use, by
whom and for what types of tasks. These are the types of questions that both system owners
(firms that provide e‐market services) and users ask. User satisfaction, intention to use a system
and other behavioural user perceptions are important. But no less important are the specific
aspects of the system (including its mechanisms and functions).
The independent constructs of the TIMES model impact, through the user‐system interactions,
the individual behavioural perceptions and outcomes and also the economic outcomes. Both
types of outcomes describe performance of individuals, groups participating in a single transaction
(e.g., dyads in bilateral negotiations), roles (e.g., buyers, sellers and auction owners),
264

22
and the mechanism performance (e.g., English auction and Vickrey auction).
5.1.5 From system performance to perceptions, and behavioural and economic outcomes
The assessment of performance used in the TIMES model is more complicated than the economic
notion of performance. The dependent variables—partly derived from the IS research—
are both objective and subjective or “perceived”. The former type seems to be emphasized in
economics, while the latter has been the focus of extensive behavioural investigations in the
area of information systems. The importance of subjective assessments lies in the fact that the
latter tend to contribute towards the adoption of the systems by individuals. We have incorporated
economic performance as an objective dependent construct in our model and two
types of behavioural variables: perceptions and behavioural outcomes. Perceptions describe
user assessments and beliefs of the system and its features and objects, including usefulness,
ease of use for such tasks and aspects as communication and learning about the counterpart
priorities. Behavioural outcomes pertain to the subjective assessments of the final results, e.g.,
satisfaction with the agreement, satisfaction with the process and relationship established
with other participants.
5.2 Dependent constructs of TIMES
The overall impact of TIMES variables on important dependent constructs is shown in Figure 2.
As we have already introduced the TIMES (independent) variables, in the section we will elaborate
on the choice of the dependent ones, as well as the impacts on them.
Figure 5. TIMES framework
5.2.1 Process
The transaction process is clearly identified in the microeconomic framework (Section 4.1);
the choices made by the individuals, and the information provided by the mechanism and
other market participants is the interaction process in which e‐market system plays a role. IS
models, with the exception of the DeLone‐McLean model (Section 3.3.3), tend not to model
265

23
the interaction process. Most models are variance models and according to Seddon et al.
(1999), “variance‐ and process‐model diagrams represent quite different concepts and cannot
be combined meaningfully in one model”. Variance models explain variation in a dependent
variable by studying its association with one or more independent variables while process
models explain an outcome by identifying the sequence of events or states that precede it.
E‐market systems make interactions between market participants possible; these interactions
have social aspects and the process perspective allows studying the different activities undertaken
by the participants. These activities lead to the behavioral and market outcomes, and
thus the introduction of process constructs that allow for the actual use of the system’s services
and their impact on the outcomes. The TIMES framework looks at the efficacy of mechanisms
and services, the development of prescriptions for advocating exchange mechanisms
and service bundles, and the need for process consideration is similar as in the IS development
(Robey and Newman 1996; Kim and Pan 2006).
The interaction process includes formulating and submitting offers (possibly accompanied
with and/or embedded in messages or contract forms) and receipt and analysis of offers made
by other participants. There may also be other activities that involve user‐system interactions,
such as, agenda setting, preference elicitation and value (utility) function construction, and
the specification of reservation levels and the best alternative to the negotiated agreement.
The process that involves these activities may lead to the users’ reassessments and change of
their strategies and tactics.
The consideration of the process categories allows including non‐monetary costs of the transaction
and study the relationship between mechanisms and systems and these costs. For example,
in one of the preliminary studies of trade negotiations conducted via an e‐negotiation
system ENS) it was found that a service that provides graphical representation of the negotiation
dynamics does not affect behavioural and market outcomes but it leads to much shorter
(334 words on average) messages exchanged (Weber et al. 2006). This indicates that graphical
service reduces users’ efforts; we would not be able to observe this relationship without explicit
consideration of process.
5.2.2 Perceptions: ease of use, usefulness and other assessments
The technology acceptance model (TAM) postulates perceived usefulness and ease of use being
important factors that help predicting actual usage of the system in question. Because of system
adaptability and adjustment potential, we are less concerned with the actual usage than
with the users’ perceptions of the services and mechanisms ease of use.
The effect of ease of use on perceived usefulness had been postulated by TAM (Davis 1989) and
widely studied since the early 1990s. The TAM model had been recently criticized for ignoring
important variables related to system features (Wixom and Todd 2005). An alternative stream
of research had been focusing on the way system features influence users’ beliefs and attitudes,
including, most notably user satisfaction (Baroudi and Orlikowski 1988; Doll et al.
1998). We propose to extend the studies to include mechanisms embedded in systems. We
should mention here that the main difference between the usability studies and the services’
ease of use and usefulness assessment is in the detailed, during‐the‐process measurement in
usability engineering (Nielsen 1994) and post‐transaction measurement in the TIMES frame‐
266

24
work.
E‐market systems may be used to engage in purely economic transactions; they may also be
used for socio‐economic transactions which involve such issues as trust, relationships and affect
(Bakos 1998). The system contribution to the users’ perceptions of effective communication,
gaining understanding of the counterparts and competitors, and trust and relationship
building may be valuable and no less important than its ease of use and usefulness. While usefulness
is considered in a narrow, transaction‐oriented sense, occasionally the social aspects
which arise during the transaction process may be considered more important than the purely
economic ones.
5.2.3 Behavioural outcomes: satisfaction, relationship and other outcomes
A model integrating important service‐based and behaviour‐based beliefs had brought together
the important theoretical constructs to provide a more comprehensive picture of system
adoption (Wixom and Todd 2005). In this model satisfaction with the way the system
provided informational support was found positively related to the perceived usefulness of the
system.
In the TIMES framework we are interested in, on one hand, the ease of use and usefulness of
services and the satisfaction with different facets of the overall process (experience). Therefore,
the construct satisfaction is a behavioural outcome associated with the achievement of
both goals established prior the transaction process and the factors that the users did not consider
beforehand. Satisfaction is an integrative evaluation of the person’s attitudes as they
form during a process (Yi 1990); it is a high level assessment that integrates other cognitive
components. Thus, we have incorporated satisfaction as an important factor of the perceived
usefulness of the system. Since we are interested in both important objective (economics) as
well as subjective (behavioural research in IS) factors our framework postulates that the satisfaction
and other behavioural outcomes may be influenced by the objective performance of
individuals.
Based on a review of IS literature and research instruments Yu (2007) identifies eight categories
of satisfaction: 1) satisfaction with the achieved outcome, 2) satisfaction with own performance,
3) satisfaction with the relationship with other participants, 4) satisfaction with information
obtained from and via the system, 5) satisfaction with the communication media,
6) satisfaction with transaction process, 7) satisfaction with the system, and 8) overall satisfaction.
We propose considering these categories because they allow for the assessment of different
dimensions of the system, its services and the process.
5.2.4 Market outcomes
Objective (market) outcomes are separated from subjective evaluations because the former
describe results achieved by the individuals and their different groupings, such as, all individuals
who participate in the same process (e.g., bidding), all those who play the same role
(e.g., buyers, sellers), and all those who use the same mechanism or system. Market outcomes
include utility values, prices, distance to efficient frontier, distance to axiomatic solutions, and
social welfare distribution. Performance measures also include agreement rate, balance of the
deal, early withdrawal rate and, in addition to utility or profit, they may also reflect equity.
267

25
Also, the recent study suggests that economic outcomes do not necessarily behave identically
as behavioural ones (Galinsky et al. 2002). For example, when high targets are stated, negotiators
tend to be dissatisfied with the objectively superior outcomes. We also note that the satisfaction
is individual while the market outcomes not necessarily so (some depend on more
than one person).
6. Experimental assessments
The verification of the TIMES framework and its construct requires several experiments and
field studies because the framework includes behavioural, economic and technical perspectives,
all of which allows for the construction of different models. By providing a common
foundation for comparative behavioural studies of information systems the framework allows
studying various systems and their components. A thorough and exhaustive TIMES assessment
is beyond the scope of this work; we hope that the two studies reported in this section will be
followed by a larger number of experimental and field studies.
6.1 The NorA project
For the purpose of comparison of e‐market systems, we need a detailed specification of the
model constructs and the formulation of the associated research instruments. The comparison
has been undertaken as part of the project called “Negotiations or auctions” (NorA) (InterNeg
2006; Yu 2007). The project focus is the comparison of two market mechanism and two e‐
market systems. The overview of the project and its relationship to the TIMES framework is
illustrated in Figure 6.
Figure 6. Research model for the NorA project
In the first study the same software platform was used with a very similar interface but with
different market mechanisms, i.e., auction and negotiation (Yu 2007). The second study compared
different system interfaces and different internal implementations of the same mechanism
(English auction) in (Kolitz and Neumann 2007). In our attempt towards model building,
we investigate the effects of mechanism and system based on general hypotheses of process
and outcome variables, rather than postulating formal hypotheses.
6.1.1 Two mechanisms
268

26
While widespread popularity of on‐line auctions is well known, however—as we mentioned in
Section 2.1—their claimed superiority is not obvious. Therefore, we would like to attempt delineating
the types of individuals and transactions (tasks) for which auctions are preferable
and those, where negotiations are more suitable. This is one purpose of the NorA project for
which the TIMES framework was used to construct models. Another purpose of the project is
to study the ease of use and usefulness of specific features and tools of the systems, and test
the relationships between the perceived ease of use and usefulness and the individual and task
characteristics.
For the purpose of illustration of the framework application we consider the following two
exchange mechanisms:
1. Multi‐attribute English auction with a buyer being the auction owner; and
2. Multi‐bilateral purchasing negotiation involving one buyer and several suppliers.
The auction bidders can submit multiple bids during the process. Between the bids they can
have an information feedback and they have the chance to adjust their bids. This characteristic
resembles negotiation to some extent, and makes the comparison easier.
The two mechanisms can be further characterised by several parameters, each taking one or
more values. The concrete mechanism implementation requires that the transaction rules and
all their parameters are specified. The main parameters describing the two mechanisms and
the values selected for illustrative purposes (and used in NorA experiments) are listed in Table
1.
Table 1. Two exchange mechanisms and their selected parameters
Parameter Auction Negotiation
Issues (auction attributes) Four Four
Issue values Given by the principal Given by the principal
Knowledge of owner’s utility Partial Partial
Offers Complete Complete or incomplete
Offer commitment Binding Binding
Messages No Yes with owner
Termination Time and reservation level Time and bilateral agreement
Transparency Full, best offer None, final offer
The difference between auction and negotiation has been the subject of numerous studies.
Thomas and Wilson (2002, 2004) found that negotiations and auctions were similar in economic
efficiency for a market of four bidders or greater competing on a single attribute. But,
when the number of bidders was reduced, negotiation produced higher revenues for the bidders
(i.e., less efficient). Therefore, we hypothesize that:
H1a. Negotiation will generate higher revenue for the bidders.
The behavior outcomes consist of satisfaction with outcome, which is the difference between
the results of the exchange in comparison to the expectations that have anchored the process,
and satisfaction with process points to the emotions expressed and the degree of conflict felt
by the participants (Suh 1999). Ivanova‐Stenzel et al. (2005) found that participants were more
269

27
satisfied with the results derived from an auction because the process is more transparent. We
expect that:
H1b(i) and H1b(ii). Auction will generate greater satisfaction with outcome and satisfaction
with process for the bidders.
Even though the mechanism is embedded into the e‐market system, the difference in the
market institution would not interfere with the user’s perceptions of the system. We posit
that:
H1c. Auction will have no differential effect on IS perceptions (i.e., usefulness, ease of
use, information quality, system reliability) relative to negotiation.
As auctions require little to no interaction among market participants and the owner, we hypothesis
that:
H1d. Auction will reduce cognitive effort for the bidders.
6.1.2 Two systems
The two mechanisms are embedded in systems generated by the Invite software platform
(Strecker et al. 2006). Invite software platform allows setting up systems supporting different
auction and negotiation processes and varying the range and type features and tools available
to the system users. One of important capabilities of the platform is the ability to easily modify
the interface of the generated systems. Figure 7 shows two screen shots, the left one comes
from the auction system and the right‐hand side comes from the negotiation system. In order
to reduce the potential impact of the system interface “look and feel” on process and its results
the differences between these two systems are minimal (the colouring is different only
for illustrative purposes). However, both systems have complete functionalities and provide
support for offer (bid) analysis, selection and comparison.
270

28
Figure 7. Screen shots (view history) of two Invite systems: English auction (InAuction)
and multi‐bilateral negotiation (Imbins)
meet2trade is a generic, flexible trading platform facilitating the creation and automation of
auction‐based markets. The platform is flexible enough to host markets from a large variety of
domains and to support various market mechanisms. It can be configured by a Market Modeling
Language (MML). This language was developed to describe electronic market parameters
and to enhance the development and creation of electronic auctions (Mäkiö and Weber 2004).
The experimental system MES allows us to conduct economic experiments on electronic markets
with rigorous control on the procedures (Kolitz and Weinhardt 2006). meet2trade follows
a client‐server architecture with a central server. The server provides the running platform
for all available markets as well as the hosting of all data (e.g. user data, account data,
product information, protocol data) and the data preparation. The clients connected to this
central server display this data and provide an interface for submission of bids and for displaying
relevant information. Table 2 provides a comparison of characteristic found in Invite and
meet2trade, and Figure 8 shows a multi‐attribute, English, auction on meet2trade.
Table 2. Two e‐market system platforms and their characteristics
Characteristics Invite meet2trade
User interface Two screens One screen
Functionality Web browser Java applet
Features Offer rating, activity table, history graph Offer rating, activity table
Figure 8. Screen shots of meet2trade (multi‐attribute English auction)
Given these different instantiation of Invite and meet2trade as independent variables, the following
general propositions are made taking into account the past research in e‐market and
271

29
related exchange systems described in the previous sections.
From the economic perspective, market outcome is the result of the mechanism that governs
the transaction process and not due to the system that mediates the exchange. We hypothesize
that:
H2a. InAuction will have no differential effect on revenue relative to meet2trade.
While the systems have different user interface, they process bidding information in the same
manner, which reduces variation on the impact of sociological aspects on outcome (Delaney
et al. 1997). We postulate that:
H2b. InAuction will have no differential effect on behavioral outcomes (i.e., satisfaction
with outcome and satisfaction with process) relative to meet2trade.
The characteristics of each system play an important on its perception according to cognitive
theory. Given that Invite provides a history graph of the user’s performance during the auction,
we posit that:
H2c(i) and H2(ii). InAuction will be perceived as more useful and as providing higher information
quality.
On the other hand, meet2trade functions on one screen such that users readily see all bidding
information. In addition, meet2trade uses Java applet technology, which does not require refresh
time that a browser would need. Thus, we expect that:
H2c(iii) and H2(iv). meet2trade will be perceived as easier to use and as being more reliable.
6.1.3 The experimental setup
Both studies employed a contract case involving: (1) an agent representing an artist, and (2)
representatives of three entertainment firms interested in signing up the artist. The case entails
four discrete issues: (1) the number of song the artist must introduce each year, (2) the
royalties given to the artist for each CD sold, (3) a contract signing bonus, and (4) the number
of promotional concerts that the artist must perform each year, all of which are comprised of
various options. The importance for each of these issues was explicitly defined for the subjects
in the negotiation.
Participants consist of university students from a major Canadian city. They play the role of
agents who represent one side of the case, and they are randomly assigned to different treatments
depending on the study. The experiments were divided into three phases: preexchange,
whereby participants answer questions about the case, their reservation values and
expectation, as well as their conflict approach; exchange, whereby subjects interact with the
system in hopes of winning a successful contact that is favorable for the principal; and postexchange,
whereby participants answer questions on their experience related with the research
variables.
6.1.4 The measures
272

30
The variables are categorized as: (1) independent variables, (2) subjective evaluations of IS
perceptions and behavioral outcomes, and (3) objective measure of market outcome.
The independent variables were manipulated according to the e‐market system (Imbins, In‐
Auction and meet2trade) following their difference in mechanism and system.
The subjective dependent variables were measured using slightly altered items from previous
research to meet the needs of this project (satisfaction with outcome and process were
adapted from (Suh 1999); cognitive effort was developed based on (Lilien et al. 2004); usefulness
and ease of use were modified from (Davis, Bagozzi et al. 1989); and information quality
and system reliability were adapted from (Wixom and Todd 2005)). Factor analysis served to
assess the quality of the seven constructs (information quality, system reliability, usefulness,
ease of use, satisfaction with outcome and satisfaction with process). In Table A of the appendix
show the factor rotation whereby convergent and discriminant validity are supported. It
also relates the reliability of the items assessed by Cronbach alphas between 0.708 and 0.896).
The objective market variable is the revenue obtained by participants after the exchange.
Since only one bidder can win the auction or negotiation in each market, this variable is determined
solely by the winners.
6.2 Results
In experiment 312 students participated; most were aged between 20 and 30 years old; and the
male‐female proportion was the same for both mechanisms. There were 20 multi‐bilateral negotiations
(three parties negotiated with a single counterpart) and 27 auctions with three bidders.
The effective samples for this study (i.e. those instances with complete questionnaires)
enabled us to conduct analysis on both objective and cognitive measures. The data analysis of
subjective measures is only for the buyer side (i.e. the music companies) because there was no
auctioneer in auction sessions, such that the sample consist of only 272 records.
6.2.1 Impact of mechanism and system
The responses collected from the participants were tested using a series of analysis of variance
(ANOVA). The results are reported in Table4 with significant effects highlighted.
The market outcome was expected to be affected by the mechanism (H1a) and not the system
(H2a). However the results show the opposite, meaning that negotiation is no better than auction,
but InAuction generates higher revenues than meet2trade.
Users related greater satisfaction with outcome and process for auction compared to negotiation
as predicted by H1b (i) and H1b (ii). There was no difference between the two auction systems,
which supports H2b (i) and H2b (ii).
H1c predicted that difference in mechanism would not result in difference in IS perceptions,
and it is supported. In terms of IS perceptions when different systems are used, meet2trade
was found to be easier to use and more reliable, supporting H2c (iii) and H2c(iv). InAuction
273

31
did not result in greater usefulness or information quality as expected in H2c (i) and H2c(ii).
The cognitive effort required in the exchange process follows H1d and H2d, which means that
auction requires less effort than negotiation and systems have no effect.
Market outcome
Table 4. Summary of experimental results of mechanism and system comparisons
Imbins InAuction meet2trade
Negotiation Auction Auction
Invite Invite meet2trade
Mechanism
ANOVA
System
ANOVA
mean (SD) mean (SD) mean (SD) F (p‐value) F (p‐value)
• revenue 4.95 (16.48) 7.27 (11.04) 0.57 (8.51) 0.44 (0.51) 6.46 (0.013)
Behavioral outcomes
• satisfaction
‐0.32 (1.86) 0.46 (1.60) 0.28 (1.33) 7.44 (0.007) 0.66 (0.42)
with outcome
• satisfaction
1.35 (1.33) 1.72 (0.84) 1.56 (0.84) 4.30(0.04) 1.48 (0.23)
with process
IS perceptions
• usefulness 1.21 (1.34) 1.31 (1.06) 1.45 (1.11) 0.57 (0.45) 0.77 (0.38)
• information
1.50 (1.16) 1.58 (0.91) 1.57 (0.99) 0.27(0.61) 0.013 (0.91)
quality
• ease of use 1.93 (1.14) 1.81 (0.98) 2.14 (0.72) 0.37 (0.54) 7.08 (0.008)
• system reliability
0.93 (1.40) 1.07 (1.35) 1.82 (0.91) 0.49 (0.48) 19.11 (0.00)
Process
• cognitive effort 0.42 (1.48) 0.92 (1.24) 0.61 (1.20) 4.84 (0.029) 2.62 (0.11)
6.2.2 Impact of process
Further analysis was performed to determine the influence of the process, in view of only cognitive
effort, on the outcome variables. We ran exploratory regression analyses with cognitive
effort as the antecedent on market and behavioral outcome, as well as IS perceptions. The results
are summarized in Table 5.
In general, the cognitive effort reported does not affect the participant’s revenue, satisfaction
with outcome or system reliability. But all e‐market systems led users to express that cognitive
effort influences their perception of information quality and reliability of the system. When
using a negotiation mechanism, users describe that less effort influences greater satisfaction
with process. In instances of Invite, less effort allows for higher perception of usefulness.
Table 5. Regression of process (cognitive effort) on outcome variables
Imbins InAuction meet2trade
System Invite Invite meet2trade
Mechanism Negotiation Auction Auction
βˆ
F
(p‐value)
2
R
βˆ
F
(p‐value)
2
R
βˆ
F
(p‐value)
2
R
274

32
Market outcome
• revenue ns ns ns
Behavioral outcomes
• satisfaction
with outcome
• satisfaction
0.53 22.91
(0.000)
with process
IS perceptions
• usefulness 0.50 19.31
(0.000)
ns ns ns
0.28 ns ns
0.25 0.31 8.71 0.10 ns
(0.004)
• information
quality
0.45 14.71
(0.000)
0.20 0.31 8.63
(0.004)
0.10 0.33 9.78
(0.002)
• ease of use 0.41 12.00 0.17 0.43 18.30 0.18 0.43 17.97
(0.001)
(0.000)
(0.000)
• system reliability
ns ns ns
0.33
0.19
6.3 Summary
The mechanism and system comparisons revealed that economic revenue acquired by bidders
is affected by the difference in system rather than mechanism. However users report contrary
results, in that they are more satisfied with the outcome and process of an auction versus a
negotiation. IS perception is rightfully unaffected by the mechanism, but a faster system with
one screen (meet2trade) impacts perceived ease of use and system reliability. Adding a history
graph (InAuction) has not enabled greater perception of usefulness or information quality.
The process, in terms of reported cognitive effort, influences perceived information quality
and ease of use for all e‐market systems, because these two construct describe what is need on
how the participants interact with the system.
7. Discussion and future work
Electronic commerce has led to the emergence of the new forms of electronic exchange
mechanisms driven by the convergence of economic and technological factors. The complexities
of these formations preclude one‐sided attempts to meaningfully study them from either
side. Instead, a comprehensive approach to investigate the components and workings and
guide the design of electronic markets demand an adoption of the integrated technoeconomical
frameworks of reference. The purpose of this paper has been to propose a model
for studying the impacts of electronic exchange mechanisms on key variables of interests,
both objective, as well as subjective ones. To this end we had reviewed and incorporated relevant
concepts from both fields of Economics and Information Systems.
The proposed TIMES model provides a framework that allows studying of types of exchange
mechanisms in their various implementations within different task, environment, and individual
contexts. These mechanisms could range from the simplest catalogue‐based models to
advanced auction and negotiation schemas. Thus, the model can accommodate continuity in
the key design principles of the mechanisms, as opposed to considering them as distinct
275

