Dynamic Routing in Self-Healing MPLS Networks - (Distributed ...

Dynamic Routing in Self-Healing MPLS NetworksKrzysztof WalkowiakChair of Systems and Computer Networks, Faculty of Electronics, Wroclaw University ofTechnology, Wybrzeze Wyspianskiego 27, 50-370 Wroclaw, Polandwalkowiak@zssk.pwr.wroc.plAbstract. Modern computer networks require self-healing capabilities. Selfhealingis the ability of a computer network to reconfigure itself around failuressuch that calls in progress are not dropped and suffer no or almost no degradationin quality of service. Provision of self-healing in MPLS (MultiProtocolLabel Switching) networks in a cost effective manner is an important problem.In this work we consider issues of dynamic routing in self-healing MPLS networks.We develop new functions and link metrics for robust establishing ofLSPs (label switched paths). We compare performance of new metrics to otherlink metrics used in routing algorithms. As the performance indicator we applythe lost flow function. Simulations show that proposed metrics provide from8% to 18% better results than the standard hop-number metric.1 IntroductionIP networks propose scalability and flexibility for rapid deployment of value added IPservices. Nevertheless, the increasing demand and explosive growth of the Internetcause that carriers require a network infrastructure that is dependable, predictable andoffers reliable network performance. Efficient routing and traffic engineering are thetwo foundations needed to accomplish survivable networking. Moreover, significantnetworking features such as path optimization, resilience and failure recovery are alsoimportant for backbone networks to deliver reliable services to customers using thenetwork for mission-critical business. The MultiProtocol Label Switching (MPLS)approach proposed by the Internet Engineering Task Force (IETF) is a networkingtechnology that enables delivering traffic engineering capability and QoS performancefor carrier networks. An MPLS node that is capable of forwarding IP packetsand supports MPLS is called a label switching router (LSR). The path through one ormore LSRs at one level of the hierarchy followed by a packet in a particular FEC(Forwarding Equivalence Class) is known as a label switched path (LSP).According to [12] self-healing is the ability of a computer network to reconfigureitself around failures such that calls in progress are not dropped and suffer no or almostno degradation in quality of service. Provision of self-healing abilities in MPLSnetworks in a cost effective manner is a significant issue.The main idea of providing self-healing capabilities in computer networks consistsin providing spare capacity available on network links. If a network link fails, thepackets will be distributed among remaining links using spare capacity of these links.

Most of self-healing methods use sharing of spare capacity for various failure events.This approach yields a cost effective self-healing restoration.Providing self-healing includes many aspects: spare capacity allocation, reroutingstrategies, how traffic is distributed, failure detection, propagation of informationabout failure, network control [1], [8]. In this work we focus on problems of sparecapacity allocation in existing MPLS networks. We consider an existing MPLS network,i.e. the network topology and link capacity are given. Our main goal is to provideroutes for all demands in order to use the network capacity efficiently in terms ofnetwork reliability. It must be noted that having the network resource and a trafficdemand, the routes assignment determines spare capacity allocation over the network.Thus, assignment of routes has an effect on the amount of flow restored when a failureoccurs [9]. Therefore, the main objective of this paper is to develop new metrics(weights) for routes calculation. Using these metrics should provide effective allocationof spare capacity that enables restoration of flow after a network failure. To makean evaluation of proposed metrics we compare performance of these metrics withperformance of other metrics used in non-reliable dynamic routing, e.g. the hopnumbermetric. Results of extensive simulation tests are presented and discussed.2 Self-healing MPLS NetworksSelf-healing MPLS networks realize fast restoration from a network failure byswitching affected LSPs over alternate routes. There are two models for path recoveryin MPLS: rerouting and protection switching. Recovery by rerouting is defined asestablishing new paths on demand for restoring traffic after the occurrence of a failure.The recovery paths may be found using network routing policies, precomputedconfigurations and network topology information. After detecting a failure, paths areestablished using signaling. Since more operations must be done, rerouting is muchslower than protection switching mechanism. However, while network resources havenot to be reserved until the failure, rerouting is easier. Protection switching recoverymechanisms use preestablished recovery paths, based upon network routing policies,the restoration requirements of the traffic on the working path, and administrativeconsiderations. When a failure occurs, the protected traffic is switched over to therecovery path(s) and restored. It must be noted that protection switching and reroutingmay be used together. For instance, in order to recover online a failed path andprovide connectivity protection switching may be used. Rerouting may be applied todetermine a new optimal network configuration and rearrange paths in offline manner[11].There are three configurations of working and recovery paths for rerouting andprotection switching in MPLS networks [2], [3], [11]: Local repair. Known also as local recovery, protects the working path against alink or neighbor node failure. Local repair can be of two types: link recovery/restoration,node recovery/restoration. The former one assumes, that the recoverypath routes around a failed link. The alternate path is found between the twoLSRs on the ends of a failed link. In latter case, the recovery path may be configuredto route around a failed neighbor node. The main advantage of local repair is

