Managing a Peer-to-Peer Storage System in a Selfish Soci<strong>et</strong>yResearch Staff : Patrick Maillé, Laszlo TokaKeywords : Peer-to-peer, pricing, incentives, game theoryApplications : N<strong>et</strong>work pricing and optimizationPartners & Funding : Institut TELECOM project DisPairSeIntroductionThe possibility of storing data online appearsas a value-adding service in the context of thesoaring “soci<strong>et</strong>y of information”. Indeed,having access to the Intern<strong>et</strong> becomes moreand more easy, with higher and higheravailable transmission rates in accessn<strong>et</strong>works, which renders transfer timesreasonable, even for large files. Moreover,online storage systems are able to cope withdocument versioning, and to protect data notonly against user device failures but alsoagainst disk failures, through the use of datareplicates stored on different disks.For those reasons, many companies nowpropose a service of online data storage.However, while creating such a storage serviceimplies owning huge memory capacities andaffording the associated energy andwarehouse costs, one can imagine using thesmaller but numerous possible storage spacesof the service users themselves, as is done inpeer-to-peer file sharing systems. In such apeer-to-peer storage system, the participantsare at the same time the providers and theusers of the service. To work properly, a peerto-peerstorage n<strong>et</strong>work needs thatparticipants offer a sufficient part of their diskspace to the system, and remain online oftenenough. However, both of those requirementsimply costs (or at least constraints) forparticipants, who may be reluctant to devotesome of their storage capacity to the systeminstead of using it for their own needs.Note that the economic models developed forpeer-to-peer file sharing systems do not applyto peer-to-peer storage services: in file sharingsystems, the data offered by a peer canbenefit to all other peers; in that sense theresource offered to the system is a publicgood. On the contrary, in a peer-to-peerstorage system the memory space offered by apeer is a private good: it can be shared amongdifferent users but each part is then devotedto only one user. Therefore the economicimplications of those systems are necessarilydifferent.Eliciting participants to contributeWe consider two possible incentive schemes,by suggesting that either each peer's use ofthe service should be limited to hercontribution level (symm<strong>et</strong>ric schemes), or thatstorage space be bought from and sold topeers by a system operator that seeks tomaximize profit.The figure presented next illustrates thequantity C i * that is exchanged by user i withdemand function d i and supply function s i inthe case of a symm<strong>et</strong>ric scheme, and thequantity of storing capacity C i o (resp. C i s ) thathe offers (resp. buys) if an operator buys fromand sells storing capacity to participants withunit price p o (resp. p s ).RealizationIn this work, we consider that users behaveselfishly, i.e. are only sensitive to the quality ofservice they experience, regardless of theeffects of their actions on the other users. Theframework of noncooperative Game Theory [1]is therefore particularly well-suited to study theinteractions among peers.We therefore focus on the incentives that canbe introduced to make participants contribut<strong>et</strong>o the system, i.e. the changes that can bebrought to the game to modify its outcomes.Which scheme to prefer?The main question addressed in this work iswh<strong>et</strong>her it is socially b<strong>et</strong>ter to impose asymm<strong>et</strong>ric scheme or to l<strong>et</strong> a profit-maximizing60 Pracom’s Annual Report <strong>2008</strong>
monopoly s<strong>et</strong> prices. The performancemeasure we consider is social welfare, i.e. th<strong>et</strong>otal value that the system has for allparticipants. Under some assumptions on thepeers utility functions, we derive a necessaryand sufficient condition for symm<strong>et</strong>ry-basedsystems to outperform revenue-orientedmanagement.The figure below compares social welfare Wobtained by using each scheme to the maximalreachable value of social welfare W*(assuming user preferences are param<strong>et</strong>erisedby two independent values, a and b, with anexponential distribution of respective mean1/µ a and 1/µ b ).for a type of management to be preferable tothe other: it appears that profit orientedpayment-based schemes may be socially b<strong>et</strong>terthan symm<strong>et</strong>ric ones under some specificcircumstances, namely if the h<strong>et</strong>erogeneityamong user profiles is high.References[1] D. Fudenberg and J. Tirole. Game theory.MIT Press, 1991[2] L. Toka and P. Maillé. Managing a Peer-to-Peer Backup System: Does Imposed FairnessSocially Outperform a Revenue-DrivenMonopoly?, in Proc. of the 4th InternationalWorkshop on Grid Economics and BusinessModels (Gecon’2007, Rennes, France), August2007.[3] P. Maillé and L. Toka. Managing a Peer-to-Peer Storage System in a Selfish Soci<strong>et</strong>y.Submitted.We obtain that user h<strong>et</strong>erogeneity tends tofavor pricing-based schemes that are moreflexible, and above a given user h<strong>et</strong>erogeneitythreshold even a monopoly-managed systemwill be socially b<strong>et</strong>ter than a system imposingsymm<strong>et</strong>ry. Those results have been presentedin [2], and a generalized version of the modelhas been submitted.ConclusionsIn this work we have addressed the problem ofuser incentives in a peer-to-peer storagesystem. Using a game theor<strong>et</strong>ical model todescribe selfish reactions of all systems actors(users and the operator), we have studied andcompared the outcomes of two possiblemanaging schemes, namely symm<strong>et</strong>ry-basedand profit oriented payment-based. Not onlythe size of the offered storage space wastarg<strong>et</strong>ed with incentives, but as the availabilityand reliability is a particularly important issuein storage systems, the model also aimed toreduce churn. By comparing the social welfarelevel at the outcome in the two cases, undersome assumptions on user preferences weexhibited a necessary and sufficient conditionPracom’s Annual Report <strong>2008</strong> 61