33
classes. Therefore, one of the key contributions of the model is that it enables the comparison
of various exchange structures in terms of the same set of key dependent factors.
One research project that will utilize TIMES to compare particular forms of auctions vs. negotiations
has been briefly discussed earlier (NorA). This project will provide empirical data for
testing the theoretical model, and provide key insights into the implications of design of market
mechanisms for different classes of tasks. In particular, the key objective process and performance
characteristics, as well as subjective perceptions and evaluations induced by adopting
auctions vs. negotiations for specific tasks will be revealed. Comparisons of other relevant
types of mechanisms within the TIMES framework will become the core part of future empirical
efforts.
While the study of electronic exchange mechanisms has been the primary motivation for developing
the TIMES model, we believe it is not limited to studying information systems for
conducting market transactions only. It can be used to study also other information systems
for which the issues of their ease of use, performance and usefulness are of interest. In this
respect, the inclusion of the abstract representation of the underlying “mechanism” in addition
to the concrete implementation‐specific features would enable studying broad classes of
systems. The proposed model thus could be potentially extended to become a powerful tool
for the design research community, as the latter is focusing on developing innovative classes
of systems. This extension of the TIMES model would however require to expand the notion of
the mechanism towards the task. The emphasis of the mechanism (and algorithms) embedded
in the system fundamentally affect the perceived usefulness of the system.
References
Al-Khaldi, M. A. (1999). "The Influence of Attitudes on Personal Computer Utilization among
Knowledge Workers, The Case of Saudi Arabia." Information and Management 36(4): 185-204.
Alter, S. (2003). "18 Reasons Why IT-Reliant Work Systems Should Replace “The IT Artifact” as the
Core Subject Matter of the IS Field." Communications of the Association for Information Systems
12(23): 366-395.
Bajari, P., R. McMillan, et al. (2002). "Auctions versus Negotiations in Procurement: An empirical
analysis." Working Paper.
Baroudi, J. and W. Orlikowski (1988). "A Short-form Measure of User Information Satisfaction: A
Psychometric Evaluation and Notes on Use." Journal of Management Information Systems 4(4): 44–
59.
Beckert, J. (2003). "Economic Sociology and Embeddedness: How shall we conceptualize Economic
Action?" Journal of Economic Issues 27(3): 769-787.
Beil, D. R. and L. Wein (2003). "An inverse-optimization-based auction mechanisms to support a
multiattribute RFQ process." Management Science 49(11): 1529-1545.
Benbasat, I. and R. Weber (1996). "Research commentary: Rethinking "diversity" in information systems
research." Information Systems Research 7(4): 389-399.
276

34
Benbasat, I. and R. W. Zmud (1999). "Empirical Research in Information Systems: The Practice of
Relevance." MIS Quarterly 23(1): 3-16.
Benbasat, I. and R. W. Zmud (2003). "The Identity Crisis within the IS Discipline: Defining and
Communicating the Discipline's Core Properties." MIS Quarterly 27(2): 183-194.
Bhargava, H. K. and S. Sundaresan (2004). "Computing as Utility: Managing Availability, Commitment,
and Pricing through Contingent Bid Auctions." Journal of Management Information Systems
21(2): 201-227.
Bichler, M. (2000). "An Experimental Analysis of Multi-attribute Auctions." Decision Support Systems
29(3): 249-268.
Bichler, M., G. E. Kersten, et al. (2003). "Towards a Structured Design of Electronic Negotiations."
Group Decision and Negotiation 12(4): 311-335.
Branco, F. (1997). "The Design of Multidimensional Auctions." Rand Journal of Economics 28(1):
63-81.
Brynjolfsson, E. and S. Yang (1996). "Information Technology and Productivity: A Review of the
Literature." Advances in Computers 43(1): 179-214.
Bulow, J. and P. Klemperer (1996). "Auctions versus Negotiations." American Economic Review
86(1): 80-194.
Che, Y.-K. (1993). "Design Competition through multidimensional auctions." RAND Journal of Economics
24(4): 668-680.
Chen-Ritzo, C.-H., T. P. Harrison, et al. (2005). "Better, Faster, Cheaper: An Experimental Analysis
of a Multi-attribute Reverse Auction Mechanism with Restricted Information Feedback." Management
Science 51(12): 1753-1762.
Davis, F. D. (1989). "Perceived Usefulness, Perceived Ease of Use and User Acceptance of Information
Technology." MIS Quarterly 13(3): 319-340.
Davis, F. D., R. P. Bagozzi, et al. (1989). "User Acceptance of Information Technology: A Comparison
of Two Theoretical Models." Management Science 35(8): 982-1003.
De Fraja, G. and J. Sakovics (2001). "Walras Retrouve: Decentralized Trading Mechanisms and the
Competitive Price." Journal of Political Economy 109(4): 842-863.
Delaney, M. M., A. Foroughi, et al. (1997). "An Empirical Study of the Efficacy of a Computerized
Negotiaiton Support System." Decision Support Systems 20: 185-197.
DeLone, W. and E. McLean (1992). "Information Systems Success: The Quest for the Dependent
Variable." Information Systems Research 3(1): 60-95.
DeLone, W. H. and E. R. McLean (2003). "The DeLone and McLean Model of Information Systems
Success: A Ten Years Update." Journal of Management Information Systems 19(4): 9-30.
Dishaw, M. T. and D. M. Strong (1996). Explaining Information Technology Utilization with the
Task-technology Fit Conctructs. First INFORMS Conference on Information Systems and Technol-
277

35
ogy, Providence, RI, INFORMS.
Dishaw, M. T. and D. M. Strong (1998). "Assessing Software Maintenance Tool Utilization using
Task–technology Fit and Fitness-for-use Models." Journal of Software Maintenance: Research and
Practice 10(3): 151-179.
Doll, W. J., A. Hendrickson, et al. (1998). "Using Davis's Perceived Usefulness and Ease-of-use Instruments
for Decision Making: A Confirmatory and Multigroup Invariance Analysis." Decision Sciences
29(4): 839-869.
Doll, W. J. and G. Torkzadeh (1988). "The Measurement of End User Computing Satisfaction Issues."
MIS Quarterly 12: 259-274.
Doll, W. J. and G. Torkzadeh (1991). "Issues and Options. The Measurement of End User Computing
Satisfaction: Theoretical and Methodological Issues." MIS Quarterly 15: 5-10.
Friedman, D. and S. Sunder (1994). Experimental Methods: A Primer for Economists. Cambridge,
Cambridge University Press.
Galinsky, A. D., T. Mussweiler, et al. (2002). "Disconnecting Outcomes and Evaluations: The Role of
Negotiator Focus." Journal of Personality and Social Psychology 83(5): 1131–1140.
Garrity, E. J., B. Glassberg, et al. (2005). "An Experimental Investigation of Web-based Information
Systems Success in the Context of Electronic Commerce." Decision Support Systems 39(3): 485-503.
Gebauer, J. and M. Ginsburg (2006). Exploring the Black Box of Task-Technology Fit: The Case of
Mobile Information Systems, University of Illinois at Urbana-Champaign.
Glass, R. L., V. Ramesh, et al. (2004). "An Analysis of Research in Computing Disciplines." Communications
of the ACM 47(6): 89-94.
Goodhue, D. L. (1995). "Understanding User Evaluation of Information Systems." Management Science
41: 1827-1844.
Goodhue, D. L. and R. L. Thompson (1995). "Task-Technology Fit and Individual Performance."
MIS Quarterly 19(2): 213-236.
Hayek, F. (1945). "The Use of Knowledge in Society." American Economic Review 35(4): 519-530.
Hevner, A. R., S. T. March, et al. (2004). "Design Science in Information Systems Research." MIS
Quarterly 28(1): 75-105.
Hurwicz, L. (1973). "The Design of Mechanisms for Resource Allocation." American Economic Review
63(2): 1-30.
Hyder, E. B., M. J. Prietula, et al. (2000). "Getting to Best: Efficiency vs. Optimality in Negotiation."
Cognitive Science 24(2): 169-204.
Iivari, J. (2003). "The IS Core - VII: Towards Information Systems as a Science of Meta-Artifacts."
Communications of the Association for Information Systems 12(37): 568-581.
Iivari, J. (2005). "An Empirical Test of the DeLone-McLean Model of Information System Success."
278

36
The DATA BASE for Advances in Information Systems 36(2): 8-27.
InterNeg (2006). Negotiations or Auctions. Montreal, InterNeg Research Centre. 2006.
ISR (2002). "Editorial Statement and Policy." Information Systems Research 13(4).
Kim, H. W. and S. L. Pan (2006). "Towards a Process Model of Information Systems Implementation:
The Case of Customer Relationship Management (CRM)." ACM SIGMIS Database 37(1): 59-
76.
Kirkegaard, R. (2004). Auctions versus Negotiations Revisited. Working Paper. Aarhus, Department
of Economics, University of Aarhus.
Kolitz, K. and D. Neumann (2007). The Impact of Information System Design on E-marketplaces. An
Experimental Analysis. Group Decision and Negotiation 2007, Mt. Tremblant, InterNeg.
Koppius, O. and E. van Heck (2002). Information Architecture and Electronic Market Performance in
Multidimensional Auctions. Erasmus Research Institute of Management. Rotterdam, Erasmus University,
Rotterdam: 38.
Kumar, M. and S.I.F. (1999). Internet Auctions. IBM Research Division, T.J. Watson Research Center.
Ledyard, J. (1993). "The Design of Coordination Mechanisms and Organizational Computing." Journal
of Organizational Computing 3(1): 121-134.
Legris, P., J. Ingham, et al. (2003). "Why Do People Use Information Technology? A Critical Review
of the Technology Acceptance Model." Information and Management 40: 191-204.
Leskela, R. L., J. Teich, et al. (2006). Decision Support for Multi-Unit Combinatorial Bundle Auctions.
Helsinki, Helsinki University of Technology.
Lilien, G. L., A. Rangaswamy, et al. (2004). "DSS effectiveness in marketing resource allocation decisions:
reality versus perception." Information Systems Research 15(3): 216-235.
Lucking-Reiley, D. (2000). "Auctions on the Internet: What's Being Auctioned, and How?" Journal of
Industrial Economics 48(3): 227-252.
March, S. T. and G. F. Smith (1995). "Design and Natural Science Research on Information Technology."
Decision Support Systems 15: 251-266.
Markus, M. L., A. Majchrzak, et al. (2002). "A Design Theory for Systems that Support Emergent
Knowledge Processes." MIS Quarterly 26(3): 179-212.
Maskin, E. and T. Sjöström (2002). Implementation Theory. Handbook of Social Choice and Welfare.
K. J. Arrow, A. Sen and K. Suzumura. Amsterdam, NL, Elsevier Science B.V. 1: 237-288.
McAfee, R. P. and J. McMillan (1987). "Auctions and Bidding." Journal of Economic Literature
25(2): 699-738.
Mount, K. and S. Reiter (1974). "The Informational Size of Message Spaces." Journal of Economic
Theory 8(1): 161-192.
279

37
Nash, J. F. (1954). "The Bargaining Problem." Econometrica 18: 155-162.
Neumann, D., M. Benyoucef, et al. (2003). "Applying the Montreal Taxonomy to State of the Art E-
negotiation Systems." Group Decision and Negotiation 12(4): 287-310.
Nielsen, J. (1994). Usability Engineering, Morgan Kaufmann.
Nunamaker, J. F. J., M. Chen, et al. (1991). "Systems Development in Information Systems Research."
Journal of Management Information Systems 7(3): 89 - 106.
Orlikowski, W. J. and C. S. Lacono (2001). "Research Commentary: Desperately Seeking the 'IT' in
IT Research - A Call to Theorizing the IT Artifact." Information Systems Research 12(2): 121-134.
Palfrey, T. (2001). Implementation Theory. Handbook of Game Theory. R. J. Aumann and S. Hart.
Amsterdam, North-Holland. 3.
Parkes, D. C. (2001). Iterative Combinatorial Auctions:
Achieving Economic and Computational Efficiency. Department of Computer and Information Science,
University of Pennsylvania.
Pikkarainen, K., T. Pikkarainen, et al. (2006). "The Measuremnet of End-user Computing Satisfaction
of online Banking Services: Empirical Evidence from Finland." International Journal of Banki Marketing
24(3): 158-172.
Rangaswamy, A. and K. Starke (2000). Computer-Mediated Neotiations: Review and Research opportunities.
Encyclopedia of Microcomputers. A. Kent and J. G. Williams. New York, NY, Marcel
Dekker. 25: 47-72.
Reiter, S. (1977). "Information and Performance in the (New)² Welfare Economics." American Economic
Review 67(1): 226-234.
Robey, D. and M. Newman (1996). "Sequential Patterns in Information Systems Development: An
Application of a Social Process Model." ACM Transactions on Information Systems 14(1): 30-63.
Roth, A. (2002). "The Economist as Engineer: Game Theory, Experimentation, and Computation as
Tools for Design Economics." Econometrica 70(4): 1341-1378.
Roth, A. E. (1995). Introduction to Experimental Economics. The Handbook of Experimental Economics.
J. H. Kagel and A. E. Roth. Princeton, NJ, Princeton University Press: 3-110.
Rothkopf, M. H., A. Pekec, et al. (1998). "Computationally Manageable Combinatorial Auctions."
Management Science 44(8): 1131-1147.
Satterthwaite, M. and A. Shneyerov (2003). "Convergence of a Dynamic Matching and Bargaining
Market with Two-sided Incomplete Information to Perfect Competition." Working Paper.
Satterthwaite, M. and A. Shneyerov (2006). "Dynamic Matching, Two-sided Incomplete Information,
and Participation Costs: Existence and Convergence to Perfect Competition." Econometrica: in print.
Seddon, P. B., D. S. Staples, et al. (1999). "Dimensions of IS Success." Communications of the AIS
2(20): 1-61.
280

38
Segev, A. and C. Beam (1999). A New Market-based Negotiation Paradigm.
Simon, H. (1996). The Sciences of the Artificial. Cambridge, Massachusetts, The MIT Press.
Simon, H. A. (1986). "Rationality in Psychology and Economics." The Journal of Business 59(4):
209-224.
Smith, V. (1982). "Microeconomic Systems as an Experimental Science." American Economic Review
72(5): 923-955.
Smith, V. (1994). "Economics in the Laboratory." Journal of Economic Perspectives 8(1): 113-131.
Smith, V. (2003). Markets, Institutions and Experiments. Encyclopedia of Cognitive Science. L.
Nadel. London, Nature Publishing Group. 2: 991-998.
Strecker, S. (2004). Multiattribute Auctions in Electronic Procurement - Theory and Experiment -.
Economics and Business Engineering. Karlsruhe, Germany, University of Karlsruhe (TH).
Strecker, S., G. E. Kersten, et al. (2006). Electronic Negotiation Systems: The Invite Prototype. Collaborative
Business MKWI’06, Passau, Germany.
Ströbel, M. and C. Weinhardt (2003). "The Montreal Taxonomy for Electronic Negotiations." Group
Decision and Negotiation 12(2): 143-164.
Subrahmanian, E. and S. N. Talukdar (2004). "Engineering of Markets and Artifacts." Electronic
Commerce: Research and Applications 3(4): 369-380.
Suh, K. S. (1999). "Impact of Communication Medium on Task Performance and Satisfaction: An
Examination of Media-Richness Theory." Information and Management 35: 295-312.
Szajna, B. (1996). "Empirical Evaluation of the Revised Technology Acceptance Model." Management
Science 42(1): 85-92.
Teich, J., H. Wallenius, et al. (2001). "Designing Electronic Auctions: An Internet-Based Hybrid Procedure
Combining Aspects of Negotiations and Auctions." Journal of Electronic Commerce Research
1: 301-314.
Thomas, C. J. and B. J. Wilson (2002). "A Comparison of Auctions and Multilateral Negotiations."
RAND Journal of Economics 33(1): 140-155.
Thomas, C. J. and B. J. Wilson (2005). "Verifiable Offers and the Relationship Between Auctions and
Multilateral Negotiations." The Economic Journal 115 (506): 1016–1031.
Torkzadeh, G. and W. J. Doll (1999). "The Development of a Tool for Measuring the Perceived Impact
of Information Technology on Work." Omega 27(3): 327-339.
Turner, M., D. Budgen, et al. (2003). "Turning Software into a Service." Computer 36(10): 38-44.
Vahidov, R. (2006). Design Researcher's IS Artifact: A Representational Framework. First International
Conference on Design Science Research in Information Systems and Technology (CD-ROM
Proceedings), Claremont, CA.
281

39
van de Ven, A. H. (1992). "Suggestions for Studying Strategy Process: A Research Note." Strategic
Management Journal 13: 169-191.
Venkatesh, V. and F. D. Davis (2000). "A Theoretical Extension of the Technology Acceptance
model: Four Longitudinal Field Studies." Management Science 46(2): 186-204.
Venkatesh, V., M. G. Morris, et al. (2003). "User Acceptance of Information Technology: Toward A
Unified View." MIS Quarterly 27(3): 425-478.
Walls, J. G., G. R. Widmeyer, et al. (1992). "Building an Information System Design Theory for
Vigilant EIS." Information Systems Research 3(1): 36-59.
Weber, M., G. E. Kersten, et al. (2006). "An Inspire ENS Graph is Worth 334 Words, on Average."
Electronic Markets 16(3): 186-200.
Wilson, D. D. (1995). "IT Investment and Its Productivity Effects: An Organizational Sociologist's
Perspective on Directions for Future Research." Economics of Innovation and New Technology 3(3-
4): 235-251.
Wixom, B. H. and P. A. Todd (2005). "A Theoretical Integration of User Satisfaction and Technology
Acceptance." Information Systems Research 16(1): 85–102.
Wolfstetter, E. (1995). "Auctions: An Introduction." Journal of Economic Surveys 10(4): 367-420.
Wolfstetter, E. (2000). Topics In Microeconomics. Industrial Organization, Auctions, and Incentives.
Cambridge, Cambridge University Press.
Yi, Y. (1990). A Critical Review of Consumer Satisfaction. Review of Marketing. V. A. Zeithaml.
Chicago, American Marketing Association: 68-129.
Yu, B. (2007). Negotiations or Auctions: Experimental Comparison of Two E-market Mechanisms. J.
Molson School of Business. Montreal, ConcordiaUniversity: 138.
Zhu, K. (2004). "The Complementarity of Information Technology Infrastructure and E-Commerce
Capability: A Resource-Based Assessment of Their Business Value." Journal of Management Information
Systems 21(1): 167 - 202.
Zwass, V. (1996). "Electronic Commerce: Structures and Issues." International Journal of Electronic
Commerce 1(2): 3 - 23.
282

40
Appendix
Information
quality
System reliability
Usefulness
Ease of use
Satisfaction
with outcome
Cognitive effort
Satisfaction
with process
Table A. Factor analysis of subjective variables 1
SR U IQ SO EU CE SP
IQ1 .162 .337 .791 .134 .009 .128 .057
IQ2 .204 .263 .836 .135 .079 .100 .136
IQ3 .267 .039 .649 .029 .377 .104 .175
IQ4 .230 .194 .762 .113 .275 .113 .117
SR1 .742 .134 .208 .094 .267 ‐.006 .186
SR2 .794 .187 .207 .027 .211 ‐.032 .128
SR3 .801 .176 .161 .109 .059 .104 .036
SR4 .812 .220 .143 .081 .084 .097 ‐.023
U1 .236 .764 .252 .101 .124 .178 .185
U2 .178 .834 .210 .203 .156 .125 .109
U3 .302 .750 .137 .176 .233 .114 .098
U4 .172 .733 .269 .114 .189 .083 .238
EU1 .225 .270 .157 .067 .776 .163 .150
EU2 .321 .276 .088 .046 .722 .135 .159
EU3 .084 .086 .215 .071 .839 .131 .049
SO1 .123 .102 .088 .889 .043 .111 .136
SO2 .066 .176 .113 .859 .050 .122 .120
SO3 .064 .123 .092 .872 .067 .072 .074
CE1 ‐.056 .100 .095 .137 .150 .855 .058
CE2 .066 .084 .241 .163 .015 .797 .165
CE3 .192 .217 ‐.007 .019 .266 .729 .225
SP1 .013 .109 .106 ‐.028 .310 .078 .698
SP2 .111 .237 .147 .321 ‐.057 .165 .710
SP3 .142 .149 .118 .175 .076 .197 .753
Reliability
(Cronbach
alpha)
0.879
0.874
0.895
0.858
0.896
0.708
0.809
1 Factor analysis was rotated by Varimax method with Kaiser normalization.
All seven factors have eigenvalues above 1.
283

Comparing Ingress and Egress Detection to Secure
Inter-domain Routing: An Experimental Analysis
Christoph Goebel
Institute of Information Systems
Humboldt Universitaet zu Berlin
10178 Berlin
Germany
Dirk Neumann
Institute of Information Systems and Management
University of Karlsruhe
76131 Karlsruhe
Germany
Ramayya Krishnan
H. John Heinz III School of Public Policy and Management
Carnegie Mellon University
Pittsburgh, PA 15213-3890
USA
July 18, 2007
Abstract
The global economy and society is increasingly dependent on computer networks linked
together by the Internet. The importance of networks reaches far beyond the telecommunications
sector since they have become a critical factor for many other crucial infrastructures
and markets. With threats mounting and security incidents becoming more frequent, concerns
about network security grow.
It is an acknowledged fact that some of the most fundamental network protocols that make
the Internet work are exposed to serious threats. One of them is the Border Gateway Protocol
(BGP) which determines how Internet traffic is routed through the topology of administratively
independent networks the Internet is comprised of. Despite the existence of a steadily
growing number of BGP security proposals, to date neither of them is even close to being
adopted.
The purpose of this work is to contemplate BGP security from a theoretical point of view
in order to take a first step toward understanding the factors that complicate secure BGP
adoption. Using a definition of BGP robustness we experimentally show that the degree of
robustness is distributed unequally across the administrative domains of the Internet, the
so-called Autonomous Systems (ASs). The experiments confirm the intuition that the contribution
ASs are able to make towards securing the correct working of the inter-domain
routing infrastructure by deploying countermeasures against routing attacks differ depending
on their position in the AS topology. We also show that the degree of this asymmetry
can be controlled by the choice of the security strategy. We compare the strengths and
weaknesses of two fundamentally different approaches in increasing the BGP’s robustness
which we termed ingress and egress detection of false route advertisements and indicate
their economic implications.
1 284