fast propagation of a failure information and simplicity. The node upstream of thefailure is responsible for initiation of the recovery process and switching the trafficto a recovery path. Global repair. The main idea of global repair, called also edge-to-edge rerouting, isto protect the working path against any link or node fault on a path, except thefailures occurring at the ingress node of the protected path. It means that the recoverypath must be link and node disjoint from the working path. It has the advantageof protecting against all link and node failures on the working path. Themain drawback is slower reaction to a failure than for local repair. Reverse backup. In the reverse backup approach the traffic of failed link is reversedfrom the point of a failure, i.e. the ingress node of the failed link, back tothe source (ingress) node of the protected working path. That node reroutes incomingtraffic to alternative recovery paths. Major benefit is relatively small time ofrestoration, since the failure notification is simplified. The main disadvantage isreduced resource utilization. Two recovery paths are needed: one from the point offailure to the ingress node of the path and the second one from the ingress node tothe egress node of the protected path.LSR4LSR4LSR4LSR7LSPworkingpathLSR7LSPworkingpathLSR7LSPworkingpathLSPrecoverypathLSR3LSPrecoverypathLSR3LSPrecoverypathLSR3LSR6LSR6LSR6LSR2LSR2LSR2LSR5LSR5LSR5LSR1LSR1reversebackupLSPLSR1(a) (b) (c)Fig. 1. MPLS rerouting strategies: (a) Global repair, (b) Local repair, (c) Reverse backupFig. 1 illustrates three configurations of working and recovery paths for reroutingand protection switching. The working LSP 1-2-3-4 connecting LSR1 and LSR 4 isbroken when the link 2-3 fails. For global repair the node 1 is informed about thefailure and is responsible for rerouting the traffic to a recovery path 1-5-6-7-4(Fig.1a). In local repair scheme, the LSR2 reroutes to a recovery path 2-7-3 to omitthe failed link (Fig. 1b). It results in a recovery path 1-2-7-3-4 for demand pair LSR1-LSR4. For reverse backup approach, LSR2 reroutes to a reverse backup LSP 2-1.Then LSR1 reroutes the traffic to a recovery path 1-5-6-7-4 (Fig. 1c).