1 Introduction
The global society and economy is increasingly dependent on the effective working of communication
networks. These networks have created connectivity linking together millions of users
including private users, businesses and public institutions. They also have become a critical
factor for infrastructures (e.g. water and electricity supply) and markets other than the telecommunications
market (e.g. the worldwide financial market) [2, 23]. Consequently, concerns about
the security of these global networks and information systems have been growing along with the
rapid increase in the number of users and the importance of the information transmitted.
Although targeted attacks on the confidentiality of information assets and the availability of particular
services have been at the center of public attention so far, the Internet infrastructure as
a whole suffers from more fundamental vulnerabilities due to its design. Security considerations
did not play a major role when the fundamental protocols were devised, which still orchestrate
global data flows in today’s Internet.
Ironically, the same mechanisms that have contributed substantially to the success of many
Internet protocols have severe drawbacks when it comes to securing them against malicious
activities. In the absence of a central authority, the different components of the Internet infrastructure
rely on distributed control protocols. This cooperative control effort can only lead
to the desired outcome if every control unit behaves in the way it is supposed to, i.e. behave
according to the protocol it implements. Consequently, the misbehavior of one control unit can
have adverse effects on the way the infrastructure works as a whole [16].
One of the key Internet protocols suffering from a fundamental lack of security is the Border
Gateway Protocol (BGP). The BGP is the core routing protocol of the Internet. It enables the
exchange of reachability information based on the Internet Protocol (IP) across administratively
independent networks, the so-called Autonomous Systems (ASs).
In this work we address the vulnerability of the BGP to routing attacks. This is a particularly
virulent type of attack targeted at the Internet infrastructure as a whole. The contributions of
this paper are threefold. Firstly, a new countermeasure approach, egress detection, is motivated.
Secondly, a simulation design is proposed that has the potential to gauge the effectiveness of
alternative countermeasures. Thirdly, a rigorous analysis showing the properties of egress detection
and comparing it with state-of-the-art ingress detection approaches. This research is
important, as it supports the design of a viable strategy to secure the global routing architecture.
This work is original as we adopt an abstract approach that allows us to compare the efficiency
of different strategies on a general level. Our approach simplifies the problem of determining the
degree of routing security by not making too detailed assumptions. Closely related earlier work
by Chan et. al [4] on the topic attempts to compare details in the protocol, which is hampered
by the fact that the introduction of a large number of additional assumptions increases the
complexity of the model and therefore both validation and interpretation of simulation results.
This work is conceived to be a primer for better understanding the dominant utility effects
underlying the adoption process, which will fuel analyses on the protocol level. Section 2 gives
an overview of related work. In section 3 we provide an overview of inter-domain routing, routing
attacks and countermeasures. In section 4 we present a basic approach to measure routing
security based on previous work by Chan et al. [4]. This measurement methodology is applied
to artificial Internet topologies in section 5 in order to conduct an extensive sensitivity analysis
both revealing fundamental coherences and comparing the effects of two different approaches
to secure inter-domain routing (ingress and egress detection) on two different security metrics
(one based on robustness and the other one based on semi-robustness). In sections 6 and 7 we
summarize our results, conclude and provide an outlook for future research.
2 285

2 Related Work
An impressive amount of work has been devoted to the specification of different BGP security
protocols recently. S-BGP, the first and most comprehensive proposal was made by Kent et
al. [13]. It uses several Public Key Infrastructures (PKIs) to enable BGP routers to authenticate
outgoing route announcements and validate incoming ones. Inter-domain hops are signed
recursively as paths are constructed in the forwarding process described in section 3. Since
S-BGP implies a significant protocol overhead, other authors tried to develop less expensive
solutions. For instance, White et al. [24] propose a centralized heuristic solution to path validation
that has the potential to reduce complexity and overhead significantly but also leads to less
accurate detection of falsified routes. Unlike S-BGP, which uses a push-approach to distribute
authentication information, IRV allows for authentication on-demand using a dedicated server
infrastructure (pull-approach) [8]. Hu et al. [10] have proposed SPV which reduces security overhead
by substituting expensive but concise public key cryptography by less expensive but more
complicated procedures based on symmetrical primitives and one-time signatures. Additional
work on inter-domain routing security includes research on ways how to make a security protocol
proposed earlier more efficient [17, 12], the specification of less comprehensive approaches
to BGP security [21, 26], and the advancement of causal research and attack detection methods
[27, 14].
Chan at al. [4] have raised the issue of adoptability of secure BGP protocols. They provide a
taxonomy of properties in order to classify the numerous secure BGP schemes that have been
proposed so far. Based on these properties they model the adoption process and compare the
adoption thresholds of the different protocols under numerous assumptions, e.g. regarding attacker
capability and traffic load. Although our approach to measure the degree of security of
AS paths follows the lines laid out by Chan et al., the level of analysis differs. While Chan et al.
dive into the analysis of different protocol classes based on detailed protocol specifications, we
follow a more general way exploring the pros and cons of two different security strategies. The
reduction of protocol details on the one hand and the abstraction from the security protocol
level to the security strategy level on the other hand allow us to make more general statements.
3 Background
3.1 Inter-domain Routing and the BGP
The Internet is a network of networks. It is currently comprised of more than 20,000 interconnected
computer networks, the so-called Autonomous Systems (ASs). We will refer to this network
of ASs as the AS topology. The physical coverage of an AS can vary substantially: Large
backbone providers like AT&T own physical networks that have Points of Presence (PoPs) 1
around the globe, whereas a university network is usually confined to some campus area.
The AS topology is a scale-free topology, i.e. a relatively small subset of the ASs, the transit
ASs, forward the traffic originated elsewhere, while most ASs, the stub ASs, serve as mere traffic
end-points. Physically, the stub ASs depend on the transit ASs to carry their traffic over long
distances.
Inter-domain routing refers to the path selection process on the AS level. The outcome of this
process is a complete prescription of how IP traffic is forwarded through the AS topology. The
information required to compute this global routing table is communicated by dedicated routers
implementing an inter-domain routing protocol. The de-facto standard protocol used for interdomain
routing is the Border Gateway Protocol (BGP) [20]. The routers which implement the
1 A Point of Presence (PoP) is a place where an AS provides the required physical infrastructure to establish
connections with other ASs.
3 286

1.1.1.0/24
r51
AS5
1.1.1.0/24 r51 AS5
AS1
r11
1.1.1.0/24 r41 AS4_AS5
AS4
r41
r42
1.1.1.0/24 r21 AS2_AS4_AS5
r21
r22
1.1.1.0/24 r42 AS4_AS5
AS2
r23
1.1.1.0/24 r31 AS3_AS4_AS5
r31
AS3
Figure 1: Propagation of route advertisements according to the BGP.
BGP are oftentimes denoted as BGP speakers.
The BGP belongs to the family of path vector protocols, i.e. the basic information exchanged
by the BGP speakers is comprised of (destination,path)-tuples. 2 The destinations are ranges
of IP addresses (oftentimes denoted as prefixes) and the paths are concatenations of AS identifiers.
IP ranges and AS numbers are allocated by the Internet Corporation for Assigned Names
and Numbers (ICANN) or the allocation task is delegated to local Internet registries. If an AS
originates a certain prefix, this information is passed on to the BGP speakers in adjacent ASs
and so forth. Consequently, each BGP speaker eventually receives a number of routes leading to
the same address space. BGP speakers performs a standardized path selection process in order
to determine which of these paths will be added to their local routing table. When the BGP
has reached a steady state, each router knows for any IP address which adjacent router an IP
packet needs to be forwarded to reach its destination. It is important to note that currently
only selected paths are forwarded to other BGP speakers, not all of them.
The BGP path selection process takes a large number of criteria into consideration. A comprehensive
description of the path selection process can be found in [20]. One of the most important
criteria is path length (in terms of AS hops). The shorter the path, the higher the probability
that it is chosen by a router.
Example 1: Propagation of Route Advertisements
If nothing but the shortest path criterion were used in the path selection process,
the prefix 1.1.1.0/24 in figure 1 would be propagated in the following way:
The prefix 1.1.1.0/24 is owned and originated by AS5. Accordingly, router r51 announces
this information to router r42 in the adjacent AS4. AS4 is connected to
AS2 and AS3 and forwards the advertisement over the BGP sessions between the
routers r42 and r31 and the routers r41 and r22. This process continues until all
2 The current specification of the BGP encompasses many more attributes which we cannot describe here due
to space constraints.
4 287

outers are aware of the fact that AS5 originates prefix 1.1.1.0/24 and a corresponding
inter-domain path. Each time the route advertisement crosses an AS border,
an AS identifier is appended to the path. Router r21 receives two different paths
leading up to the same prefix. Considering the shortest path criterion, it will prefer
the path AS4 AS5 over the path AS3 AS4 AS5. Only the announcement containing
the shorter path is then forwarded to router r11 in AS1.
After the routing tables maintained by the routers have reached a converged state,
all traffic from AS1 to AS5 will travel on the path leading through AS2 and AS4,
which is emphasized in figure 1.
3.2 BGP Routing Attacks
It is widely known that the BGP is vulnerable to a whole range of possible attacks [16]. When
considering attacks on a complex system like the inter-domain routing infrastructure, it is advisable
to put oneself into the position of an attacker and explore possible ways how to achieve
certain attack goals. Using the scope of an attack’s impact as a means of differentiation, we
distinguish two general types of attacks on the BGP: Attacks on single BGP sessions (small scale
attacks) and attacks intended to affect the global routing infrastructure as a whole (large scale
attacks) [3]. Particular sessions may be attacked in order to terminate them or to eavesdrop on
the control messages exchanged. Carrying out large scale attacks on the routing infrastructure
may, for instance, be conducted with the intention to alter the global routing table, i.e. in order
to change the paths inter-domain traffic travels on. We denote this kind of attack a routing
attack. While the ultimate goal of routing attacks can be manifold, their consequences mostly
include disruption of reachability, so that effective traffic delivery is prevented [3].
Both small and large scale attacks can only be carried out if the attacker finds a way to gain control
over BGP speakers or to at least tamper with BGP control messages on their way from one
BGP speaker to another. Thus, BGP security is inherently linked to classical security problems
concerning host and network security. If we assume that attackers are always in the position to
control one or several BGP speakers, providing BGP security appears to be a question of achieving
stronger protocol robustness. Generally, a distributed protocol is robust, if the misbehavior
of one or several of the entities do not have any negative impact on the outcome.
Taking into account the distributed nature of the Internet infrastructure, router and session
security can only be considered as a first line of defense. This is because as long as humans
are involved, computer and network security can never be fully guaranteed, no matter how sophisticated
the security technologies and policies are. Security technology should at least aim
at putting the right ’speed bumps’ into place, such that the velocity and impact of electronic
attacks are decreased to a level where other protection mechanisms can operate [18]. 3
In terms of security, inter-domain routing exhibits weakest link properties: To date, if an attacker
succeeds in compromising the BGP speakers of one AS, many other ASs may be hurt by
a routing attack. Therefore, even the deployment of host and network security on the Internet
backbone can only reduce the probability of a threat occurance, it cannot effectively delimit
the extent of the damage caused in case the threat is realized. As a consequence, securing the
inter-domain routing infrastructure requires on the one hand, enhancing the security of BGP
sessions and routers and on the other hand making the protocol itself more robust.
In this paper, we concentrate on two important types of routing attacks, namely prefix hijacking
and path spoofing. Prefix hijacking takes advantage of the fact that current BGP implementations
do not prevent a compromised or miss-configured router from illegitimately announcing
prefixes it does not originate. The router in adjacent ASs are not able to distinguish the real
3 Intentional misbehavior (i.e. attacks) is only one part of the BGP security problem. In fact, the incidents that
have been recorded so far were due to unintentional misconfiguration of BGP speakers [15]. The configuration of
BGP speakers is error-prone, since it is still done manually to a significant extent. Unintentional misconfiguration
cannot be prevented by more rigid access controls or the protection of BGP sessions.
5 288

1.1.1.0/24
r51
AS5
1.1.1.0/24 r51 AS5
AS1
r11
1.1.1.0/24 r41 AS4_AS5
AS4
r41
r42
legitimate advertisement:
1.1.1.0/24 r21 AS2_AS4_AS5
illegitimate advertisement:
r21
r22
1.1.1.0/24 r42 AS4_AS5
1.1.1.0/24 r21 AS2_AS3
AS2
r23
hijacked prefix
legitimate advertisement:
1.1.1.0/24 r31 AS3_AS4_AS5
illegitimate advertisement:
1.1.1.0/24 r31 AS3
r31
1.1.1.0/24
AS3
Figure 2: Prefix Hijacking
from the fake owner, leaving a significant probability that the fake advertisement is selected and
subsequently forwarded. Since AS path length is one of the most important criteria in the best
path selection algorithm, the fake announcements are more likely to be adopted by routers in
ASs which are further away from the legitimate originator of the prefix. 4 Let us illustrate the
working of prefix hijacking.
Example 2: Prefix Hijacking
Using the same topology as in example 1, suppose that router r31 in AS3 has been
compromised and is used for a prefix hijacking attack (figure 2). It injects a route
announcement claiming that AS3 originates the prefix 1.1.1.0/24. All routers located
in ASs other than AS5 and AS3 not know which AS legitimately originates
the prefix. Router r21 cannot validate these route advertisements and will therefore
be likely to adopt the advertisement containing the shorter path. Consequently, the
traffic that previously used to travel on the inter-domain path leading over AS2 and
AS4 will now be forwarded to AS3. Hence the traffic will never be delivered to the
intended destination.
Path spoofing works in a similar way as prefix hijacking but with the slight difference that not
the originating AS but the remaining ASs present on the AS path are faked. A router can
change route announcements received from other routers by inserting or deleting ASs from the
path attribute. Malicious ASs can take advantage of this possibility forcing traffic to make a
detour.
4 If an AS illegitimately claims to be the origin of some IP address space and the prefix it announces is more
specific than the one announced by the legitimate AS, prefix hijacking is particularly effective. Since BGP performs
longest prefix matching, prefixes that are more de-aggregated, i.e. longer, are preferred. Prefix de-aggregation
forces routers to adopt the fake announcement even if they have received the legitimate announcement on a shorter
path before. We do not consider this particular type of prefix hijacking in this paper.
6 289

Example 3: Path Spoofing
Considering the scenario depicted in figure 2, router r31 could apply path spoofing
by faking a direct connection to AS5. This results in forwarding the advertisement
1.1.1.0/24 r31 AS3 AS5 to router r23. Since r21 would then receive two paths with
equal length leading to the same origin, there is a good chance that this path spoofing
attack has the same effect as the prefix hijacking attack described above.
3.3 BGP Security Countermeasures: Ingress and Egress Detection
As described above, the protection of BGP routers and sessions is an important first step toward
securing the inter-domain routing infrastructure. AS operators have begun to deploy security
mechanisms in their networks, e.g. the BGP TTL Security Hack [7] and the MD5 signature option
[9]. However, the fundamental lack of robustness inherent to the BGP cannot be removed
by these individual security efforts. Effective protection against large scale attacks requires a
concerted security effort by many AS operators [3].
We describe two different approaches to prevent illegitimate route announcements from adversely
affecting inter-domain routing. The first enables BGP speakers to validate the legitimacy of received
route announcements. The second prevents BGP speakers from announcing illegitimate
routes. In the following we will call these two approaches ingress and egress detection, respectively.
3.3.1 Ingress Detection
The need for a comprehensive BGP security architecture has lead to an abundance of proposals
[3]. Almost all of them follow the ingress detection approach, as they enable routers implementing
the same ingress detection protocol to exchange authoritative routing information. This
information can be used by routers to validate incoming route advertisements, e.g. to check if
the origin of an announced route is correct. Comparing two of the most discussed comprehensive
BGP security proposals, S-BGP [13] and IRV [8], we find that both of them basically call for
an additional communication channel between BGP speakers. Although different techniques
are applied to facilitate the transfer of authentication data, the degree of security achieved is
approximately the same [3].
3.3.2 Egress Detection
Hitherto, egress detection has not received much attention as a viable solution to the routing
security problem so far - albeit the general idea ’taking care of one’s own front garden’ is
certainly not new. An implementation of egress detection includes some security mechanism
that prevents or at least complicates the announcement of false routes. For instance, the routes
being announced by an AS could be double-checked by a trusted and secure third party within
the AS before they are released. The trusted instances could in turn be bound to validate all
route announcements with certificates distributed through a PKI as proposed by Aiello et. al [1].
Egress detection has the advantage that no additional communication channels between routers
located in different ASs would be needed. Therefore, issues such as communication overhead
and Internet-wide protocol standardization can be avoided.
An obvious disadvantage of egress detection is that it cannot be used by a subset of ASs to
achieve guaranteed routing security of the connecting inter-domain paths. With ingress detection
deployed in all ASs on a path, none of the legitimate traffic flowing on that path can be diverted
by prefix hijacking or path spoofing. However, this does not imply that egress detection performs
worse with respect to the number of paths that can be secured given a set of adopters. The
protection effort simply cannot be targeted onto particular routes the way it could be if ingress
detection was used as a countermeasure.
7 290

The egress detection approach appears to be particularly attractive regarding the fact that many
false announcements are actually caused by the unintended misconfiguration of routers and not
by malicious behavior [15, 14].
4 Experimental Setup
This section introduces a basic quantitative methodology to measure the security level interdomain
routing. Using this measurement method, we can experimentally show important features
of inter-domain routing security depending on the employed countermeasure strategy,
either ingress or egress detection. We focus on topological characteristics and the dynamics of
selected security metrics depending on the number of secure BGP adopters.
4.1 Preliminaries
Emphasizing the interdependencies of topological features, router configurations and policies,
the whole path selection process needs to be taken into account. As earlier research has shown,
models of inter-domain routing which equate routing and administrative domains do not yield
accurate results. BGP speakers would have to be modeled independently from each other in
order to reconstruct or accurately predict inter-domain paths [5]. Furthermore, the impact of
intra-domain routing on inter-domain paths needs to be explicitly taken into account [19]. 5
Policy routing is another major factor determining how traffic is routed through the AS topology
[22]. The business relationships between AS operators find their expression in the way traffic
engineering is done: The standardized path selection process running on every BGP router is
influenced by ingress and egress filters configured in a way that business objectives are achieved.
One typical routing policy implemented by transit AS operators consists of announcing only
customer routes to peers minimizing their own transit fees [11].
Clearly, the emphasis of this work is not on the prediction of realistic inter-domain paths. As
this would require the inclusion of router-level topologies, intra-domain routing mechanisms and
most importantly as many protocol details as possible into the model. The purpose of our
work is to quantify the effects illegitimate route announcements have on the robustness of interdomain
routes. This study is designed to reveal how effective ingress and egress detection are
in preventing the most straight-forward routing attacks depending on how many ASs support
ingress and egress detection, on their position in the network and on the order of deployment.
With respect to topology detail the model is restricted to the AS level, ignoring the effects
of individual router behavior and intra-domain routing. For simplicity, we make the following
assumptions on the simulation set-up. The path selection process is reduced to a single but
crucial criterion, namely the AS path length.
The routers in each AS are aware of one AS path leading from their AS to all other ASs in the
Internet, i.e. n − 1 paths in total where n is the number of ASs in the Internet. There exist
n(n − 1) AS paths at a time, which form the global routing table.
4.2 Threat Model
Operationalizing our questions, we will set up a basic threat model. In our model attackers
are assumed to have two basic techniques at their disposal, prefix hijacking and path spoofing.
In the absence of any countermeasures deployed in the ASs on the path, the success of either
attack depends on the shape of the topology in the region around that path. The attack can be
5 Intra-domain routing refers to the way IP traffic is routed through the network topology of an AS. Intradomain
routing also has an important influence on the way inter-domain paths are selected because it determines
at which border gateway router an IP packet leaves the AS. The chosen egress point can in turn have a significant
impact on the inter-domain path a packet travels on in the following.
8 291

thwarted if certain ASs on the path implement ingress detection or if egress detection is implemented
by the infiltrated AS. Let us explain these relationships using the generic configuration
depicted in figures 3 and 4.
m
m
. . .
. . .
j ... p k
. . .
. . .
j p‘ k
p ...
. . .
alternative path
. . .
alternative path
(a)
(b)
Figure 3: Prefix Hijacking
Node m represents the compromised AS. Consider a legitimate path from AS j to AS k emphasized
by a bold line, where AS k legitimately advertises a prefix. If no countermeasures
have been implemented, the nodes in the network cannot distinguish the legitimate originator
k from the illegitimate originator m and will thus accept the announcement from the AS which
is closer in terms of AS hops. We assume that if m announces a path which is as long as the
legitimate path, the nodes will pick the illegitimate path. 6 An attacker controlling AS m can
thus change the paths if all ASs on the bold path except k because at least AS p will accept m’s
announcements and forward all traffic intended for k to m instead. If AS m has deployed egress
filtering, the path is totally secure from attacks using m’s routers: All route advertisements are
’double-checked’ using authoritative information kept by AS m before they leave it and false
ones are prevented from passing the egress filter. Now consider the case when both, j and k
use ingress detection. This boils down to j and k being able to communicate authentication
information. Therefore AS j knows who is the legitimate originator of the prefix and will therefore
choose a path other than the bold path leading to AS k given it is aware of such a path.
This result can be generalized: Falsified AS paths connecting two endpoints will never carry the
traffic between those endpoints if both implement ingress detection. This is only a first step
towards routing security, however. Note that the attacker has at least succeeded in preventing
the traffic from flowing on the original path. 7 This is why we define this security property of a
legitimate inter-domain path semi-robustness.
Definition : Semi-Robustness of Inter-domain Paths
A legitimate inter-domain path is semi-robust if it cannot be changed in a way that its traffic
passes through or terminates in the AS that launched the routing attack. Either there exists an
alternative path that can be used instead, or the corresponding traffic does not flow at all.
Complete path robustness can only be assured if all ASs on the bold path that have deployed
ingress filtering and at the same time no AS is connected to m via a shorter AS path than with
k. We define inter-domain path robustness as follows:
Definition : Robustness of Inter-domain Paths
6 Note that this can be considered a worst case assumption. Since we only consider the relative results of our
measurements, this decision is not very important with respect to our conclusions.
7 Either the router at the two ends of the AS path know an alternative legitimate path through the AS topology
or the connection is disrupted.
9 292