For local repair, global repair and reverse backup the recovery path may be preestablishedor dynamically sought after notification of a failure. The former approachprovides effectiveness and small recovery time. Resources of pre-established recoverypaths may be reserved in order to increase reliability. The latter approach is veryflexible. However, since some traffic may not be restored due to limited networkresources, it doesn’t guarantee high reliability [3].3 Functions for LSPs ReroutingMPLS supports two methods of route selection: hop-by-hop routing, and explicitrouting [10]. In this work we apply the explicit routing. Explicit routing assumes thata single LSR, usually the LSP ingress node, specifies several or all LSRs in the LSP.Therefore, it’s a kind of source-destination routing. According to [7] sourcedestinationrouting enables providing traffic engineering and policy routing betterthen traditional hop-by-hop destination oriented routing. Links that are congested canbe avoided with routing decisions made at the LSP ingress node. Moreover, the use ofdynamic link metric instead of static link metric facilitates selection of good pathsthat meets the required QoS parameters.We focus on two rerouting methods: local repair and reverse backup. Recall that inboth schemes the node upstream of the failure is responsible for initiation of the recoveryprocess and switching the traffic to a recovery path. Therefore, that node is thepotential bottleneck of the restoration process. More precisely, the flow of the failedlink must be rerouted using other links leaving the considered node, excluding thefailed link. The residual capacity of these links is an upper bound of the flow that canbe restored. We define residual (spare) capacity of a link as a difference between linkcapacity and flow of that link. The link flow is a sum of all LSPs’ flows using thatlink. Since the residual capacity is used for rerouting of failed LSPs, it is very importantto locate resources of residual capacity effectively.The central idea of our approach is to develop new objective functions for workingroutes assignment. Such functions should indicate preparation of the network to thererouting. Next, these functions can be used to develop link metrics for route calculation.We consider an existing MPLS, i.e. we do not consider capacity planning andtopological design. We make an assumption that for restoration of failed LSPs thedynamic rerouting after the occurrence of a failure is applied. We model MPLS networkas a directed graph G=(N,A,C) where N is a set of n nodes (vertices) representingnetwork switches, A is a set of m arcs (directed edges) representing network linksand C is a vector of link capacity.We denote by o : A → V and d : A → V functions defining the origin and destinationnode of each arc. For each A in ( a)= k ∈ A | d(k)= d(a),k ≠ aa ∈ we call { }the set of incoming arcs of d(a) except a, and out ( a)= { k ∈ A | o(k)= o(a),k ≠ a}theset of outgoing arcs of o(a) except a.To mathematically represent the problem, we introduce the following notationsf Represents the total flow on arc a.a

c The capacity of arc a.aoutg v = ∑ f ii:o(i)=vAggregate flow of outgoing arcs of v.ing v = ∑ f i Aggregate flow of incoming arcs of v.i:d(i)=voute v = ∑ c i Aggregate capacity of outgoing arcs of v.i:o(i)=vine v = ∑ c i Aggregate capacity of incoming arcs of v.i:d(i)=vWe model the MPLS flow as non-bifurcated multicommodity (m.c) flow denotedby f = [ f 1,f 2 ,..., f m ]. f is a vector of flows in all arcs. We call a m.c. flow f validif for every arc a ∈ A the capacity constraint holds. For more information on m.c.flows refer to [5].For the sake of simplicity we introduce the following function⎧0for x ≤ 0(1)ε ( x)= ⎨⎩xfor x > 0To analyze the local repair and reverse backup strategies we consider an arck ∈ A . We assume failure of k. Recall that considered rerouting methods assume thatflow on the arc k must be rerouted by the source node of the arc k. Therefore, sparecapacity of outgoing arcs of o(k) except k is a potential bottleneck of the rerouting.Notice that iff ≤ ∑ ( c − f )(2)kii∈ out( k)then flow of the failed arc k can be restored using spare capacity of other arcs leavingthe origin node of k. Recalling definition of g o(k)and e o(k)we can reformulateout out(2) in the following waygouto k≤ eout ( ) o(k)i− ck(3)Otherwise, iff > ∑ ( c − f )(4)kii∈ out( k)ithen some flow of the failed link k cannot be restored, since spare capacity of otherarcs leaving the origin node of k is too small. It means that these arcs block the 100%out outrestoration and some flow of arc k is lost. Recalling definition of g o(k)and e o(k)wecan reformulate (4) in the following waygouto k> eout ( ) o(k)− ck(5)