A legitimate inter-domain path is robust if it is entirely resistant to the invalid behavior of
routers located in ASs that are not positioned along the path.
Inter-domain path robustness trivially implies inter-domain path semi-robustness: If an AS path
is robust, the traffic flowing on that path cannot be diverted such that it is impossible not to
reach the destination. The reverse is, however, not true.
m
m
. . .
. . .
spoofed link
. . .
. . .
spoofed link
j ... p k
j
...
p
q
k
. . .
alternative path
. . .
alternative path
(a)
(b)
Figure 4: Path Spoofing
Figure 4 depicts the generic situation in which an attacker uses path spoofing to divert traffic
from its legitimate path. AS m announces routes indicating that it is directly connected to k
while in reality it is not. Figure 4a) illustrates that path spoofing results in almost the same
threat for the legitimate path except that the distance between mislead AS j and m has to be at
least one AS hop smaller than the distance between AS j and k. Consequently, the situations in
which path spoofing is successful are less numerous than the situations in which prefix hijacking
is successful. If ingress detection was deployed both in j and k, k could tell j that it is not
directly connected to m and j would therefore not accept route announcements claiming that
this is the case. AS j would then forward traffic destined to k via another path. In fact, the
attacker would have to spoof a direct connection to the nearest AS on the legitimate path that
has not deployed ingress detection to maximize the chances of a successful attack. The legitimate
path can only be completely secured against path spoofing attacks if all ASs on the legitimate
route, except the two ASs that are one and two hops away from k, implement ingress detection.
4.3 Security Metrics
Based on the threat model we can now formally define the security metrics we use to determine
the security level of each AS in the topology.
Consider the generic configurations depicted in figures 3 and 4; false route announcements from
AS m can have adverse effects on the AS path connecting AS j and k. Let the AS topology
by denoted by G(V, E) where V stands for the set of ASs and E for the set of connections in
the topology. Based on this topology the global routing table is constructed. We represent this
construction process as a function denoted by Ψ. Consequently, Ψ maps the topology G(V, E)
on a global routing table satisfying the shortest path criterion. Let ψ denote a possible global
routing table, i.e. the relation Ψ(G(V, E)) ↦→ ψ holds. Any intermediate AS p on a legitimate
path from AS i to AS j therefore has the property p ∈ R ij where R ij is given by ψ.
We define a binary function describing the security level of that path, which takes k, j, m,
G(V, E) and ψ as arguments.
Definition : Security Indicator Function
The function sec(j, k, m, G(V, E), ψ) is equal to 1 if false route announcements injected by BGP
speakers in m will have no influence at all on the legitimate path from j to k, and equal to 0
10 293

otherwise.
We now define the probability that a legitimate path from j to k is vulnerable to an attack from
m. It is implicitly assumed that all ASs are equally likely to be used for routing attacks. The
probability is referred to as the route security metric in the following way.
Definition : Route Security Metric
s route (j, k, G(V, E), ψ) = 1 ∑
n m∈V,m/∈R jk
sec(j, k, m, G(V, E), ψ)
Using the route security metric we can measure the security of all n(n − 1) routes provided
the network topology G(V, E) and an instance of the global routing table ψ determined by the
function Ψ.
We compute the security metric describing the security level of AS i by averaging the route
security metrics of the (n − 1) routes leading form i to all other ASs j ∈ V and back.
Definition : AS Security Metric
s node (i, G(V, E), ψ) = 1
n−1
∑
j∈V,j≠i sroute (i, j, G(V, E), ψ)
Regarding these definitions the security measurement model can be represented as a function S
mapping the (n × 1)-vector of binary security protocol adoption decisions to another (n × 1)-
vector of node security metrics s node .
Definition : Security Measurement Function
S(e, G(V, E), ψ) = s node where e ∈ {0, 1} n and s node ∈ [0, 1] n
The route security metric is apparently a worst case metric, as it only assesses a route as secure
if it cannot be influenced by any means. In contrast to that the AS security metric is not a worst
case metric, as it rather represents the expected value of the event that the communication of a
particular AS is not affected by false routes announcements.
Example 4: Calculation of the Security Metrics
We demonstrate the calculation of the security measurement function for prefix hijacking
under the requirement of full path robustness using the topology introduced
in the earlier examples (n = 5). Assume that AS1 and AS5 have deployed ingress
detection. In order to calculate the AS security metric for AS1 we need to compute
the route security metric of all four legitimate routes which connect AS1 to the four
other ASs. The calculation of the route security metric of the legitimate path from
AS1 to AS2 implies evaluation of the security indicator function three times using
m = {AS3, AS4, AS5}. It does not have to be evaluated for m = AS1, AS2 since
these ASs lie on the legitimate path. It turns out that AS1’s prefixes can be hijacked
using AS3 and AS4, but not using AS5.
Using the formula for the route security metric we get
s route (AS1, AS2, G(V, E), ψ) = 1 5
= 0.6
∑
m∈{AS3,AS3,AS5}
sec(AS1, AS2, m, G(V, E), ψ)
Applying the same reasoning to the remaining three legitimate AS paths we obtain
s route (AS1, AS3, G(V, E), ψ) = 0.4
s route (AS1, AS4, G(V, E), ψ) = 0.6
s route (AS1, AS5, G(V, E), ψ) = 0.8
11 294

Therefore, the AS security metric of AS1 evaluates to
s node (AS1, G(V, E), ψ) = 1
n − 1
∑
j∈{AS2,AS3,AS4,AS5}
= 1 (0.6 + 0.4 + 0.6 + 0.8)
4
= 0.6
s route (AS1, j, G(V, E), ψ)
Note that for the full evaluation of the security measurement function this procedure
would have to be repeated for the remaining four ASs. For the other ASs we get
s node (AS2, G(V, E), ψ) = 0.55
s node (AS3, G(V, E), ψ) = 0.6
s node (AS4, G(V, E), ψ) = 0.55
s node (AS5, G(V, E), ψ) = 0.65
Thus the security measurement function evaluates to
S((1, 0, 0, 0, 1) T , G(V, E), ψ) = (0.6, 0.55, 0.6, 0.55, 0.65) T
The average AS security metric - which we will use as a proxy for the network’s
routing security - is the average of all AS security metrics, i.e. 0.59.
4.4 Simulation Procedure
The initial step of the simulation is the generation of a global routing table with n(n − 1) entries
for a given network topology. 8 As aforementioned, the sole criterion for path selection is path
length. In case there exist several paths with minimal length, one path is chosen with equal
probability. Thus, the function Ψ is inherently indeterministic. Throughout our experiments,
we used only one out of many possible global routing configurations. This simplification was
necessary due to computational constraints. We used artificial AS topologies of different shapes
and sizes for the experiments. In terms of the shape of topologies an important criterion of
differentiation is the distribution of node degrees. We used scale-free topologies as well as
topologies where the distribution of the node degree follows a normal distribution with little
variance, i.e. where all nodes had approximately the same node degree. In order to determine
the effect of topology size on our results, we compared the results from measurements on 20-
node topologies and 4000-node topologies. As the differences are almost negligible, we report
the results of measurements using a 4000-node Internet-like, i.e. scale free topology generated
by Inet-3.0 [25].
After the global routing table ψ has been constructed, the simulation applies the measurement
function S(e, G(V, E), ψ), which maps a vector of adoption decisions onto a vector of AS security
levels given the AS topology defined by G(V, E). In addition to the adoption decisions e, the
topology G(V, E) and the global routing table ψ, we considered three more independent variables
in our model not included in the formal definitions for notational convenience:
• The security scheme implemented by the adopters, i.e. ingress or egress detection
• The capability of the attackers, i.e. prefix hijacking, path spoofing or both attacks 9
• The security goal underlying the security metric, i.e. either semi-robustness or (complete)
robustness
8 Thus, exactly one path connecting each AS to all other ASs is stored although there may be many more.
9 If the attacker is able to use both kinds of attacks we assume that she first tries to use prefix hijacking and
if this should not be successful uses path spoofing.
12 295

Independent Variable
configuration of adoption decisions
AS topology
global routing table
security scheme
attacker capability
security goal
Values
ascending and descending deployment a
scale-free and random, 20 and 4000 ASs
shortest path routing b
ingress and egress detection
prefix hijacking, path spoofing and both attacks
semi-robustness and (complete) robustness
a Ascending deployment refers to a sequential deployment process starting with the AS that has the least interdomain
paths leading through it until full adoption is reached. Descending deployment proceeds in the opposite
way.
b For each topology we based the simulation on one random instance of the global routing table only. We do
not expect the indeterministic nature of the routing table construction to have a significant impact on our results.
Table 1: Summary of independent variables
Altogether 32 simulation treatments were run with different parameter configurations. Table 1
gives an overview of the independent variables and corresponding values we considered in our
simulation study.
5 Experimental Results
5.1 Fundamental Vulnerability
The benchmark model assumes that no countermeasures have been implemented. For this
benchmark, we measure the fundamental vulnerability of ASs. In terms of our notation, this
boils down to evaluating the security measurement function S with the parameter e = (0, ..., 0).
Shortest path selection in a scale free topology naturally results in few ASs with high node degree
being responsible for carrying the traffic of a much larger portion of ASs with low outdegree.
Subsequently, we counted the number of (legitimate) inter-domain paths leading through each
AS and formed groups of ASs with the same number of paths leading through them.
mean
0.9 1
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.0001 0.001 0.01 0.1 1
mean
0.9 1
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.0001 0.001 0.01 0.1 1
path fraction
path fraction
(a)
(b)
standard deviation
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
0.0001 0.001 0.01 0.1 1
path fraction
standard deviation
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
0.0001 0.001 0.01 0.1 1
path fraction
(c)
(d)
Figure 5: Mean and standard deviation of AS security metrics without countermeasures
Figure 5 shows plots of the mean (a) and standard deviation (c) of the AS robustness metric for
13 296

each of these AS groups if attackers are only capable of prefix hijacking. Plots (b) and (d) of
figure 5 show the same measurements applied to path spoofing. We restricted the presentation
of results to the robustness metric. The conclusions drawn from the depicted results are the
same for the semi-robustness metric. The measurement results for the third attack mode (prefix
hijacking combined with path spoofing) coincides with the results for prefix hijacking since prefix
hijacking is always more effective than path spoofing in the absence of countermeasures. Each
impulse in the plots represents one group of ASs with an equal number of inter-domain paths
leading through them. Accordingly, all stub ASs are allocated to the same group and this group
is represented by the first impulse from the left. 10 We normalized the horizontal axis to the unit
interval and applied a log scaling in order to improve visibility.
The plots indicate that there is a significant positive correlation between the number of interdomain
paths leading through ASs and their vulnerability to routing attacks. We obtained a
correlation coefficient of 0.1763 for prefix hijacking and 0.1205 for path spoofing. Thus, the
security level of an AS with more inter-domain paths going through it tends to be lower for
both kinds of attacks and this relationship is more significant for prefix hijacking. The plots
also show that vulnerability to prefix hijacking attacks is generally much higher than to path
spoofing. The observed correlation would be much higher if the security metrics of ASs with a
relatively small number of inter-domain paths leading through them were less scattered. The
plots of the standard deviations across the different groups show that the security metrics are
less scattered in groups of ASs that have more transit paths. The standard deviation of the stub
ASes’ security metric is particularly high, i.e. there are significant differences among their levels
of vulnerability.
In the Internet-like topology the ASs are more specialized on either serving as transit ASs or
mere traffic end-points (aka stubs). Given that paths are constructed solely based on the shortest
path criterion, this does not come as a surprise. If flat topologies are used instead of scale-free
ones, the number of inter-domain paths leading through the ASs is distributed more evenly
because there are fewer network hubs accommodating shortest paths. In our experiments using
this kind of topology we observed the same correlation between the number of paths leading
through an AS and its security metrics.
5.2 The Effects of Ingress Detection
To measure the effect of the deployment of ingress detection in the AS topology, we aggregated
the AS security metrics by computing the average over all ASs (including the non-adopters).
The aggregate metric therefore serves as a metric for the resilience against routing attacks of
the whole network. Figure 6 shows how the mean AS security metrics based on path robustness
evolve for all three attack modes if ingress detection is deployed in an increasing number of ASs.
While figure 6a) shows the measurement results obtained when deployment was performed in
ascending order of the number of inter-domain paths leading through the ASs, figure 6b) shows
the results obtained for the opposite deployment order, i.e. descending deployment.
Figure 6a) indicates that the mean AS security metrics for path robustness significantly increase
only if more than 90% of the AS population deploy ingress detection. Furthermore, the metrics
hardly move until about half of the ASs take part in the effort. Remember that this approximately
accounts for the number of stub ASs in the sample. Therefore we can conclude that
stub ASs contribute almost nothing to the overall robustness of the network if they are alone to
deploy ingress detection.
From the comparison of figures 6a) and 6b) one can observe that if the security protocol is
deployed in ascending order, the marginal mean AS security metric increases at a lower rate in
10 For the 4000 node topology we used in our experiment, the algorithm responsible for constructing the global
routing table terminated with more than half of the ASs being stubs.
14 297

mean AS security metric (robustness)
1
0.9
0.8
0.7
0.6
0.5
only prefix hijacking
only path spoofing
both attacks
0.4
0 0.2 0.4 0.6 0.8 1
adopting fraction
mean AS security metric (robustness)
1
0.9
0.8
0.7
0.6
0.5
only prefix hijacking
only path spoofing
both attacks
0.4
0 0.2 0.4 0.6 0.8 1
adopting fraction
(a)
(b)
Figure 6: Mean robustness for ascending (a) and descending (b) deployment of ingress detection
mean AS security metric (semi-robustness)
1
0.9
0.8
0.7
0.6
0.5
only prefix hijacking
only path spoofing
both attacks
0.4
0 0.2 0.4 0.6 0.8 1
adopting fraction
mean AS security metric (semi-robustness)
1
0.9
0.8
0.7
0.6
0.5
only prefix hijacking
only path spoofing
both attacks
0.4
0 0.2 0.4 0.6 0.8 1
adopting fraction
(a)
(b)
Figure 7: Mean semi-robustness for ascending (a) and descending (b) deployment of ingress
detection
the beginning and at a higher rate toward the end compared to the case where deployment takes
place in the opposite order (descending deployment). The major transit ASs therefore have a
crucial influence on the overall routing security in the AS topology. Without their participation
hardly any additional inter-domain paths can be made robust against routing attacks. If they
deploy ingress detection right at the beginning, they pave the way for a faster increase of the
mean path robustness metric as the ASs with fewer or no transit paths leading through them
join.
If semi-robustness is measured instead of complete robustness, the mean security metrics evolve
differently. Figure 7 depicts the development of the mean security metric based on semirobustness.
Comparing 6b) and 7b) we see that if the ASs deploy ingress detection in descending
order, the mean security metrics for semi-robustness and complete robustness almost coincide. 11
Therefore the number of paths which are semi-robust but not robust at the same time is very
small during the whole deployment process. If the deployment order is reversed, however, this
number becomes much greater. The explanation for this observation can be found in the fact
that semi-robustness is a weaker security requirement. More importantly, it shows that if the
11 In fact, the former is always slightly greater than the latter.
15 298

aim is to only enhance the overall semi-robustness of the network, the order of deployment plays
a much less significant role. If complete robustness is required, descending deployment is clearly
superior to ascending deployment in terms of overall resilience against routing attacks.
Let us now have a closer look at some of the more subtle differences between the curves in
figure 6a) and 7a). We have already noted that the importance of the deployment order is less
significant if we consider semi-robustness instead of complete robustness. This can be attributed
to the fact that a path can be made semi-robust against prefix hijacking if ingress detection is
deployed at both ends of the path whereas complete robustness in the worst case requires deployment
of security filters along the whole inter-domain path. However, this does not explain
why the mean semi-robustness metric in figure 7a) tends to be higher than the one in figure
7b). Apparently, the ASs with fewer legitimate paths leading through them contribute more to
the increase of the mean AS security than the major transit ASs do. If one considers the initial
distribution of security levels described in section 5.1, the reason for this effect can be explained
by the relatively low baseline vulnerability of transit ASs. Since many of their inter-domain
paths are robust (or at least semi-robust) even in case no countermeasures are deployed in the
network, there is little additional security to be gained among them. Those ASs accomodating
fewer or no transit paths gain more on average if more ASs deploy ingress detection. Since
the fundamental distribution of vulnerability has the same properties both for semi-robustness
as well as complete robustness, this explanation given above applies even more generally. The
reason that the effect cannot be observed in figure 6a) is that transit ASs play such an important
role in providing (complete) robustness of their inter-domain paths. Thus they prevent stub ASs
from realizing security gains as long as they have not deployed ingress detection.
Another interesting observation from the comparison of the plots is that the mean semi-robustness
against prefix hijacking increases particularly quickly if deployment takes place in increasing order.
This can be explained by the fact that assuring semi-robustness against prefix hijacking
only requires the deployment of ingress detection at both ends of the legitimate path. However,
if the attacker is able to path spoof, this might not be enough. In this case the structural
position of the path within the topology remains an important factor in determining the degree
of semi-robustness. For example, if the infiltrated AS (1) spoofs a link to one of the neighbors
of the destination AS which has not deployed ingress detection and (2) the spoofed path is not
longer than the legitimate one, the path is not considered semi-robust according to our simulation.
Thus, resilience against path spoofing in many cases requires the deployment of ingress
detection in intermediate ASs. 12
Our next step was to divide the set of ASs into stub and transit ASs and compute the mean AS
security metric of both groups separately. We did so in order to find out more about the mutual
dependencies of the two groups.
Figure 8 depicts the mean security metrics of stub and transit ASs based on complete robustness
(figure 8) and on semi-robustness (figure 9) given the respective group is the first deploying
ingress detection. The gap between the last measurement point in these diagrams and complete
security (i.e. a mean AS security metric of 1) represents the security increase of the observed
group that would be achieved if the other group also adopted ingress detection. In economic
terms, the effect of the action of one entity on the utility of another entity which has no influence
on the effect itself is called externality. Borrowing from the economics literature we therefore
denote this effect security externality. The comparison of figures 8a) and b) reveals that stub
ASs can achieve almost no increase of their mean security metric without the adoption of transit
ASs whereas the transit ASs can achieve a significant increase without the adoption of the
stub ASs. Accordingly, stub ASs in terms of complete robustness depend more strongly on the
adoption of the transit ASs than vice versa.
Figure 9 shows the development of the mean security metric of the two groups if the weaker
security requirement semi-robustness is the objective. 9a) and b) indicate that the asymmetry
12 You may want to refer to the description of the simulation model to clarify this point.
16 299

mean AS security metric (robustness)
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
only prefix hijacking
only path spoofing
both attacks
0.2
0 0.2 0.4 0.6 0.8 1
adopting fraction
mean AS security metric (robustness)
1
0.95
0.9
0.85
0.8
0.75
0.7
0.65
0.6
0.55
only prefix hijacking
only path spoofing
both attacks
0.5
0 0.2 0.4 0.6 0.8 1
adopting fraction
(a)
(b)
Figure 8: Comparison of the robustness of stub (a) and transit ASs (b)
mean AS security metric (semi-robustness)
1
0.9
0.8
0.7
0.6
0.5
0.4
only prefix hijacking
only path spoofing
both attacks
0.3
0 0.2 0.4 0.6 0.8 1
adopting fraction
mean AS security metric (semi-robustness)
1
0.95
0.9
0.85
0.8
0.75
0.7
0.65
0.6
0.55
only prefix hijacking
only path spoofing
both attacks
0.5
0 0.2 0.4 0.6 0.8 1
adopting fraction
(a)
(b)
Figure 9: Comparison of the semi-robustness of stub (a) and transit ASs (b)
of dependencies is less significant compared to the case where complete robustness of the paths
is required. We have noted earlier that the deployment of ingress detection in intermediate ASs
is oftentimes not required in order to make a path semi-robust. Hence, the stub ASs can secure
a significant fraction of the paths connecting them to each other without the participation of
transit ASs in this case.
All mean security metrics depicted in figures 8 and 9 are higher for the group of transit ASs.
This is because the more inter-domain paths lead through an AS, the more secure it usually
is. Since stub ASs are present on the minimum number of paths, namely just the n − 1 paths
connecting each of them with all other ASs in the Internet, their average security metric is lower
than the one of the remaining ASs on average.
5.2.1 The Effects of Egress Detection
The development of the mean AS security metric changes substantially if egress instead of ingress
detection is deployed in the topology.
Figure 10 shows how the this metric evolves as egress detection is deployed sequentially in ascending
and descending order. In both deployment modes the mean robustness increases with
each adopting AS. According to our model, deploying egress detection in an AS is equivalent
17 300

mean AS security metric (robustness)
1
0.9
0.8
0.7
0.6
0.5
only prefix hijacking
only path spoofing
both attacks
0.4
0 0.2 0.4 0.6 0.8 1
adopting fraction
mean AS security metric (robustness)
1
0.9
0.8
0.7
0.6
0.5
only prefix hijacking
only path spoofing
both attacks
0.4
0 0.2 0.4 0.6 0.8 1
adopting fraction
(a)
(b)
Figure 10: Mean robustness for ascending (a) and descending (b) deployment of egress detection
mean AS security metric (robustness)
1
0.9
0.8
0.7
0.6
0.5
0.4
only prefix hijacking
0.3
only path spoofing
both attacks
0.2
0 0.2 0.4 0.6 0.8 1
adopting fraction
mean AS security metric (robustness)
1
0.95
0.9
0.85
0.8
0.75
0.7
0.65
0.6
only prefix hijacking
0.55
only path spoofing
both attacks
0.5
0 0.2 0.4 0.6 0.8 1
adopting fraction
(a)
(b)
Figure 11: Comparison of the mean robustness of stub (a) and transit ASs (b)
to eliminating this AS as a potential threat for the inter-domain paths linking together other
ASs. Therefore the linear shape of the curves depicted in figure 10 indicates that each AS can
be used to conduct successful attacks on at least some portion of the paths. Comparing figures
10a) and 10b) more closely, we see that - although the metrics increase in an almost linear way
in both deployment modes - securing the transit ASs first causes the metric to rise more quickly
in the beginning. This shows that attacks launched from core ASs can be expected to have more
severe consequences in terms of the number of inter-domain paths affected.
Comparing the plots of figure 10 with the ones obtained for ingress detection (figure 6), we
observe that switching the deployment mode has much more severe consequences for the development
of the mean security metrics in the case of ingress detection. Furthermore we find
that irrespective of the deployment order, the egress detection approach dominates the ingress
detection approach: For all analyzed configurations of the input vector e (except zero and full
adoption) it is greater.
Since the results for semi-robustness do not differ from the mean security metric evolves, we
omit them. The explanation for this similarity is straightforward given the premises of our simulation:
Deploying egress detection in an AS causes the potential adverse influence of the AS on
legitimate paths in the topology to vanish completely. Thus, the difference between robustness
18 301