We define the flow of k lost in the node o(k) as flow that cannot be restored usingother arcs leaving o(k) due to limited resources of spare capacity as followsLoutkout out( g − ( e − c ))( f ) =ε (6)o(k)o(k)outout outThe Lkfunction depends on the flow g o(k)leaving the node o(k). Lkis notdependent directly on the flow f k of the arc k. It means that changing the flow on otheroutarcs leaving the node o(k) modifies value of the function Lk. Therefore, we formulatea function LN v of lost flow leaving the node v as a sum of Loutoutkfunctions overall arcs leaving that nodeLNoutvNote that functionskout out( g −(e −c))out( f ) = ∑ L ( f ) = ∑ε (7)ii:o(i)= vi:o(i)= voutLkandoutLN v provide a good metric for reverse backup rerouting.In the same way as functionsandinLN vLNinvinkLinkvvoutLkandoutLN v( g − ( e − c ))ind ( k )ind ( k )i, we can define functions( f ) =ε (8)kin in( g −(e −c))in( f ) = ∑ L ( f ) = ∑ε (9)ivi:d ( i)= v i:d(i)= vThe function L denotes lost flow that cannot be restored using arcs entering theinnode d(k) due to limited spare capacity of these arcs. Correspondingly, LN v is a lostflow entering the node v.outinoutL k ( f ) , L k ( f ) , LN v ( f ) and LN v ( f ) are continuous, non-decreasing,piece-wise linear and convex functions for any valid flow f . The formal proof ofthese properties can be found in [13]. Moreover, function L k ( f ) is a lower boundof the lost flow for reverse backup rerouting of arc k.It must be noted that authors of [8] have introduced similar approach for local reroutingof ATM network. They formulated a problem of working routes assignmentwith the objective function of lost flow using the k-shortest path based rerouting.For local repair all LSPs using k must be rerouted around k after failure of this arc.In order to estimate the amount of the restored flow, the maximum flow algorithm canbe applied. The maximum flow criterion denotes the theoretical maximal reroutingcapacity. In this method, the failed arc k is removed from the network. Next, maximumflow between the origin and destination node of k is calculated. Another performanceindicator for local repair is k-shortest paths (KSP) algorithm, which finds k-successively shortest disjoint paths in a graph. Authors of [4] compared these twoinvioutinL k

strategies using simulation methods. KSP restoration offers performance 99.9% ofthat from Max Flow. The advantage of KSP is time complexity of O(nlogn) comparedto maximum flow O(n 3 ) using the centralized restoration by a single processor computation.Let MF(k) denote flow of the failed arc k restored by the maximum flow method.We formulate a function of lost flow due to a failure of arc k using maximum flowrestoration method as followsLMFk( f − MF())( f ) =ε k(10)kFunction L MF k ( f ) provides a theoretical minimal value of flow lost after the localrepair of k. It can be interpreted as a performance indicator used here to measure thenetwork performance in terms of self-healing capabilities.4 Definitions of Link MetricsMain goal of this work is to develop and evaluate new link metrics for computation ofLSPs. Routes should be set up in such a way that the network will be prepared torerouting of failed LSPs. Since the residual capacity will be located efficiently, thenetwork will be better arranged to the rerouting of all failed LSPs. As mentionedabove, we apply the explicit routing of MPLS that enables the ingress LSR of LSP tospecify all LSRs in the LSP. We use a partial information model proposed in [6]. Inthis model the information available to the routing algorithm is the total bandwidthusage for each link. It means that each LSR in the network has the knowledge on thenetwork topology, capacity and flow in each link. Author of [7] proposed to apply anextended version of link state routing protocol to obtain these information. We concentrateon additive metrics, i.e. the path’s length is equal to the sum of the correspondingmetrics of the links along that path. Another possibility, not considered inthis work, is to use nonadditive metrics. In this case the value of a path is the minimum(or maximum) link metric along that path.We make an assumption that all LSPs are not known a priori. Therefore, the dynamiconline routing of LSPs is required. Each new LSP’s route is determined usingthe shortest path algorithm applying selected link metrics using the information onlink capacity and link aggregate flow. Applying functions defined in the previoussection, we develop three new metrics that can be used for computation of LSPs’shortest routes.In order to make easier the consideration we introduce a new function⎧0ϖ ( x)= ⎨⎩1Next we define the following functionforforx ≤ 0x > 0(11)