and semi-robustness does not have any influence on the way the mean security metric evolves. 13
Figure 11 shows the development of the mean security metric based on complete robustness of
the groups of stub and transit ASs separately. The plots reveal that the security externalities
we have already observed in the ingress detection case also exist in the egress detection case.
However, the dependency of the stub on the transit ASs as compared to ingress detection is far
less significant. The stub ASs can significantly improve their own security by adopting egress
detection while they can achieve almost no additional security if they deploy ingress detection.
6 Summary and Discussion
This paper provides several contributions to the discussion of how tomorrow’s Internet infrastructure
can be protected against routing attacks, which opens up new avenues for protocol
research. The contributions can be summarized as follows:
First and most importantly, taking a step back by questioning the approach that takes ingress
detection for granted has brought us to the definition and motivation of an alternative approach
which we call egress detection. Egress detection requires less communication overhead and fewer
standardization efforts compared to ingress detection and therefore represents a much less complex
and less expensive solution. On the other hand, egress detection fails in terms of securing
specific paths. In order to quantify the effects ingress and egress detection have we adapt a
simulation model that was introduced by Chan et al. [4].
While the security measurement methodology used in our paper is based on the same principle,
the model we use is less detailed with respect to attacker characteristics 14 and makes less stringent
assumptions 15 . We have noted earlier that the model we use does not address many of the
complexities of the BGP. In order to compare the efficiency of ingress and egress detection on
a strategic level our simulation can yield important insights. The implementation of our model
is capable of showing the effects of prefix hijacking and path spoofing attacks separately. This
way we have shown that the inter-domain routing infrastructure is about twice as vulnerable to
prefix hijacking than to path spoofing attacks and that if attackers have both types of attack at
their disposal, path spoofing does not significantly increase the impact of attacks.
Furthermore, we have defined two different security metrics based on different robustness requirements
(robustness and semi-robustness of inter-domain paths).
Our measurement study has revealed that the vulnerability of the different ASs is correlated
with the number of inter-domain paths leading through them and therefore depends on the
roles the ASs play in provisioning connectivity. The results suggest that stub ASs are more
vulnerable to routing attacks than transit ASs. Furthermore, we have shown experimentally
how ingress and egress detection differ with respect to how efficiently they enhance the routing
security of the whole network. We have discovered that deploying ingress and egress detection
in increasing and decreasing order of the number of paths that lead through each AS is an
effective way to reveal existing dependencies between ASs. By separating the set of ASs in the
sample topology into stub and transit ASs, we have directly shown the existence of what we call
security externalities, i.e. that the action of deploying either countermeasure in one AS has a
positive effect on the security metric of other ASs. If ingress detection is used, security benefits
13 The only difference between the robustness and semi-robustness plots is that the semi-robustness curve possesses
a higher intercept. This can again be explained by the fact that semi-robustness is the weaker requirement.
We have seen earlier that there are a few more paths that are semi-robust than there are robust paths given no
countermeasures are deployed.
14 Chan et al. use two different attacker models based on different degrees of knowledge that the attacker has
about the topology. We adopted what Chan et al. denote as the strong attacker model, i.e. attackers are expected
to know how the global routing table looks like before the attack.
15 Chan et al. use an AS security metric based on the assumption that all inter-domain paths leading through
an AS are relevant for that AS, whereas we only consider the n-1 paths that connect that AS with all other ASs
in the topology. They also introduce a weighing factor differentiating the importance of different paths based on
some estimate of for traffic load while we treat all paths as equal.
19 302

Property Ingress Detection Egress Detection
type of externality positive network positive pure
mean (semi-)robustness metric convex linear
asymmetric dependency stub ASs depend on transit ASs a none
subnet (semi-)robustness practical impractical
rank of dominance b 2 1
a Only applies if complete robustness is required.
b The term dominance refers to the number of paths that are (semi-)robust given a fraction of adopters. A
lower number translates to a higher rank, i.e. egress detection dominates ingress detection with respect to this
definition.
Table 2: Comparison of ingress and egress detection
are restricted to the set of adopters while they spill over from adopters to non-adopters in case
egress detection is used. Therefore we are dealing with two different kinds of externality, namely
network externalities in the case of ingress detection and pure externalities in the case of egress
detection.
The results of our experimental analysis suggest that security externalities are more or less
asymmetric depending on the countermeasure used (ingress or egress detection), how security
is measured (requiring semi-robustness or complete robustness) and the kind of ASs that the
countermeasure is deployed in. If ingress detection is used, stub ASs can achieve no significant
increase of their path robustness levels without the help of transit ASs whereas transit ASs are
much less dependent on the stub ASs. However, this asymmetry does not exist if semi-robustness
instead of complete robustness is being measured. Therefore we argue that a more explicit definition
of the security goal is needed as the objective regarding robustness requirements may
influence adoption incentives. Do we only want to make sure that traffic cannot be hijacked by
the attacker (then semi-robustness would be a sufficient security requirement), or do we want
to maintain the legitimate paths in any event (then complete robustness would be required)?
Another important result of the experiments is that using egress detection as a countermeasure
more overall security can be achieved by fewer adopters, i.e. that the egress detection approach
dominates the ingress detection approach with respect to efficiency. However, egress detection
has the downside that neither robustness nor semi-robustness of particular paths can be guaranteed
if it is only deployed in the ASs on these paths. 16 Thus, complete robustness of a subset
of the inter-domain paths can only be effectively achieved by the use of ingress detection. We
call this property subnet (semi-)robustness.
We have summarized our findings concerning the comparison of ingress and egress detection in
table 2.
7 Conclusions and Future Work
Inter-domain routing security has been on the agenda of the networking community for quite
some time now. Although an impressive number of proposals for security patches lie on the table,
there has still not been any substantial progress with respect their adoption. All comprehensive
security solutions put forward to make the Internet’s infrastructure more robust to routing
attacks suffer from high adoption barriers due to technical issues (protocol standardization and
router overhead) which are in turn amplified by network effects. We conclude that these ingress
detection schemes would have to be deployed by a significant number of AS operators in order
to reliably prevent serious outages in the case of routing attacks. Furthermore, our analysis has
16 This statement has to be seen within the bounds of our model, of course. In reality inter-domain routing can
be a highly dynamic process and therefore the guaranteed robustness of a particular end-to-end connection may
require adoption of ingress detection by a different set of ASs in the course of time.
20 303

evealed strong asymmetries in the dependency relationship of the ASs at the top of the Internet
hierarchy and the ones at the bottom: Stub ASs for instance are more dependent on transit ASs
to deploy ingress detection than vice versa. This dependency is very relevant since the major
transit ASs have little incentive to take part in the security effort given the fact that their paths
are pretty safe from routing attacks anyway. Following the ingress detection approach may thus
lead to a situation where adoption is prevented due to strategic considerations on both sides.
Egress detection could represent an alternative security approach. It is less demanding in terms
of overhead and standardization efforts. According to our results, deployment of egress detection
in a given set of ASs provides more routing security on average than the deployment of
ingress detection. Furthermore, the asymmetries in the dependency structure do not come into
existence if the egress detection approach is being followed. Also, since many routing attacks
occur rather by chance than by intuition, egress detection is an inexpensive alternative to ingress
detection when it comes to preventing unintended misconfiguration. However, egress detection
cannot be effectively used to make a particular set of inter-domain paths more robust, which
might be regarded as crucial by AS operators eager to protect the most important connections
first. Furthermore, egress detection suffers from pure externalities and thus adoption is not
individually incentive-compatible.
More efforts should be devoted to the specification of more realistic simulation models. For
instance, the behavior of the BGP could be modeled more realistically (e.g. using C-BGP [19]).
Furthermore, the effects of policy routing on path selection have been ignored throughout this
work and may also be a worthwhile topic of future research with respect to the security background
[22].
Although we have outlined the general functionality of egress detection, a technical analysis
would have to follow in order to determine whether corresponding mechanisms are feasible at all
and if so how they could be implemented. One could also imagine to mix the two approaches, i.e.
to deploy ingress and egress detection in parallel, in order to further optimize robustness for a
given set of adopters. From a technical viewpoint it would be interesting to see what additional
infrastructure is needed to implement egress detection on top of ingress detection.
Routing attacks certainly are no purely BGP-specific threat. They should be included into security
considerations regarding all networked systems. Recent research in related areas like sensor
networks show interesting parallels [6].
What makes this work particularly interesting is the inherent connection of network security and
economics. Some of the results we obtained can actually be used as a foundation of economic
argumentation and provide decision support for policy makers. We have seen that, depending on
the technological approach of increasing the robustness of inter-domain routing, different types
of externalities come into existence. This in turn leads to different economic incentives and corresponding
instruments to overcome the economic adoption barriers. A promising research avenue
would therefore be to further investigate the interdependency between technological features of
BGP security schemes, economic incentives and appropriate policies.
References
[1] W. Aiello, J. Ioannidis, and P. McDaniel. Origin Authentication in Interdomain Routing.
In 10th ACM Conference on Computer and Communications Security, 2003.
[2] The President’s Critical Infrastructure Protection Board. The National Strategy to Secure
Cyberspace, 2003.
[3] K. Butler, T. Farley, P. Mcdaniel, and J. Rexford. A Survey of BGP Security Issues and
Solutions. Technical report, AT&T Labs - Research, Florham Park, NJ, February 2004.
21 304

[4] H. Chan, D. Dash, A. Perrig, and H. Zhang. Modeling Adoptability of Secure BGP Protocols.
In Raymond A. Marie, Peter B. Key, and Evgenia Smirni, editors, SIGCOMM, pages
389–390. ACM, 2006.
[5] H. Chang, R. Govindan, S. Jamin, S. J. Shenker, and W. Willinger. Towards Capturing
Representative AS-Level Internet Topologies. Computer Networks, 44(6):737–755, 2004.
[6] M. Feldman, J. Chuang, I. Stoica, and S. Shenker. Hidden-Action in Multi-Hop Routing.
In CECOMM: ACM Conference on Electronic Commerce, 2005.
[7] V. Gill, J. Heasley, and D. Meyer. The Generalized TTL Security Mechanism (GTSM).
RFC 3682, February 2004.
[8] G. Goodell, W. Aiello, T. Griffin, J. Ioannidis, P. Mcdaniel, and A. Rubin. Working Around
BGP: An Incremental Approach to Improving Security and Accuracy of Interdomain Routing.
In NDSS, April 2003.
[9] A. Heffernan. Protection of BGP Sessions via the TCP MD5 Signature Option. RFC 2385,
March 2002.
[10] Y. Hu, A. Perrig, and M. A. Sirbu. SPV: Secure Path Vector Routing for Securing BGP.
In SIGCOMM, pages 179–192, 2004.
[11] G. Huston. Interconnection, Peering, and Settlements. http://www.potaroo.net, 2003.
[12] S. Kent, C. Lynn, J. Mikkelson, and K. Seo. Secure Border Gateway Protocol (S-BGP) -
Real World Performance and Deployment Issues. February 2000.
[13] S. Kent, C. Lynn, and K. Seo. Secure Border Gateway Protocol (S-BGP). IEEE Journal
on Selected Areas in Communications, 18, 2000.
[14] Ratul Mahajan, David Wetherall, and Tom Anderson. Understanding BGP Misconfiguration.
In SIGCOMM, pages 3–16. ACM, 2002.
[15] S. Misel. Wow, AS7007! Merit NANOG Archive, 1997.
[16] S. Murphy. BGP Security Vulnerabilities Analysis. RFC 4272, January 2006.
[17] D. M. Nicol, S. W. Smith, and M. Zhao. Efficient Security for BGP Route Announcements.
Technical Report TR2003-440, Dartmouth College, Computer Science, Hanover, NH, May
2003.
[18] A. Odlyzko. Economics, Psychology, and Sociology of Security, 2003.
[19] B. Quoitin and S. Uhlig. Modeling the Routing of an Autonomous System with C-BGP.
IEEE Network, 2005.
[20] Y. Rekhter and T. Li. A Border Gateway Protocol 4 (BGP 4). RFC 4271, January 2006.
[21] L. Subramanian, V. Roth, I. Stoica, S. Shenker, and R. H. Katz. Listen and Whisper:
Security Mechanisms for BGP. In NSDI, pages 127–140. USENIX, 2004.
[22] H. Tangmunarunkit, R. Govindan, S. Shenker, and D. Estrin. The Impact of Routing Policy
on Internet Paths. In IEEE INFOCOM, December 2001.
[23] European Union. Network and Information Security: Proposal for a European Policy Approach.
COM(2001)298 final, June 2001.
22 305

[24] R. White. Deployment Considerations for secure origin BGP (soBGP). Internet Draft,
2002.
[25] J. Winick and S. Jamin. Inet-3.0: Internet Topology Generator, July 11 2002.
[26] H. Yu, J. Rexford, and E. W. Felten. A Distributed Reputation Approach Cooperative
Internet Routing Protection. In Workshop on Secure Network Protocols, 2005.
[27] X. Zhao, D. Pei, L. Wang, D. Massey, A. Mankin, S. F. Wu, and L. Zhang. Detection of
Invalid Routing Announcement in the Internet. In International Conference on Dependable
Systems and Networks (DSN), 2002.
23 306

A Market Mechanism for Energy Allocation in Micro-CHP Grids
Carsten Block, Dirk Neumann, Christof Weinhardt
Universität Karlsruhe (TH), Germany
{block, neumann, weinhardt}@iism.uni-karlsruhe.de
Abstract
Achieving a sustainable level of energy production and
consumption is one of the major challenges in our
society. This paper contributes to the objective of
increasing energy efficiency by introducing a market
mechanism that facilitates the efficient matching of
energy (i.e. electricity and heat) demand and supply
in Micro Energy Grids. More precisely we propose
a combinatorial double auction mechanism for the
allocation and pricing of energy resources that especially
takes the specific requirements of energy producers and
consumers into account. We describe the potential role
of decentralized micro energy grids and their coupling
to the large scale power grid. Furthermore we introduce
an emergency fail over procedure that keeps the micro
energy grid stable even in cases where the auction mechanism
fails. As the underlying energy allocation problem
itself is NP-hard, we derive a fast heuristic for finding
efficient supply and demand allocations. In addition we
show the applicability of this approach through numerica.
Keywords: Combined Heat and Power; Market
mechanism; Double Auction; Combinatorial allocation
1 Introduction
Over the last years fossil energy prices were rising remarkably.
E.g. the price for Crude oil increased from approximately
20 USD in 1995 to more than than 65 USD
as of today [25] while the prices for other fossil resources
evolved similarly [13]. On the one hand rising energy
prices indicate the increasing scarcity of these resources
and thus implicate a need for change in our energy consumption
habits, on the other hand high prices foster the
development of alternative technologies for energy generation.
A study from Emerging Energy Research [16]
states that today wind power technology already allows
electricity generation at cost below those of conventional
coal fired power plants if carbon penalties of more than
30 EUR per tonne are applied. 1 Similarly Krewitt and
Schlomann [17] state that energy produced from renewable
resources is already competitive if external cost are
included.
Taking this ongoing change process as given one
needs to determine the challenges that come along with
this development. In particular the effects on the future
energy generation and distribution infrastructures need
to be addressed [6]. Traditionally energy generation is
a centralized business, mainly because the required (fossil)
resources were (and still are) available “on the spot”
and central generation allows to exploit some economies
of scale. Throughout the last years the development
of new energy generation technologies fosters the utilization
of renewable energy resources. But compared
to large scale power plants, energy generation through
photovoltaic systems, biogas plants or even windparks is
(i) a much more decentralized and (ii) a much less predictable
business. Thus this development puts a lot of
pressure on the current energy distribution infrastructure.
Our proposition to address this issue is to develop
semi-autonomous micro energy grids that bind the
distributed energy generation facilities to the local demands.
The local netting of energy demand and supply
inside a microgrid could already reduce parts of the randomness
and thus should enable the microgrid to appear
as a rather predictable “citizen” in the large scale grid
where it buys peak demands from the outside outside
world and offers free generation capacities as balancing
power to others. Furthermore a combined usage of heat
1 In May 2006 carbon prices in the European Cap-and-Trade
System already reached 30 EUR per tonne of CO 2
1
307

and power inside the microgrid will increase the overall
energy efficiency.
Besides difficulties in predicting and coordinating
distributed energy supply, energy consumption forecasting
– especially for private households – is problematic as
well. This is mainly due to two reasons: On the one hand
old metering technology limits readouts in practice to annual
cycles, on the other hand flat fee tariffs (e.g. 18 Ct
/ kWh of electricity) do not set any economic incentives
for consumers to shift energy consumption from peak
hours to periods of low consumption. The result is that
consumers use energy more or less “at random” while energy
providers are more or less blind on the energy flows
inside their low voltage distribution networks. At the
same time it is known from field studies that about 50%
of the private households’ electricity consumption is dedicated
to the operation of refrigerators, freezers, (water)
heaters, washing mashines or dryers [24] and thus could
follow – at least within certain boundaries – a planned
schedule.
Our solution for addressing the aformentioned challenges
is to develop combined heat and power micro
energy grids (CHPMEG) that are “loosely” connected
to the conventional large-scale power grid minimizing
energy transmission losses while maximizing generation
efficiency ratios. Within a CHPMEG energy producers
and energy consumers trade their energy (i.e. heat
and electricity) supplies and demands on local marketplaces.
Arbitrage agents ensure that the resulting energy
prices within the microgrid are levelled to those on
external marketplaces such as the European Energy Exchange
(EEX). As markets provide an efficient matching
of demand and supply on average, a technical balancing
mechanism (spinning reserve) complements the market
mechanisms to ensure stable operation of the local energy
grid, even in those cases where the market-based allocation
might fail without compromising the economic
2 http://www.sesam.uni-karlsruhe.de
tract making on the platform[8]. In this paper we focus
on the hybrid control mechanism for CHPMEG that is
developed as part of this project. The applicability of
incentives set by the markets. The work presented in
this paper is part of the project “Self-organization and
Spontaneity in Liberalized and harmonized Markets” 2
(SESAM). The main focus of the project ist to develop
an electronic energy market platform that also features
mechanisms for secure and non repudiable communication
between the market participants, permits the utilization
of electronic trading agents and provides an electronic
law mediator who supervises the electronic con-
our approach is demonstrated through numerica. The
remainder of this paper is structured as follows. Section
2 describes the typical setup of a CHPMEG and describes
the importance and role of energy markets within
the microgrids. Based on this several requirements for
the operation of CHPMEG are deduced which serve as
benchmarks for the literature review in section 3. In section
4 the formal market model is introduced along with
a fast heuristic for implementation purposes. Section 5
concludes with an application case study and an outlook
on future work.
2 Motivational Scenario
In this section we describe an application for a market
based micro CHP grid which typically consists of several
components connected to each other trough heat pipes
and power / communication lines. Usually five key components
can be identified in a micro grid as shown in
figure 1 [1, 18].
1. Consumers are private households and small business
that require electricity and / or heat.
2. Producers are photovoltaic systems, heaters,
micro-turbines, biogas plants, etc. They generate
electricity, heat or both.
3. Spinning Reserves provide ad-hoc energy reserves
that is used to stabilize voltage and frequency of
the power grid. Battery banks or spinning wheels
usually provide facilities for quick electricity storage
and retrieval and thus serve as spinning reserves.
3 Heat-wise grid stability is not too critical,
still therms can take over the role of heat regulators
as they are able to temporarily store certain
amounts of heat.
4. The coupling point is the connection between CH-
PMEG and the large scale grid. At this point in-
3 Balancing power from the large scale grid can complement the
migrogrid’s spinning reserves. Still, distinct spinning reserves are
oftentimes build into the micro grid as well in order to gain a
certain degree of independence from power outages in the large
grid.
308

ound and outbound power flows are metered and
regulated.
5. The controller is the core of the CHPMEG. It
is responsible for matching the consumer’s energy
demandy by dispatching appropriate generators for
heat and / or electricity, i.e. it has to determine an
optimal production schedule given the consumption
requests.
Although figure 1 depicts consumers and producers
as distinct entities these roles can switch dynamically.
A household might generate and sell excess electricity
from its own photovoltaic system during daytime while
it might consume power during night time. We require
all entities within a CHPMEG, i.e. producers, consumers
and spinning reserves to be connected to the controller
through power lines, local area network (LAN) and –
where applicable – through heat pipes as well. Furthermore
we assume that heat pipe and power line capacities
are sufficiently dimensioned to meet all transmission demands
that might occur.
As mentioned above, the central component of this
microgrid is the Controller which determines energy production
and consumption of all grid participants and ensures
that (i) the electricity grid is always balanced in
terms of voltage and frequency and (ii) the heat grid
always provides at least as much heat as required, storing
or disposing overproduction in therms. Furthermore
the controller should provide incentives to consumers to
shift their energy consumption from peak hours to time
periods with low energy demands if possible. Likewise
on the producer side the controller needs to determine
an operation schedule that always favors the producer
with the lowest production cost for generating the required
energy. If e.g. a photovoltaic system and a microgas
turbine are ready to produce power for the microgrid,
the photovoltaic system will likely have comparably
lower production cost during sunshine periods and thus
should produce the energy saving the microturbine from
burning gas.
This idea could be extended even further: The controller
needs to make sure that not only the cheapest
producers inside the microgrid generate the required energy
but that the overall cheapest producers will take
over this responsibility, i.e. if the spot market price for
electricity e.g. at the EEX is lower than the marginal
production cost of the most expensive currently running
generator in the microgrid, it should be turned off and
the electricity should be purchased from the EEX instead
(and vice versa). The same procedure should be applied
if – during peak hours – the local generation capacity is
insufficient to meet the local demands. From a technical
perspective is important to note that many generators,
especially combustion engines are not able to operate at
below certain load levels. Thus the CHPMEG controller
hast to take minimum and maximum production constraints
of the respective energy producers into account
as well.
Finally the netting of energy demand and supply
within the CHPMEG should reduce the randomness
in energy production and consumption to the outside
world. At best the CHPMEG is able to announce future
energy shortages (or energy reserves) as early as possible
to the outside world, making its behavior as predictable
as possible.
In this section we collected several different requirements
that CHPMEGs and especially the CHPMEG
controllers should meet. These can be summarized as
follows:
R1
R2
R3
R4
R5
R6
R7
Online Mechanism (instantaneous matching
of demand and supply)
Bundled allocation of electricity and heat
Price signals to indicate scarcity
Min and max. allocation constraints
Stable operation (balanced demand & supply)
Coupling with large-scale power grid
Forecasts for demand and supply
In the following section we will review related approaches
for realizing this control component before we
introduce our market based controller in section 4.
3 Related Work
The idea to employ market mechanisms for distributed
environments is not new. Since the 60s researchers have
motivated the use of markets as a means to cope with
incentival problems especially for distributed computing.
But it was only recently that the idea of decentralization
spilled over from computer grids to power grids [4, 10].
309

Figure 1: Schematic Layout of a Micro CHP Grid
Also relatively new is the idea to use markets for the
coordination of decentralized resources. Among the first,
Buyya, Wolski et al. [5], and Subramoniam et al. [23]
motivated the use of auctions and negotiations for Computational
Grid. In their first paper, Wolski et al. [27]
suggested the use of traditional auction formats such as
English auctions. Eymann et al. [12] introduce a decentralized
bargaining system for resource allocation. Regev
et al. [21] propose the use of a Vickrey auction for allocation
computational resources in distributed systems.
Although often suggested in literature, the effectiveness
of traditional bargaining and auction systems in
grid environments is conceivably delimited as the trading
objects are traded as unbundled standardized commodities.
Consequently, these auction formats fail to express
demand on bundles – exposing especially owners of CHP
generators to the risk of selling only e.g. electricity without
selling the co-generated heat.
Reviewing the requirements upon the mechanism,
it becomes obvious that the previous described mechanisms
fail to satisfy these requirements. Most of the
mechanisms proposed in literature do not meet the immediacy
requirement in a way that they are either iterative
in nature or NP hard to compute. Both properties
rule out a use in highly interactive markets for different
reasons. While iterative mechanisms are unsuited due
to the fact that they frequent user feedback is needed,
NP hard mechanisms consume too much time to be of
use in large scale markets. Beside the requirement of immediacy,
the negligence of capacity constraints for bids
diminish the use of the proposed market mechanisms.
To account for predictability of the system time
attributes, Wellman et al. model single-sided auction
protocols for the allocation and scheduling of resources
under consideration of different time constraints . Conen
goes one step further by designing a combinatorial
bidding procedure to improve the resource scheduling offering
different running, starting, and ending times for
different resources. Furthermore, Bapna et al propose
an auction that allows to bid on bundles and on time
slots. In addition, the authors also suggest a heuristic
that approximates the outcome of the auction in polynomial
time. However, these approaches are single-sided
and hence do not create competition on both sides. Furthermore
in electricity grids it might not be clear, when
a consumer wants to start consuming energy and how
long his consumption will take. Thus integrating time
schedules into the auction mechanism does not solve the
310

Requirements R1 R2 R3 R4 R5 R6 R7
√
√
Eymann et al. 2003
√
–
√
– – – –
Buyya et al. 2001
√
–
√
– – – –
Wolski et al. 2003
√
–
√
– – – –
Regev et al. 1998
–
√ √
– – – –
Wellman et al. 2001 –
√ √
– – – –
Conen 2002 –
√ √
– – – –
Parkes et al. 2001 –
√ √
– – – –
Biswas et al. 2003 –
– – – –
Bapna et al. 2006 –/ √ √ √ √ √ √
– – – –
√
Schnizler et al. 2006
√
–
√
– –
Entchev 2003
– – –/ √ √
√ √ √
– –
Engler 2005
– –
–
Legend: R1: online mechanism; R2: bundle allocation; R3: price signals; R4: capacity
constraints; R5: stable operation; R6: coupling with large grid; R7: forecasts
Table 1: Literature Overview
problem of providing precise estimates on future consumption
and production of the grid.
Demanding competition on both sides suggests the
development of a combinatorial exchange. In literature,
Parkes et al. introduce the first combinatorial exchange
as a single-shot sealed bid auction . As payment scheme,
Vickrey discounts are approximated. Biswas and Narahari
propose an iterative combinatorial exchange based
on a primal/dual programming formulation of the allocation
problem. By doing so, the preference elicitation
problem can be alleviated, as the bidders can restrict
their attention to some preferred bundles in contrast
to all 2n1 possible combinations. Obviously, both
approaches do not account for immediacy demands are
thus not directly applicable for the problem at hand.
The MACE mechanism developed by Schnizler et
al. satisfies the bundling requirement and also allows the
modeling of capacity constraints but is NP hard and cannot
account for application scenarios with more than 500
bids and asks within a reasonable time span . Accordingly,
this mechanism can be used for batch applications,
but not for those applications that require immediacy as
is the case in the motivating scenario.
Enchev [11] proposes a different approach in that he
uses a fuzzy logic based optimization algorithm for adjusting
the energy production to a given (and possibly
changing) demand pattern. This approach appears to be
suitable for ensuring stable operation of the micro electricity
grid, but as there are no price signals provided to
indicate the scarcity of resources, consumers do not have
an incentive to optimize there consumption schedules.
Engler et. al. [10] exploit the potentials of an
agent based system for controlling a micro electricity
grid. They show that the system operates stable under
realistic conditions and even recovers automatically from
exogenous shocks (brownouts). Still in this approach
heat is not considered as a resource, furthermore price
signals are missing so that also in this case consumers
do not have an incentive to optimize their consumption.
This paper intends to tailor a mechanism for the
simultaneous allocation of electricity and heat resources
by coupling two separate call markets together [9] so that
it accounts for the aforementioned requirements.
4 Modeling an Auction Based
Energy Allocation Mechanism
The base model
Based on the aforementioned requirements we propose
a double auction mechanism that is capable of coallocating
electricity and heat and additionally takes
minimum and maximum allocation constraints for both
goods into account (if specified by the bidders). For the
description of our model we use the following notation
311

with the superscripts e and h indicating the variable’s
dedication to electricity or heat respectively:
i
j
h i
h j
e i
e j
b h i
a h j
βi
h
βi
h
αj
h
αj
h
x h i
yj
h
Index for buyers i = 1 . . . I
Index for sellers j = 1 . . . J
Quantity of heat requested by buyer i
Quantity of heat offered by seller j
Quantity of electricity requested by buyer i
Quantity of electricity offered by seller j
Maximum price per heat unit
buyer i is bidding
Minimum price per heat unit seller j
is asking for
Market allocation for buyer i‘s bid on heat
Min. heat allocation constraint of buyer i
Market allocation for seller j‘s ask for heat
Minimum allocation constraint of seller j
Binary decision variable for buyer i‘s bid on heat
Binary decision variable for seller j‘s ask for heat
Using this notation we can formulate our market
model as a mixed-integer maximization problem. In particular
the total welfare, i.e. the difference between all
allocated buyers bids b i and all allocated sellers asks
a j is maximized, subject to the resulting solution being
feasible. A solution is feasible only if the consumed
amount of heat is leq the generated amount of heat (1)
and the generated amount of electricity equals the consumed
amount of electricity (2). Furthermore if a buyer’s
bid or a seller’s ask is allocated by the market for either
heat or electricity it also has to be allocated for the other
product, i.e. combinatorial bids on both products cannot
be split by the auctioneer but have to be allocated
simultaneously. Lastly a consumer can set minimum allocation
constraints for heat and / or electricity (5)-(8)
that have to be taken into account as well.
max
s.t.
∑
βi h x h i b h i + βi e x e i b e i − ∑ αj h yj h a h j + αjy e j e a e j
j∈J
i∈I
∑
β h i x h i h i ≥ 0 (1)
αj h yj h h j − ∑
j∈J i∈I
∑
βi e x e i e i = 0 (2)
j∈J
α e jy e j e j − ∑ i∈I
y h j = y e j ∀j ∈ J (3)
x h i = x e i ∀i ∈ I (4)
0 ≤ α e j ≤ α e j ≤ 1 ∀j ∈ J (5)
0 ≤ α h j ≤ α h j ≤ 1 ∀j ∈ J (6)
0 ≤ β e i ≤ β e i ≤ 1 ∀i ∈ I (7)
0 ≤ β h i ≤ β h i ≤ 1 ∀i ∈ I (8)
x h i , x e i ∈ {0, 1} ∀i ∈ I (9)
y h j , y e j ∈ {0, 1} ∀j ∈ J (10)
Assume a seller j who operates a combined heat
and power (CHP) generator. He is able to produce
e j = 10 units of electricity which he wants to sell at
a price of at least a e j = 20 ct/unit. As his generator
always co-generates heat when producing electricity,
j wants to sell the heat (h j = 30 units) simultaneously
at a price of at least a h j 5 ct/unit. Furthermore, as
his generator is not able to operate at a capacity of <
50% of the norm capacity, he wishes to make sure that
– in case his offer is matched – the allocated amount
of electricity is ≥ αj
e = 50% of the originally specified
amount. For the heat on the other hand, seller
j might have a thermae available capable of storing
the produced heat if it cannot be sold, thus αj h = 0.
In summary, the ask of seller j will consist of a tuple
A(e j ; a e j ; αe j ; h j; a h j ; αh j ) = A(10; 20; 50; 30; 5; 0).
Similarly a buyer i might want to buy e i = 10units
of electricity for at most b e i = 25ct/unit but he is
neither interested in heat nor in partial allocations.
Thus buyer i‘s bid will be B(e i ; b e i ; βe i ; h i; a h i ; βh i ) =
A(10; 25; 100; Nil; Nil; Nil).)
In general, combinatorial allocation problems as the
one described above belong to the complexity class of
NP-hard problems but this does not mean that every
possible problem instance will be NP hard. Instead there
usually exist problem instances that can be easily computed
in polynomial time. We will further exploit this
312

particular property for the design of our heuristic in the
extended model described in the following section. For
the example above we can easily solve the allocation
problem by awarding the 10 units of electricity offered
by seller j to consumer i ignoring the heat that j offered
as well (αj h = αj h = 0). Having solved the allocation
problem like this, the pricing for the transaction still
needs to be determined. In general one possible solution
is to employ a Vickrey Clarke Groves (VCG) like
mechanism, which is the only class of mechanisms that
provides allocative efficient, individual rational and incentive
compatible solutions [14]. But as solving VCG
problems is NP-hard and furthermore VCG mechanisms
cannot be budget balanced [19], thus we will use the
maximum execution principle to determine prices. The
idea is described in more detail in the next section.
The extended model
In this section we describe a heuristic that is capable
of efficiently finding near-optimal solutions solving the
aforementioned allocation and pricing problem in polynomial
time by particularly exploiting the fact that in
micro electricity grids, most of the consumers and some
of the producers do not need to allocate electricity and
heat at the same time. In other words, in most of the
cases separate double auctions for heat and electricity
would be sufficient, only e.g. for producers that rely on
CHP generators, submitting bundle offers might be advantageous.
Thus our solution is to set up two slightly adapted
open book call markets [9] running parallel and independently
from each other. Table 2 depicts the orderbooks
of an electricity and a heat call market that are meant
to run in parallel. Call market trading basically means
that incoming orders are collected in an open order book
up to a predefined point in time. Then all orders are executed
in a single multilateral trade at a price that best
matches the aggregated asks and bids. On the buy side,
all bids with limit prices greater or equal to the matching
price are executed and vice versa for the sell side.
If e.g. the ask side of the orderbook is longer, meaning
more ask than buy orders exist at the clearing price,
those ask orders are executed at the same fraction (pro
Our
rata) ensuring for each order that a potential minimum
allocation constraint αj e (αh j ) is not violated.
solution for coupling the two markets together
thus allowing combinatorial bids is simple. Besides the
two orderbooks we use an extra table to keep track of the
combinatorial constraints. In our case Table 3 is used to
record these constraints 4 . The first row states that the
order with order id oid e = 1 and the one with order id
oid h = 2 have to be either executed together or not at
all (AND order). Now, when the time for matching the
markets is reached, each of the markets determines its
optimal set of matched orders independently from the
other market. In this case we extended the previous
example. The electricity market is matched as before
but for the heat market we now have an additional ask
(oid 4) and an additional buy order (oid 5). With the
only bidder in this auction offering to buy heat at most
6 ct per unit and two potential sellers offering 10 units
of heat for at least 4 ct/unit and 30 units for at least
5 ct/unit respectively, the price will be set to 5 ct/unit
in order to maximize the exchange volume. As a result,
seller j = 3 sells 10 units of heat, seller j = 1 sells 20
units of heat, and buyer i = 4 buys 30 units of heat in
this auction, each for 5 ct/unit.
As soon as this task is accomplished both markets
indicate the results of the matching in the bundling constraint
table. Orders for which the bundling constraint is
violated are then deleted from both, the order books and
the bundling constraint table before the market matching
is executed again. This procedure is repeated until
no violations occur anymore. In our case, the bundling
constrained is not violated, thus the matching process is
completed.
A controller that implements the mechanism as described
so far would already fulfill requirements R2 (bundle
allocation), R3 (price signals), and R4 (capacity constraints)
but, as stated before, this is not sufficient for
operating a micro CHP grid. In particular, this mechanism
does not describe when the allocated resources
need to be available and for how long. We overcome this
limitation by discretizing time into distinct, consecutive
slots of a predefined duration d, e.g. d = 15min. Then
n bundle call auctions are set up similar futures markets
at the stock exchanges one for each of the next n time
slots (n being sufficiently large). We adjust the auctions’
execution time intervals intervals t such that t ≤ d. Also
we ensure that a future market is closed well timed be-
4 Non-combinatorial orders go directly into the respective markets
and are thus not recorded in this table.
313

Orderbook Electricity
Orderbook Heat
Ask Bid Ask Bid
oid j e j a e j αj e oid i e i b e i βi e oid j e j a e j αj e oid i e i b e i βi
e
1 1 10 20 5 3 2 10 25 1 4 3 10 4 .2 5 4 30 6 1
2 1 30 5 nil
Table 2: Sample Orderbooks
oid e oid h type match e match h
1 2 AND true true
Table 3: Bundling Constraints
fore the underlying time slot starts. Like this all market
participants can submit several orders to the respective
market adjusting their combined electricity and heat demands
and supplies for the next n time slots.
Setting up our mechanism like this has several advantages.
Market prices on the respective markets indicate
the energy scarcity in the respective time slot and
thus help users decide whether to use (or not to use)
energy in certain time slots. In case a user decides to
change his consumption schedule he can then immediately
submit orders to sell energy in case of excess capacities
or to buy energy in case of increased demand. Thus
micro grid particpants can continuously adjust their demand
and supply profiles over time, which in turn meets
our requirement R1 (online mechanism).
The two requirements R6 (coupling with large-scale
power grid) and R7 (forecasts for demand and supply)
are now relatively easy to satisfy by introducing a special
market participant namely, a so-called arbitrage agent.
Located at the coupling point (c.f. Figure 1 (4)) of the
microgrid this agent can buy electricity from the largescale
grid (e.g. at the EEX) and immediately sell it
inside the microgrid markets at a higher price if electricity
is cheaper in the outside world and vice versa. On
average this agent (i) levels market prices in both markets,
(ii) ensures that generators in the microgrid are
only turned on if they can be operated at competitive
costs, and (iii) communicates the future aggregated energy
demand (supply) of the microgrid to the large-scale
grid through placing his orders on the outside market.
With the arbitrage agent in place our mechanism
now fulfills all requirements but one for operating a micro
energy: Stable operation (R5)). We claim that one
average markets are very good of matching demand and
supply. But in particular on average means not always.
We already described the procedure for allocating energy
resources in case of unbalanced orderbooks, which
basically condenses to rationing one side of the market.
In case supply is rationed, nothing much will happen besides
some generators being operated at a reduced workload
or even being turned off. But in case of demand rationing
this means at worst that energy production and
consumption become unbalanced and thus a brownout
or a blackout might occur.
In order to cope with this challenge we need to introduce
another (market) institution. Consider the point
in time where a market will be closed because the underlying
timeslot starts. At this moment energy trading
becomes impossible for the market participants leaving
them with the risk that they can no longer intervent (i.e.
buy or sell) in case predicted and actual demand (supply)
diverge due to unforeseen events (e.g. a defect generator).
In such a situation the whole micro grid, especially
the electricity subgrid, is likely to become unstable and
will eventually collapse.
This is where the spinning reserves (c.f. Figure 1
(3)) take over control as emergency coordinators. Spinning
reserves do not trade their energy capacities on the
above mentioned future markets instead they provide
balancing power on demand throughout the currently
running time slot. In other words, the spinning reserves
are responsible to level out energy (i.e. heat and electricity)
demand and supply on the spot beyond the market’s
capabilities ensuring a stable operation of the micro grid
and thus fulfilling our last requirement R5.
In order to set an incentive for regular consumers
and producers not to rely on the spinning reserves, the
effectively provided balancing power (heat) is charged
314

ex post to those market participants who originated the
market imbalance by deviating from their originally negotiated
energy consumption (production) level. The
spinning reserves are price sufficiently above the market
price in order to provide market participants with a
strong incentive to negotiate exactly the amount of energy
that they are going to consume (produce) during a
certain time slot. In other words, truth revelation becomes
a dominant strategy for the market participants
the shorter the remaining time until the start of a slot
becomes.
5 Conclusion & Outlook
In this paper we describe a market based system for coordinating
combined heat and power micro energy grids.
We use bundled open book call markets to let consumers
and producers of the microgrid negotiate the allocation
their future energy (i.e. heat and electricity) demand
and supply. Furthermore we propose the utilization of
an arbitrage agent for connecting the micro (electricity)
grid to the large-scale grid. We also describe a way to
employ spinning reserves for stabilizing the micro grid
on the spot without inferencing the principal of market
based coordination.
Overall the approach for operating micro energy
grids described in this paper is very interesting as it proposes
several new research avenues:
• The efficiency of the mechanism introduced in this
paper needs to be further evaluated, especially the
average gap between optimal allocations and those
allocations generated by the heuristic needs to be
determined in more detail.
• Subsequently, agent based simulations could be used
to further study the system’s dynamics, e.g. its reactivity
to exogenous shocks such as a sudden generator
breakdown [15].
• Electronic agents need to be developed that automate
the process of energy trading. Here a clear
advantage of our model is its simplicity. The fact
that e.g. consumer agents can be implemented for
procuring only heat or electricity reduces complexity
and allows the recource to related literature from
algorithmic trading research.
• Trading strategies for different types of producers
(i.e. wind turbines vs. micro CHP generator) and
consumers need to be developed.
• Innovative preference elicitation methods are required
in order to allow e.g. consumers to only specify
target temperatures for certain times of the day
thus leaving the task of determining the resulting
optimal heat consumption profile to the agents.
• Different energy prices for different timeslots set an
incentive for demand side optimization. In this area,
new models need to be developed that support consumers
in finding optimal or at least more efficient
consumption schedules without reducing their quality
of life.
• The potentials and dynamics of hierarchically organized
energy grids as well as potential yo-yo effects,
which might occur if a large number of microgrids
are coupled together hierarchically, need further investigation.
6 Acknowledgement
This work was supported in part by the German Federal
Ministry of Education and Research (SESAM - Selforganization
and Spontaneity in Liberalized and harmonized
Markets, Grant 01-AK-703).
References
[1] S. Abu-Sharkh, R. J. Arnold, J. Kohler, R. Li, T. Markvart,
J. N. Ross, K. Steemers, P. Wilson, and R. Yao.
Can microgrids make a major contribution to uk energy
supply? Renewable and Sustainable Energy Reviews,
10(2):78–127, April 2006.
[2] R. Bapna. When snipers become predators: Can mechanism
design save online auctions? Communications of
the ACM, 46(12):152–158, 2003.
[3] S. Biswas and Y. Narahari. Iterative dutch combinatorial
auctions. Annals of Mathematics and Artificial
Intelligence, page forthcoming, 2003.
[4] B. Buchholz and U. Schluecking. Energy management
in distribution grids european cases. Power Engineering
Society General Meeting, 2006. IEEE, pages 2 pp.+,
2006.
315