outvout out( g −(e −c))outτ ( f ) = ∑ϖ(12)vi:o(i)=vvNote that τ ( f ) (12) is a derivative of LN v ( f ) (7) for a valid flow f exceptvoutpoints g = e − c ) for all i : o(i)= v . In these points the function τ ( f ) is( voutiequal to the left-sided derivative of the function LN v ( f ) . Similar we can define thevinfunction τ ( f )voutioutin in( g −(e −c))outvinτ ( f ) = ∑ϖ(13)vi:d ( i)=vThe new metrics for the shortest route path calculation are defined as followsLNlivout1+τo(i)vi= (14)Lliout= 1+0.5( τ + τ )(15)ino( i)d ( i )Llimaxout= 1+max( τ , τ )(16)o(i)ind ( i )Metrics (14-16) are formulated in order to locate resources of spare capacity in thebest way in terms of the network reliability. To make a comprehensive evaluation ofthese three new metrics, we compare their performance with other four metricsl i1 = 1(17)fil =fi(18)f /( cil1/( cil− f ) fi(19)=( c − f )ii− f ) 1(20)=( c − f )iMetrics (17-20) directly don’t take into account any reliability criteria. The metric(17), called the hop-number metric, is widely used in many routing protocols.i

5 ResultsWe made extensive tests over 10 various networks with the number of nodes varyingfrom 10 to 18 (see Table 1).Table 1. Parameters of tested networksName Number ofnodesNumber oflinksAverage nodedegree (avnd)Number ofpathsNumber oftests1034 10 34 3.40 90 41038 10 38 3.80 90 41042 10 42 4.20 90 51046 10 46 4.60 90 51450 14 50 3.57 182 121456 14 56 4.00 182 101462 14 62 4.43 182 101866 18 66 3.66 306 91874 18 74 4.11 306 91882 18 82 4.56 306 9The methodology of tests is as follows. All LSPs, one by one, are processed. EachLSP is represented by the origin node, destination node and bandwidth requirement.A shortest path algorithm according to the considered metric finds the working route.If it is impossible to establish a new LSP due to capacity constraints, the LSP is rejected.When all LSPs have been processed, the following functions representing theperformance indicators are calculatedMFMFL ( f ) = ∑ LFi( f )i∈A(21)outoutLN ( f ) = ∑ LNi( f )i∈A(22)MFRecall that the function L defines the theoretical minimal value of flow lost ina network using the local repair due to a single failure of any arc. The functionoutLN is an estimation the flow lost in the network due to a single failure of any linkfor reverse backup rerouting strategy. Both functions can be considered as the performanceindicators used here to measure network performance. These two functionsindicate the preparation of the network to the restoration process. If value of functionsMF outL and LN is low, more flow is restored after a failure. As a third performanceindicator we used the number of rejected calls (NRC), i.e. the number of LSPs, whichare not set up due to limited resources of the capacity. The same set of LSPs is processedfor all metrics (14-20) defined in previous section.We can see in Table 2 that results for metrics (18-20) are much worse than forMFother four metrics. Values of function L for metrics (14-16) are lower than for theouthop-number metric (17) respectively by 8%, 12% and 10%. For LN function

metrics (14-16) have even better performance comparing to the hop-number metricand the difference is respectively 14%, 18%, and 14%. However, the hop-numbermetric provides the lowest number of rejected LSPs given by the NRC function.Table 2. Results of all testsMetricsFunction 1 f f/(c-f) 1/(c-f) LN L LmaxMFL 221394 314687 337261 377719 204078 195365 200236outLN 110104 198341 231895 252241 94159 90702 94689NRC 322 499 647 725 360 355 361To present the results on Fig. 2 and Fig. 3, the amount of lost flow calculated accordingto the function L (21) is normalized with the respect to the link capacity.MFNLF is defined as a unit of normalized lost flow where 100 NLF is equal to the capacityof all links in the network. Fig. 2 presents values of the normalized lost flowfunction for all tested networks. Performance of metrics (14-17) is shown. We cansee that for various networks (in terms of the avnd parameter denoting the averagenode degree) new metrics provide better performance than the hop-number metric.Normalized Lost Flow NLF504540353025201510503,30 3,50 3,70 3,90 4,10 4,30 4,50 4,70avndFig. 2. Graph showing performance of metrics LN, L, Lmax, hop-number as a function of theaverage node degreeOn the Fig. 3 we can see performance of metrics (14-17) for the network 1882 as afunction of the network average link utilization (avlu). The avlu parameter is a quotientof the overall network flow and capacity of all links. The performance is denotedby normalized lost flow. When the load is not heavy, metrics (14-16) provide muchbetter performance than the metric (17). For more saturated networks results of allmetrics are much more similar.LNLLmax1