[5] R. Buyya, H. Stockinger, J. Ghiddy, and D. Abramson.
Economic models for management of resources in peerto-peer
and grid computing. In International Conference
on Commercial Applications for High Performance
Computing, Denver, CO, 2001.
[6] D. Coll-Mayor, M. Paget, and E. Lightner. Future intelligent
power grids: Analysis of the vision in the european
union and the united states. Energy Policy,
35(4):2453–2465, April 2007.
[7] W. Conen. Economically coordinated job shop scheduling
and decision point bidding - an example for economic
coordination in manufacturing and logistics. In Workshop
on Planen, Scheduling und Konfigurieren, Entwerfen,
Freiburg, 2002.
[8] M. Conrad, J. Dinger, H. Hartenstein, D. Rolli,
M. Schller, and M. Zitterbart. A peer-to-peer framework
for electronic markets. In R. Steinmetz and K. Wehrle,
editors, Peer-to-Peer Systems and Applications, volume
Lecture Notes in Computer Science 3485. Springer,
2005.
[9] N. Economides and R. Schwartz. Electronic call market
trading. Journal of Portfolio Management, 21(3):10–18,
1995.
[10] A. Engler, C. Hardt, J. Jimeno, J. Ruela, and J. Oyarzabal.
Agent based micro grid management system.
In International Conference on Future Power Systems,
November 2005.
[11] E. Entchev. Residential fuel cell energy systems performance
optimization using ”soft computing” techniques.
Journal of Power Sources, 118(1-2):212–217, May 2003.
[12] T. Eymann, S. Sackmann, and G. Mueller. Hayeks
katallaxie - ein zukunftsweisendes konzept fuer
die wirtschaftsinformatik. Wirtschaftsinformatik,
45(5):491–496, 2003.
[13] P. Gerling, H. Rempel, U. Schwarz-Schampera, and
T. Thielemann. Reserven, ressourcen und verfügbarkeit
von energierohstoffen. Technical report, Bundesanstalt
für Geowissenschaften und Rohstoffe, 2005.
[14] J. Green and J.-J. Laffont. Characterization of satisfactory
mechanisms for the revelation of preferences for
public goods. Econometrica, 45(2):427.438, 1977.
[15] Haque, Nadim, Jennings, Nicholas, Moreau, and Luc.
Scalability and robustness of a network resource allocation
system using market-based agents. Netnomics,
7(2):69–96, August 2005.
[16] A. Klein. Comparative costs of energy in europe 2007.
Clean Power Generation Advisory CPG 725-070110,
Emerging Energy Research, January 2007.
[17] W. Krewitt and B. Schlomann. Externe kosten der
stromerzeugung aus erneuerbaren energien im vergleich
zur stromerzeugung aus fossilen energietraegern. Technical
report, Deutsches Zentrum fuer Luft- und Raumfahrt
and Fraunhofer Institut fuer System- und Innovationsforschung,
2006.
[18] R. Lasseter, A. Akhil, C. Marnay, J. Stevens, J. Dagle,
R. Guttromson, S. A. Meliopoulous, R. Yinger, and
J. Eto. The certs microgrid concept - white paper on
integration of distributed energy resources. Technical
report, U.S. Department of Energy, April 2002.
[19] R. B. Myerson and M. A. Satterthwaite. Efficient mechanisms
for bilateral trading. Journal of Economic Theory,
28:265–281, 1983.
[20] D. C. Parkes, J. Kalagnanam, and M. Eso. Achieving
budget-balance with vickrey-based payment schemes
in exchanges. In Proceedings of the 17th International
Joint Conference on Artificial Intelligence, pages 1161–
1168, 2001.
[21] O. Regev and N. Nisan. The popcorn market - an
online market for computational resources. In First
international conference on Information and computation
economies, Charleston, South Carolina, 1998. ACM
Press.
[22] B. Schnizler, D. Neumann, D. Veit, and C. Weinhardt.
Trading grid services - a multi-attribute combinatorial
approach. European Journal of Operational Research,
In Press, Corrected Proof:forthcoming, 2006.
[23] K. Subramoniam, M. Maheswaran, and M. Toulouse.
Towards a micro-economic model for resource allocation
in grid computing systems. In Subramoniam, editor,
Proceedings of the 2002 IEEE Canadian Conference
on Electrical & Computer Engineering, volume 2, pages
782–785 vol.2, 2002.
[24] V. VDEW. Stromverbrauch der privaten haushalte –
zehn prozent des stroms für kommunikation und information.
Online Article, May 2006.
[25] J. A. Villar and F. L. Joutz. The relationship between
crude oil and natural gas prices. Technical report, Energy
Information Administration, 2006.
[26] M. P. Wellman, W. E. Walsh, P. Wurman, and
J. MacKie-Mason. Auction protocols for decentralized
scheduling. Games and Economic Behavior, 35(271-
303):271–303, 2001.
[27] R. Wolski, J. Brevik, J. Plank, and T. Bryan. Grid
resource allocation and control using computational
economies. In Grid Computing - Making The Global
Infrastructure a Reality, page chapter 32. John Wiley &
Sons, 2004.
316

Lebenslauf
Dr. Dirk Neumann, M.A. (Univ. Wisc) Tel. (privat) 0721-6238678
Rintheimer Str. 78 Tel. (univ.) 0721-6088377
76131 Karlsruhe Email: neumann@kit.edu
Geboren am 15. März 1974
Geboren in Frankfurt/Main
Familienstand ledig
1. Ausbildung
15.07.2004 Promotion zum Dr. rer. pol. in Wirtschaftswissenschaften (summa
cum laude)
Universität Karlsruhe (TH) Fakultät für Wirtschaftswissenschaften
Thema der Dissertation „Market Engineering – A Structured
Design Process for Electronic Markets”. Doktorvater: Prof. Dr.
Christof Weinhardt
08.06.2000 Diplom in Ökonomie an der Justus-Liebig Universität Gießen
(Note: 1,3: 1. von 83)
16.05.1999 Master of Arts (M.A.) in Economics an der University of Wisconsin,
Milwaukee (GPA: 3,97/4,00)
18.08.1998 Immatrikulation für den Masterstudiengang Economics
an der University of Wisconsin, Milwaukee
29.07.1996 Vordiplom in Wirtschaftswissenschaften
an der Justus-Liebig Universität, Giessen
03.10.1994 Immatrikulation für den Diplomstudiengang Wirtschaftswissenschaften
an der Justus-Liebig Universität, Giessen
14.06.1993 Abitur an dem katholischen Privatgymnasium St. Lioba
in Bad Nauheim, Leistungskurse Mathematik und Physik
August 1984 St. Lioba Gymnasium, Bad Nauheim
August 1980 Frauenwald Grundschule, Bad Nauheim, Nieder-Mörlen
317

2. Berufserfahrung
1.10.2004 – heute Wissenschaftlicher Assistent (C1) am Lehrstuhl für Informationsbetriebswirtschaftslehre
der Universität Karlsruhe (TH)
1.6.2000–30.9.2004 Wissenschaftlicher Mitarbeiter
2001 – 2004 Wissenschaftlicher Mitarbeiter (BAT IIa) am Lehrstuhl
für Informationsbetriebswirtschaftslehre bei Prof. Christof
Weinhardt, Fakultät für Wirtschaftswissenschaften Universität
Karlsruhe (TH)
2000 – 2001 Wissenschaftlicher Mitarbeiter (BAT IIa) am Lehrstuhl
für BWL-Wirtschaftsinformatik bei Prof. Christof Weinhardt,
Fakultät für Wirtschaftswissenschaften Justus Liebig Universität
Giessen
1996 – 2000 Wissenschaftliche Hilfskraft
1999 – 2000 Stud. Hilfskraft Tutor und Übungsgruppenleiter für
Wirtschaftsinformatik (Programmierung Visual Basic), Prof.
Weinhardt, Universität Gießen
1996 – 1998 Stud. Hilfskraft Tutor und Übungsgruppenleiter für
Wirtschaftsinformatik (Programmierung Visual Basic), Prof.
Weinhardt, Universität Gießen
1993 – 1994 Grundwehrdienst in Gotha und Stadtallendorf
Panzeraufklärungseinheit
318

3. Eingeladene Vorträge
• Entwicklungsperspektiven des Grid Computing, im Rahmen des 7.@kit-Kongresses
"Vernetztes Rechnen – Softwarepatente – Web 2.0" 2007, Potsdam, 21.06.2007.
• A Decision Support System for Designing Auctions, im Rahmen des Research Symposiums
"Hot Topics in Negotiation Support Systems" an der HICSS Conference 2006,
Po’ipu, Hawaii, 04.01.2006.
• A Framework for Assessment of e-Negotiation Systems – The Case of Online Auctions
and E-negotiations, im Rahmen des Research Symposiums "Hot Topics in Negotiation
Support Systems" an der HICSS Conference 2006, Po’ipu, Hawaii, 04.01.2006.
• Computer Aided Market Engineering, im Rahmen der Vortragsreihe der John Molson
School of Business in Montreal 2005, Montreal, Kanada, 12.10.2005.
• P2P – Technical and Economic Foundations, Tutorial zusammen mit Prof. Dr. Klemens
Böhm im Rahmen der 7th International IEEE Conference on E-Commerce Technology,
München 21.07.2005.
• Market Engineering – Wissensbasierte Unterstützung der Konzeptphase, E-Commerce
Kompetenzzentrum Abschluss-Kolloquium, Dresden, 18.11.2004.
• A Market Engineering – Knowledge Based Auction Design Markets, negotiations and
dispute resolution in the new economy, Montreal, Canada, October 12-13, 2004, Montreal,
Kanada, 12.10.2004.
• Market Engineering and Auction Design: Pricing Per Column (PPC), im Rahmen der
CORS/INFORMS Tagung in Banff 2004, Banff, Kanada, 16.05.2004.
• eNegotiation –Taxonomy and Design, im Rahmen des IFIP Workshops “B2B Markets –
Beyond MRO” in Zürich 2001, Zürich, Schweiz, 03.10.2001.
• The London Classification Scheme – An Application on Commercial eNegotiations, im
Rahmen des eNegotiation Workshops in Montreal 2001, Montreal, Kanada, 05.08.2001.
4. Lehre
Vorlesungen
• Vorlesung “Grundzüge der Informationswirtschaft”, im Wintersemester 2005/2006
Pflichtvorlesung für Informationswirte, ca. 80 Studierende.
• Vorlesung “Programmieren in Visual Basic”, im Wintersemester 2001/2002 (6 Termine),
Justus-Liebig-Universität Gießen
Pflichtvorlesung im Grundstudium für Wirtschaftswissenschaftler, ca. 300 Studierende.
• Vorlesung “Information and Market Engineering” (engl.), im Sommersemeter 2006
– Vorlesung im Rahmen des Executive Master Programs „Information Engineering“ der
Hector School
• Vorlesung “Management of Business Networks” (engl.), im Wintersemester
2004/2005 – Vorlesung enthält eine Fallstudie, die mit den kanadischen Partnern der
Concordia University abgehalten wird.
Wahlpflichtvorlesung, ca. 15 Studenten.
• Vorlesung “Market Engineering” (engl.), im Sommersemester 2005
Wahlpflichtvorlesung, ca. 20 Studenten.

• Vorlesung “Market Engineering” (engl.), im Sommersemester 2006
Wahlpflichtvorlesung, ca. 25 Studenten.
• Vorlesung “Market Engineering” (engl.), im Sommersemester 2007
Wahlpflichtvorlesung, ca. 25 Studenten.
Seminare
• eNegotiation I, SS 2001
• eNegotiation II, WS 2001/2002
• Management of Business Networks, SS 2005
• Peer-to-Peer Economics, SS 2005
• Grid Economics, WS 2005/2006
• Mechanismen in verteilten Netzen (Mechanisms in Distributed Networks), in Kooperation
mit Prof. Dr. Klemens Böhm, WS 2005/2006
• Information Pricing, SS 2006
• Supply Chain Management & Auktionen, SS 2006
• Service Science. Management and Engineering, WS 2006/2007
• Computer Online Game in Finance – Entwicklung eines browserbasierten Wirtschaftssimulationsspiels,
WS 2006/2007
• Automatisierte Verhandlungen und Algorithmen, SS 2007
• Energie und Märkte, WS 2007/2008
• Mobilfunk und Telekommunikation, WS 2007/2008
Tutorien Vorlesungen
• Tutorium (50 Teilnehmer) zur Vorlesung „Grundzüge der Informationswirtschaft“,
Universität Karlsruhe (TH), Prof. Weinhardt, Wintersemester 2000/2001
• Tutorium (50 Teilnehmer) zur Vorlesung „Grundzüge der Informationswirtschaft“
(Hälfte der Termine), Universität Karlsruhe (TH), Prof. Weinhardt, Wintersemester
2001/2002
• Tutorium (50 Teilnehmer) zur Vorlesung „Grundzüge der Informationswirtschaft“
(Hälfte der Termine), Universität Karlsruhe (TH), Prof. Weinhardt, Wintersemester
2002/2003
• Tutorium (15 Teilnehmer) zur Vorlesung „Koordination und Netzwerkeffekte im
eBusiness“, Universität Karlsruhe (TH), Prof. Weinhardt, Sommersemester 2002
• Tutorium (15 Teilnehmer) zur Vorlesung „Koordination und Netzwerkeffekte im
eBusiness“, Universität Karlsruhe (TH), (Hälfte der Termine) Prof. Weinhardt, Sommersemester
2003
• Zwei Tutorien (je 25 Teilnehmer) zur Vorlesung „Wirtschaftsinformatik I“, Justus-
Liebig-Universität Gießen, Prof. Weinhardt, Wintersemester 1996/1997
• Drei Tutorien (je 25 Teilnehmer) zur Vorlesung „Wirtschaftsinformatik I“, Justus-
Liebig-Universität Gießen, Prof. Weinhardt, Wintersemester 1997/1998
• Zwei Tutorien (je 25 Teilnehmer) zur Vorlesung „Wirtschaftsinformatik I“, Justus-
Liebig-Universität Gießen, Prof. Weinhardt, Wintersemester 1999/2000
320

• Ein Tutorium (19 Teilnehmer) zur Vorlesung „Einführung in die BWL“, Justus-
Liebig-Universität Gießen, Prof. Glaum, (3 Termine) Wintersemester 2000/2001
Betreute Diplomarbeiten
• Dr. Bjoern Schnizler, Iterative Kombinatorische Auktionen – iBundle vs. SAAPB. Abgabe
im Dezember 2003 (Von 2004-2007 Assistent am Lehrstuhl Prof. Weinhardt), Note:
1,3.
• Thorsten Bendrich, Computational Mechanism Design im Umfeld kombinatorischer
Auktionen. Abgabe im Januar 2004, Note: 2,0.
• Dr. Henner Gimpel, Auktionsverfahren im Aktienprimärmarkt – Experimentelle Evaluierung
der Informationseffizienz. Abgabe im Mai 2003. Ausgezeichnet mit dem 1.
Hochschulpreis 2003 des Deutschen Aktieninstituts (DAI) in der Kategorie Diplomarbeiten.
(Von 2003-2007 Assistent am Lehrstuhl Prof. Weinhardt), Note: 1,0.
• Dr. Daniel Rolli, Eine dynamische Workflow-Standard-Erweiterung im Hinblick auf die
Steuerung elektronischer kombinierter Verhandlungen. Abgabe im März 2003 (von
2003-2006 Assistent am Lehrstuhl Prof. Weinhardt), Note: 1,3.
• Tunc Berdan, Auctions versus Negotiations in Procurement., an der Concordia University,
Montreal in Kooperation mit Gregory Kersten. Abgabe im März 2006, Note: 1,3.
• Christoph Goebel, Securing Interdomain Routing: An Incentive Analysis, an der Carnegie
Mellon University, Pittsburg in Kooperation mit Ramayya Krishnan. Abgabe im
Juli 2006. (Seit 10/2006 Assistent am Lehrstuhl von Prof. Dr. Günther, Humboldt Universität
Berlin), Note: 1,0.
• Jochen Stoesser, Generic Market Platforms: Adaption and Application of meet2trade to
Emissions Trading, an der University of Newsouth Wales, Sydney in Kooperation mit
Fethi Rabhi. Abgabe im Juli 2006 (Seit 09/2006 Assistent am Lehrstuhl Prof. Weinhardt),
Note: 1,3.
• Kai Beckhaus, Innovation in Value Chain Management, in Kooperation mit AT Kearney,
Abgabe im März 2007, Note 1,0.
• Sibylle Stacklies, Innovation in Value Chain Management, in Kooperation mit AT
Kearney, Abgabe im April 2007, Note 1,0.
• Christian Wiedenhoff, Einflussgrößen innerhalb der Transportplanung auf eine gesamtkostenminimale
Gestaltung des Logistiksystems Audi Neckarsulm, in Kooperation
mit Audi AG. Abgabe im Januar 2007, Note 1,3.
• Thorsten Hau, An Allocation Scheme for the Retail Business – How Auctions Can Ensure
Allocative Efficiency, in Kooperation mit Lux International. Abgabe im Dezember
2006 (Seit 01/2007 Assistent am Lehrstuhl von Prof. Dr. Brenner, St. Gallen), Note:
1,0.
• Lutz Jansen, Information Management Design Decomposition, in Kooperation mit ETH
Zürich und Forschungszentrum Informatik (FZI), Abgabe im Dezember 2006, Note:
1,3.
• Christine Rössle, Decentralized Online Resource Allocation for Dynamic Web Service
Applications, bei SUN Singapur und Nanyang Technological University in Kooperation
mit Simon See, Abgabe im Juni 2007, Note: 1,3.
• Philipp Bodenbenner, Pay-your-Bid Mechanisms for Resource Allocation in Grids, bei
SUN Singapur und Nanyang Technological University in Kooperation mit Simon
See, Abgabe im Juni 2007, Note: 1,0.
321

• Marius Romanescu, Product delivery to Customers using company-internal Markets,
Abgabe im April 2007, Note: 1,0.
• Benno Schwaiger, Das Similarity-based-Cost-Estimation Verfahren zur Schätzung der
IT-Kosten bei Kooperationen, in Kooperation mit BMW, Abgabe im August 2007.
• Christian Schaub, Efficient Management of Information Resources Over Ad-Hoc Data
Grids, laufend, an der University of Newsouth Wales, Sydney in Kooperation mit Fethi
Rabhi.
• Philip Caune, Evaluating different Market Mechanisms in Grids, laufend, an der Singapure
Management University (SMU), Singapore in Kooperation mit Steven Miller.
• Tim Püschel, Economically Enhanced Resource Management, laufend, am Barcelona
Supercomputing Center (BSC), Barcelona in Kooperation mit Jordi Torres.
• Michael Becker, Risk Management in Grids, laufend, an der Cardiff University, Wales
in Kooperation mit Omer Rana.
• Melanie Mossmann, Combinatorial Auctions in Grids, laufend, bei France Telecom‚
Paris, Frankreich in Kooperation mit Ruby Krishnaswamy.
• Matthias Deindl, Agent-Based Demand Management in Micro Energy Grids, laufend,
an der Concordia University, Montreal in Kooperation mit Rustam Vahidov.
• Tobias Heitfeld, Situated Decision Support Systems for Complex Services, laufend, an
der Concordia University, Montreal in Kooperation mit Rustam Vahidov.
5. Forschungsschwerpunkt und Drittmittel
Aktuelle Forschungsschwerpunkte
• Grid Computing und Internet Economics
• Marktmechanismen in verteilten Netzen
• Peer-to-Peer Economics
• IT-Service-Management
• Business Service Management
• Autonomic Computing
• Virtuelle Währungssysteme
• Incentive Engineering und Service Science
• Anreizgestaltung in kooperativen Systemen und Netzwerken
• Geschäftsmodelle
• Prozessmanagement/ Kontinuierliche Verbesserungsprozesse
• Situated Negotiation Support Systeme
• Supply Chain Management und Market Engineering
• Koordination in Wertschöpfungsnetzen
• eProcurement Mechanismen
• Anwendungen des Operation Research
322

Laufende Forschungsprojekte
1/8 Self-Organizing ICT Resource Management (SORMA)
Förderung
EU FP 6 IST 5th Call als STREP
Förderzeitraum 01.08.2006 – 31.07.2008
Beantragtes 2.700.000,00 €
Gesamtvolumen
Beantragter 609.034,63 €
eigener Anteil
Persönliche Rolle Koordinator/Projektmanager
Projektpartner Forschungszentrum Informatik, Karlsruhe
Cardiff University, Großbritannien
Swedish Institute of Computer Science, Schweden
Research Center at the BFM Bayreuth
Universitat Politècnica de Catalunya, Spanien
University of Reading, Großbritannien
Hebrew University, Israel
Sun Microsystems, Singapur
Correlation Systems, Israel
TXT e-Solutions, Italien
Barcelona Supercomputing Center, Spanien
Kurzbeschreibung Konzeption, Implementierung und Einführung eines Open Grid
Marktes für Grid Ressourcen.
Webseite: http://www.sorma-project.eu/
2/8 Billing the Grid
Förderung
Gefördert von der Landesstiftung Baden-Württemberg als Landesschwerpunktprogramm
Förderzeitraum 01.11.2005 – 31.10.2007
Gesamtvolumen 280.000,00 €
Eigener Anteil 70.000,00 €
Persönliche Rolle
Projektpartner
Kurzbeschreibung
Projektleiter
Universität Karlsruhe (TH)
Konzeption und Implementierung von Zahlungs- und Abrechnungsfunktionalitäten
für Grids im universitären Bereich.
Webseite: http://www.billing-the-grid.org/
3/8 Efficient Management of Information Resources Over Ad-Hoc Data Grids
(ADAGE)
Förderung
DEST-ISL Competitive Grants Programme – Australien
Ergänzungsantrag zu SORMA
Förderzeitraum 01.01.2007 – 31.12.2008
Gesamtvolumen 812.863,00 $
Eigener Anteil 16.000,00 $
Persönliche Rolle
Projektpartner
Workpackage Leader
University of New South Wales, Sydney, Australien
University of Sydney, Australien
SIRCA, Australien
Cardiff University, Großbritannien
University of Lyon, Frankreich
323

Kurzbeschreibung
Vienna University of Technology, Österreich
Das Ergänzungsprojekt zu SORMA zielt darauf ab, effiziente
Methoden und Technologien für ad-hoc Data Grids zu entwickeln.
Webseite: http://cgi.cse.unsw.edu.au/~soc/adage/
4/8 Selbstorganisation und Spontaneität in liberalisierten und harmonisierten Märkten
(SESAM)
Förderung
BMBF-Forschungsschwerpunktprogramm "Internetökonomie"
Förderzeitraum 01.07.2003 – 31.09.2007
Gesamtvolumen 3.063.927,00 €
Eigener Anteil 63.170,34 € (p.a)
Persönliche Rolle Projektverantwortlicher am Institut
Projektpartner ---
Kurzbeschreibung Das interdisziplinäre Projekt “SESAM” bietet Infrastruktur und
ökonomische Mechanismen, um einen dezentralen Energiehandel
zu ermöglichen.
Webseite: http://www.sesam.uni-karlsruhe.de/
5/8 Electronic Negotiations, Media and Transactions for Socio-Economic Interactions
Förderung
Social Sciences and Humanities Research Council of Canada
(SSHRC)
Förderzeitraum 01.04.2002 – 30.10.2007
Eigener Anteil ca. 12.000,00 €
Persönliche Rolle Projektverantwortlicher am Institut
Projektpartner Concordia University Montreal, Kanada
University of Ottawa, Kanada
McMasters University, Hamilton, Kanada
University of Hawaii, USA
TU München
Universität Wien, Österreich
University of Taiwan, Taiwan
Kurzbeschreibung Analyse der Auswirkung von e-Negotiation-Protokollen auf das
Verhalten menschlicher Verhandlungsführer. Design geeigneter
e-Negotiation-Protokolle in verschiedenen Szenarien.
Webseite: http://interneg.concordia.ca/enegotiation/index.html
6/8 Georgia Tech Austauschprogramm
Förderung
ISAP Programm, DAAD
Förderzeitraum 21.08.2006 – 03.08.2007
Gesamtvolumen ca. 30.650,00 €
Eigener Anteil ca. 30.650,00 €
Persönliche Rolle Projektleiter
Projektpartner Georgia Institute of Technology, Atlanta
Kurzbeschreibung Studentenaustausch mit dem Computer Science Department der
Georgia Tech. Jährlicher Austausch von 3 Karlsruher Studenten
im Master Programm.
Webseite:
http://www.wiwi.uni-
324

karlsruhe.de/studium/auslandstudium/bewerbung.html
7/8 CatNets
Förderung
EU IST FET Open als STREP
Förderzeitraum 01.09.2004 – 30.08.2007
Gesamtvolumen 1.414.000,00 €
Eigener Anteil 284.275,00 €
Persönliche Rolle Projektmitarbeiter
Projektpartner Universität Bayreuth
Universitat Politecnica de Catalunia, Spanien
Universita delle merche Ancona, Italien
University of Cardiff, Großbritannien
ITC-irst Treno, Italien
Universität Mannheim
Kurzbeschreibung Untersuchung von dezentralen versus zentralen Marktmechanismen
in Grid Märkten.
Webseite: http://www.catnets.org/
8/8 Moving Business to the Grid – An Application for the Automotive Industry
(Biz2Grid)
Förderung
BMBF Projekt
Förderzeitraum 01.06.2007 – 31.05.2010
Gesamtvolumen 1.180.000,00 €
Eigener Anteil 401.052,00 €
Persönliche Rolle Projektmitarbeiter/Antragsteller
Projektpartner Forschungszentrum Informatik, Karlsruhe
BMW München
IBM Deutschland GmbH
Philipps-Universität Marburg
Webseite: http://www.biz2grid.de/
6. Gutachtertätikeiten
Begutachtung von Beiträgen der Zeitschriften:
• Concurrency and Computation: Practice and Expertise
• International Journal of Electronic Markets
• Journal of Grid Computing
• Journal of Group Decision and Negotiation
• Journal of Managament Information Systems (JMIS)
• Journal of Organizational Computing and Electronic Commerce
• Wirtschaftsinformatik
Begutachtung von Beiträgen der Konferenzen und Workshops:
• Agent-based Grid Computing Workshop
• Cooperative Information Systems
• EC-Web Conference
325