6 Concluding RemarksThe two main contributions in this paper are the development of new metrics fordynamic LSPs calculation and a comparative analysis of these metrics and other metricswith respect to MPLS constraints. According to obtained results the networkusing new metrics is better prepared to the restoration process. Therefore, applyingnew metrics leads to improved performance of the network in terms of self-healingcapabilities. The experimental data is reasonably well explained by the hypothesisthat new metrics take into consideration issues of network reliability and spare capacityallocation. Other tested metrics don’t consider directly spare capacity allocation.Another conclusion is that performance of metrics (14-16) is better for less congestednetworks. As average link utilization increases, the hop-number metric gives resultscomparable to results of metrics (14-16). Furthermore, the comparison of metricsshows that the performance doesn’t depend on the average node degree parameter.An additional advantage of this paper is that proposed approach can be applied toboth centralized and distributed control scenarios of self-healing networks.Normalized Lost Flow NLF403530252015105LNLLmax100,42 0,47 0,52 0,57 0,62 0,67 0,72avluFig. 3. Graph showing performance of metrics LN, L, Lmax, hop-number as a function of theaverage link utilization for the network 1882For design of computer networks we can use offline or online algorithms. Abovewe discussed the possibility of applying new objective functions and new metrics fordynamic, online routing. For offline algorithms functions developed in section 3 canbe applied as objective functions in the optimization problem of static routes assignment.For more information refer to [13], [14].In future work we plan to compare our approach to the path caching frameworkproposed in [7]. Furthermore, we want to extend our research on nonadditivie metrics(e.g. available bandwidth, residual capacity) and compare performance of developedabove additive metrics with nonadditive metrics.

References1. Anderson, J., Doshi, B., Dravida, S., Harshavardhana, P.: Fast Restoration of ATM Networks.IEEE JSAC, 1 (1994) 128-1382. Calle, E., Jove, T., Vila, P., Marzo J.: A dynamic multilevel MPLS protection domain. ThirdInternational Workshop on Design of Reliable Communication Networks DRCN 2001, Budapest(2001) 289-2943. Chen, T., Oh, T.: Reliable Services in MPLS. IEEE Comm. Mag., 12 (1999) 58-624. Dunn, A., Grover, W., MacGregor, M.: Comparison of k-Shortest Paths and Maximum FlowRouting for Network Facility Restoration. IEEE JSAC, 1 (1994) 88-995. Kasprzak, A.: Topological Design of Wide Area Networks. Wroclaw University of TechnologyPress, (2001)6. Kodialam, M., Lakshman, T.: Restorable Dynamic Quality of Service Routing. IEEE Comm.Mag., 6 (2002) 72-807. Medhi, D.: QoS Routing Computation with Path Caching: A Framework and NetworksPerformance. IEEE Comm. Mag., 12 (2002) 106-1138. Murakami, K., Kim, H.: Virtual Path Routing for Survivable ATM Networks. IEEE/ACMTrans. Networking, 2 (1996) 22-399. Murakami, K., Kim, H.: Optimal Capacity and Flow Assignment for Self-Healing ATMNetworks Based on Line and End-to-End Restoration, IEEE/ACM Trans. Networking, 2(1998) 207-22110. Rosen, E., Viswanathan, A., Callon, R.: Multiprotocol Label Switching Architecture. RFC3031, (2001)11. Sharma, V., Hellstrand, F. (ed.): Framework for MPLS-based Recovery. RFC 3469 (2003)12. Van Landegem, T., Vankwikelberge, P., Vanderstraeten, H.: A Self-Healing ATM NetworkBased on Multilink Principles, IEEE JSAC, 1(1994) 139-14813. Walkowiak, K.: Algorithms for assignment of virtual paths in survivable ATM networks.Ph.D. Thesis, Sci. Papers Division of Systems and Computer Networks of Wroclaw Univ.Technolog., Report Preprinty 2/00 (2000)14. Walkowiak, K.: Assignment of virtual paths in survivable ATM networks for localdestinationrerouting. Third International Workshop on Design of Reliable CommunicationNetworks DRCN 2001, Budapest (2001) 273-280