• ECIS
• EuroSys
• Gecon Workshop
• Grid Economics Workshop
• HICSS
• Multi-Konferenz Wirtschaftsinformatik
• Wirtschaftsinformatik Tagung
• Workshop on eBusiness – ICIS Pre-Conference
Enagagement in internationalen Veranstaltungen
• Chair des Workshops "Economic Models for Distributed System", im Rahmen der Mardigras
Conference 2008, Baton Rouge, Louisiana.
• Mitglied im Programm-Kommittee an der Konferenz "EC Web Conference", 2008
• Mitglied im Programm-Kommittee der Teilkonferenz "Betriebswirtschaftliche Anwendungen
des Grid Computing" im Rahmen der Multikonferenz Wirtschaftsinformatik
2008 (MKWI '08), in München, 26.-28. Februar 2008.
• Mitglied im Programm-Kommittee der Teilkonferenz "Märkte und Geschäftsmodelle
für Service-orientiertes Computing und Grid-Netzwerke" im Rahmen der Multikonferenz
Wirtschaftsinformatik 2008 (MKWI '08), in München, 26.-28. Februar 2008.
• Mitglied im Programm-Kommittee an der Konferenz “Group Decision and Negotiation”,
Coimbra, Portugal, 2008.
• Chair des Workshops "Economic Models and Algorithms in Grid System", im Rahmen
der Grid 2007, Austin, Texas.
• Tutorial Chair an der Wirtschaftsinformatik Konferenz WI 2007 “eOrganisation: Service-,
Process-, Market-Engineering” Karlsruhe, 2007.
• Track Chair der Session “Autonomous Bidding Agents for Auctions and Negotiations”
im Rahmen der Group Decision and Negotiation Konferenz (GDN07) in Montreal.
• Mitglied im Programm-Kommittee am Workshop "Grid Economics and Business Models"
(Gecon 2007), Rennes, Frankreich, 2007.
• Mitglied im Programm-Kommittee an der Konferenz “Group Decision and Negotiation”,
Montreal, Kanada, 2007.
• Mitglied im Programm-Kommittee am Workshop "FinanceCom" im Rahmen der ICIS,
Montreal, Kanada 2007.
• Invited Speaker für das Tutorial “P2P – Technical and Economic Foundations” zusammen
mit Prof. Dr. Klemens Böhm im Rahmen der 7th International IEEE Conference
on E-Commerce Technology, München 21. Juli 2005.
• Mitglied des CT1 Technical Collaboration Steering Committee der Europäischen
Kommission für den Bereich Grid Computing.
• Mitglied im Programm-Kommittee an der Konferenz "EC Web Conference", 2007
326

• Mitglied im Programm-Kommittee am Workshop "Agent-based Grid Computing" for
7 th IEEE International Symposium on Cluster Computing and the Grid (CCGrid07) in
Rio de Janeiro, Mai 2007.
• Mitglied im Programm-Kommittee der Konferenz "Group Decision and Negotiation“
(GDN07) in Montreal.
• Mitglied im Programm-Kommittee am Workshop "Agent-based Grid Computing" im
Rahmen der 6 th IEEE International Symposium on Cluster Computing and the Grid
(CCGrid06) in Singapore, 16. – 19. Mai 2006.
• Mitglied im Programm-Kommittee am Track "Business Applications of P2P and Grid
Computing" for Multikonferenz Wirtschaftsinformatik 2006 (MKWI '06), in Passau, 20.
– 22. Februar 2006.
• Mitglied im Programm-Kommittee am "4th Workshop on Grid Economics & Business
Models for Grid” im Rahmen der Grid Asia Konferenz 2006, in Singapore, 16. Mai
2006.
• Mitglied im Programm-Kommittee der Konferenz "EC Web Conference", 2006
• Mitglied im Programm-Kommittee am Workshop "Smart Grid Technologies“ (SGT06),
Dublin, Ireland,
• Eingeladener Teilnehmer für das Dagstuhl Seminar “Electronic Market Design” im Juli
2002. These workshops consist of gatherings of leading experts in one research area for
a week at the Dagstuhl Castle in Germany. Participation is by invitation only.
• Eingeladener Teilnehmer für das Dagstuhl Seminar “Market and Negotiation Engineering”
im November 2006. These workshops consist of gatherings of leading experts in
one research area for a week at the Dagstuhl Castle in Germany. Participation is by invitation
only
• Organisation des Habilitanden Workshop im November 2005, in Schloss Rauischholzhausen
Schriftenverzeichnis
Beiträge in referierten Fachzeitschriften
• D. Neumann, J. Stoesser, C. Weinhardt (2008): Briding the Adoption Gap – Developing
a Roadmap for Trading Grids, Electronic Markets – The International Journal, 18(1),
(akzeptiert am 26. August 2007), forthcoming.
• B. Schnizler, D. Neumann, D. Veit, Ch. Weinhardt: A Multi-attribute Combinatorial
Exchange for Trading Grid Services, European Journal of Operational Research
(EJOR), (akzeptiert am 04. November 2006), forthcoming
• C. Weinhardt, D. Neumann, C. Holtmann (2006) Computer Aided Market Engineering,
Communications of the ACM, 49(7): 79.
• D. Neumann, C. Orwat, C. Holtmann (2006) Grid Economics, Wirtschaftsinformatik,
48(3): 206-209
• Ch. Weinhardt, B. Weber, D. Neumann, T. Hendershott, R. Schwartz (2006) Financial
Market Engineering, Electronic Markets – The International Journal, 16(2), 98-100
327

• D. Rolli, M. Conrad, D. Neumann, C. Sorge (2006) Distributed Ascending Proxy Auction
– A Cryptographic Approach, Wirtschaftsinformatik, 48(1): 7-15.
• T. Eymann, O. Ardaiz, M. Catalano, P. Chacin, I. Chao, F. Freitag, M. Gallegati, G.
Giulioni, L. Joita, L. Navarro, D. Neumann, O. Rana, M. Reinicke, R. Schiaffino, B.
Schnizler, W. Streitberger, D. Veit, F. Zini (2006), Catallaxy-based Grid Markets, Multi-Agent
Systems and Grid, Special Issue on Smart Grid Technologies & Market Models,
1(4): 297 - 307
• Ch. Weinhardt, C. Holtmann, D. Neumann (2003) Market Engineering, Wirtschaftsinformatik
45(6): 635-640
• D. Neumann, M. Benyoucef, S. Bassil, J. Vachon (2003) Applying the Montreal Taxonomy
to State of the Art E-Negotiation Systems, Group Decision and Negotiation
12(4): 287-310
• C. Lattemann, D. Neumann (2002) Clearing und Settlement im Wandel – Eine Perspektive
für den europäischen Wertpapierhandel, Zeitschrift für die gesamte Kreditwirtschaft:
1159-1164.
• M. Budimir, C. Holtmann, D. Neumann (2001) Design of a Best Execution Market,
Revue Bancaire et Financière/Bank en Financiewezen: 66(2-3): 138-146.
Monographien
• D. Neumann (2007) Market Engineering - A Structured Design Process for Electronic
Markets, in Universitätsverlag Karlsruhe
Referierte Konferenzbeiträge
1. J. Stoesser, D. Neumann, C. Weinhardt, Market-Based Pricing in Grids: On Strategic
Manipulation and Computational Cost, Journal of AIS Sponsored Theory Development
Workshop, Montreal, Canada.
2. G. Kersten, E. Chen, D. Neumann, R. Vahidov, Technology Assessment and Comparison:
The Case of Auction and E-negotiation Systems, Journal of AIS Sponsored Theory
Development Workshop, Montreal, Canada.
3. R. Vahidov, D. Neumann, Situated Decision Support for Service Level Agreement Negotiations,
Hawaii International Conference on System Sciences (HICSS 2008), 7-10
Januar 2008, Waikoloa, Hawaii.
4. J. Stoesser, P. Bodenbenner, S. See, D. Neumann, A Discriminatory Pay-as-Bid Mechanism
for Efficient Scheduling in the Sun N1 Grid Engine, Hawaii International Conference
on System Sciences (HICSS 2008), 7-10 Januar 2008, Waikoloa, Hawaii.
5. C. Block, D. Neumann, C. Weinhardt, A Market Mechanism for Energy Allocation in
Micro-CHP Grids, Hawaii International Conference on System Sciences (HICSS 2008),
7-10 Januar 2008, Waikoloa, Hawaii.
6. M. Becker, N. Borissov, V. Deora, O. Rana, D. Neumann, Using k-pricing for Penalty
Calculation in Grid Markets, , Hawaii International Conference on System Sciences
(HICSS 2008), 7-10 Januar 2008, Waikoloa, Hawaii.
328

7. R. Riordan, B. Blau, C. Weinhardt, D. Neumann, Collaborative Continuous Service
Engineering, Hawaii International Conference on System Sciences (HICSS 2008), 7-10
Januar 2008, Waikoloa, Hawaii.
8. T. Püschel, N. Borissov, M. Macias, D. Neumann, J. Guitart, J. Torres, Economically
Enhanced Resource Management for Internet Service Utilities, The 8th International
Conference on Web Information Systems Engineering (WISE 2007), 3-7 Dezember
2007, Nancy, France.
• J. Stoesser, D. Neumann, GreedEx – A Scalable Clearing Mechanism for Utility Computing,
Networking and Electronic Commerce Research Conference (NAEC) 2007, 18-
21 October 2007, Riva Del Garda, Italy.
• L. Amar, J. Stoesser, A. Barak, D. Neumann, Economically Enhanced MOSIX for Market-based
Scheduling in Grid OS, Workshop on Economic Models and Algorithms for
Grid Systems (EMAGS) 2007, 19 - 21 September 2007, Austin, USA.
• T. Püschel, N. Borissov, D. Neumann, M. Macías, J. Guitart, J. Torres, Extended Resource
Management Using Client Classification and Economic Enhancements, eChallenges
e-2007 Conference, 24-26 October 2007, The Hague, Netherlands.
• J. Stoesser, C. Roessle, D. Neumann, Decentralized Online Ressource Allocation for
Dynamic Web Service Applications, IEEE Joint Conference on E-Commerce Technology
(CEC'07) and Enterprise Computing, E-Commerce and E-Services (EEE '07), 23-26
July 2007, Tokyo, Japan.
• D. Neumann, J. Stoesser, C. Weinhardt, Bridging the Grid Adoption Gap – Developing
a Roadmap for Trading Grids, 20th Bled eConference, Merging and Emerging Technologies,
Processes, and Institutions, June 4-6, 2007, Bled, Slovenia – Best Theme Paper
Award.
• A. Anandasivam, D. Neumann, C. Weinhardt, Rightsizing of Incentives for collaborative
e-Science Grid Applications with TES, 20th Bled eConference, Merging and
Emerging Technologies, Processes, and Institutions, June 4-6, 2007, Bled, Slovenia.
• P. Bodenbenner, J. Stoesser, D. Neumann, A Pay-as-Bid Mechanism for Pricing Utility
Computing, 20th Bled eConference, Merging and Emerging Technologies, Processes,
and Institutions, June 4-6, 2007, Bled, Slovenia.
• D. Neumann, C. Block, C. Weinhardt, Y. Karabulut, Knowledge-Driven Selection of
Market Mechanisms in e-Procurement, European Conference on Information Systems
(ECIS) 2007.
• J. Stoesser, D. Neumann, A. Anandasivam, A Truthful Heuristic for Efficient Scheduling
in Network-centric Grid OS, European Conference on Information Systems (ECIS)
2007.
• D. Neumann, J. Stoesser, A. Anandasivam, N. Borissov, SORMA – Building an Open
Grid Market for Grid Resource Allocation, The 4th International Workshop on Grid
Economics and Business Models (GECON 2007), 28 August 2007, Rennes, France.
• D. Neumann, B. Schnizler, I. Weber, Ch. Weinhardt, Second Best Combinatorial Auctions
– The Pricing-Per-Column Auction, Hawaiian International Conference of System
Sciences (HICSS) 2007.
• A. Anandasivam, D. Neumann, Token Exchange System as incentive mechanism for
the e-Science Grid community, Workshop on Agents for Grid Computing 2007
(AGC2007), Rio de Janeiro, Brasilien.
329

• E. Chen, G. Kersten, D. Neumann, R. Vahidov, C. Weinhardt, B. Yu, A Framework for
E-market System Assessment and Design, Group Decision and Negotiation, 2007,
Montreal, Kanada.
• K. Kolitz, D. Neumann, The Impact of Information System Design on eMarketplaces –
An Experimental Analysis, Group Decision and Negotiation, 2007, Montreal, Kanada.
• D. Neumann, Engineering Grid Markets, N. Jennings, G. Kersten, A. Ockenfels, C.
Weinhardt (Eds.): Dagstuhl Seminar Proceedings 06461 Negotiation and Market Engineering,
2006, Internationales Begegnungs- und Forschungszentrum (IBFI), Schloss
Dagstuhl, Germany.
• G. Kersten, E. Chen, D. Neumann, R. Vahidov, C. Weinhardt, On Comparison of Mechanisms
of Economic and Social Exchanges: The Times Model N. Jennings, G. Kersten,
A. Ockenfels, C. Weinhardt (Eds.): Dagstuhl Seminar Proceedings 06461 Negotiation
and Market Engineering, Internationales Begegnungs- und Forschungszentrum
(IBFI), Schloss Dagstuhl, Germany, 2006.
• D. Neumann, Self-Organizing ICT Resource Management – Policy-based Automated
Bidding, eChallenges e-2006 Conference, Barcelona.
• J. Stoesser, A. Anandasivam, N. Borissov, D. Neumann, Economic Virtualization of
ICT Infrastructures, Cracow Grid Workshop 2006.
• D. Neumann, C. Holtmann, Ch. Weinhardt, Using Blueprints for Engineering Electronic
Market Services, The First International Workshop on Service Intelligence and Service
Science (SISS 2006), October 17, 2006, Hong Kong.
• D. Neumann, S. Lamparter, B. Schnizler, Automated Bidding for Trading Grid Resources,
European Conference on Information Systems ECIS 2006, Göteborg.
• B. Schnizler, D. Neumann, D. Veit, Ch. Weinhardt, Moving Markets to the Grid, 68.
wissenschaftlichen Jahrestagung des Verbandes der Hochschullehrer für Betriebswirtschaft,
2006, Dresden.
• T. Eymann, D. Neumann, M. Reinicke, B. Schnizler, W. Streitberger, D. Veit, On the
Design of a Two-Tiered Grid Market Structure, Multi-Konferenz Wirtschaftsinformatik,
2006, Passau.
9. M. Kunzelmann, D. Neumann and C. Weinhardt, Zwischen Limit und Market Order,
10th Symposium on Finance, Banking, and Insurance, 2005, Karlsruhe.
10. D. Rolli, M. Conrad, D. Neumann, C. Sorge, An Asynchronous and Secure Ascending
Peer-to-Peer Auction, ACM SIGCOMM'05 Workshop on the Economics of Peer-to-Peer
Systems, 2005, Philadelphia, PA.
• D. Neumann, B. Schnizler, Transforming Auction Mechanisms into Protocols, Group
Decision and Negotiation, 2005, Vienna, Austria.
11. B. Schnizler, D. Neumann, D. Veit, C. Weinhardt, A Multiattribute Combinatorial Exchange
for Trading Grid Resources Proceedings of the 12th Research Symposium on
Emerging Electronic Markets (RSEEM), 2005, Amsterdam, Netherlands.
12. D. Neumann, J. Maekioe, C. Weinhardt, CAME – A Toolset for Configuring Electronic
Markets, European Conference on Information Systems (ECIS), 2005, Regensburg.
• B. Schnizler, D. Neumann, C.Weinhardt, Resource Allocation in Computational Grids –
A Market Engineering Approach, Proceedings of the Workshop on eBusiness (WeB
2004) – ICIS Pre-conference, Washington.
330

• D. Rolli, D. Neumann, C. Weinhardt, A Minimal Market Model in Ephemeral Markets,
Proceedings of 1st International Workshop on Theory Building and Formal Methods in
Electronic/Mobile Commerce (TheFormEMC), Toledo, 2004, Spain.
• D. Neumann, C. Holtmann, Embodiment Design in Market Engineering, Proceedings of
the Research Symposium on Emerging Electronic Markets (RSEEM 2004), Dublin, Irland,
S. 85-96.
• D. Neumann, Market Engineering - A Structured Design Process for Electronic Markets
Dissertation, Fakultät für Wirtschaftswissenschaften, 2004, Universität Karlsruhe.
• C. Holtmann, D. Neumann, Market and Firm - Two Sides of a Coin, Proceedings of the
10th Research Symposium on Emerging Electronic Markets (RSEEM 2003), Bremen.
• C. Holtmann, D. Neumann, Konzeption eines Bezugsrahmens für die interdisziplinäre
Analyse und Gestaltung elektronischer Märkte, 65. wissenschaftlichen Jahrestagung des
Verbandes der Hochschullehrer für Betriebswirtschaft, 2003, Zürich.
• D. Neumann, C. Holtmann, T. Honekamp, Market Integration and Metamediation:
Perspectives for Neutral B2B E-Commerce Hubs in: Ch. Weinhardt, C. Holtmann
(Hrsg.) E-Commerce - Netze, Märkte, Technologien, 2002, Physica, Heidelberg, S. 67-
86.
• D. Neumann, C. Holtmann, H. Gimpel, Ch. Weinhardt, Agent-based bidding in Electronic
Markets – A Prototypical Approach J. Monteiro, P. M. C. Swatman und L. V.
Tavares (Hrsg.) Towards the Knowledge Society: e-Commerce, e-Business and e-
Government, 2002, Kluwer Academic Publishers, S.553-567.
• D. Neumann, C. Holtmann, H. Weltzien, C. Lattemann, Ch. Weinhardt, Towards a Generic
E-Market Design J. Monteiro, P. M. C. Swatman und L. V. Tavares (Hrsg.) Towards
the Knowledge Society: e-Commerce, e-Business and e-Government, 2002,
Kluwer Academic Publishers, S. 289-305.
• D. Neumann, Ch. Weinhardt, Domain-independent eNegotiation Design: Prospects,
Methods, and Challenges In Proceedings of the 3rd DEXA Workshop e-Negotiations,
2002, Aix-en-Provence, France, S.680-684.
Sonstige Beiträge
1. D. Neumann, Engineering Grid Markets, in H. Gimpel, N.R. Jennings, G. Kersten, A.
Ockenfels, C. Weinhardt (Hrsg), 2008 Negotiation and Market Engineering, Lecture
Notes in Business Information Processing (LNBIP), Springer.
2. G. Kersten, R. Kowalczyk, H. Lai, D. Neumann, M. Chhetri, Shaman: Software and
Human Agents in Multiattribute Auctions and Negotiations, in H. Gimpel, N.R.
Jennings, G. Kersten, A. Ockenfels, C. Weinhardt (Hrsg), 2008 Negotiation and Market
Engineering, Lecture Notes in Business Information Processing (LNBIP), Springer.
3. G. Kersten, E. Chen, D. Neumann, R. Vahidov, C. Weinhardt, On Comparison of
Mechanisms of Economic and Social Exchanges: The Times Model, in H. Gimpel, N.R.
Jennings, G. Kersten, A. Ockenfels, C. Weinhardt (Hrsg), 2008 Negotiation and Market
Engineering, Lecture Notes in Business Information Processing (LNBIP), Springer.
4. C. Block, D. Neumann, A Decision Support System for choosing Market Mechanisms
in e-Procurement, in H. Gimpel, N.R. Jennings, G. Kersten, A. Ockenfels, C. Weinhardt
331

(Hrsg), 2008 Negotiation and Market Engineering, Lecture Notes in Business Information
Processing (LNBIP), Springer.
5. D. Neumann, D. Veit, C. Weinhardt, Grid Economics: Market Mechanisms for Grid
Markets. Grid Computing: Konzepte, Technologien, Anwendungen. T. Barth and A.
Schüll ‘(‘Hrsg), 2006, Viehweg Verlag: 64-83.
6. Weinhardt, C., D. Neumann, M. Kunzelmann, Design und Evaluierung von Relative
Orders zur Reduktion impliziter Transaktionskosten, in W. Bessler (Hrsg), 2006 Börsen,
Banken und Kapitalmärkte, 205-226.
7. D. Neumann, I. Weber, Market Engineering and the PPC-Auction, CORS/INFORMS
Joint International Meeting, May 19-22 2004, The Banff Center, Banff, Alberta, Kanada.
8. C. Holtmann, D. Neumann, Ch. Weinhardt, e-Commerce im Wertpapierhandel – Der
private Investor im Fokus, Branchengutachten im Auftrag des Deutschen Bundestags,
2001, im Rahmen des TAB-Projekts "e-Commerce", 2. Phase, vorgelegt dem Büro für
Technikfolgenabschätzung (TAB) beim Deutschen Bundestag, Berlin.
332