Performance and Availability in Peer-to-Peer Content Distribution Systems: A Case for a Multilateral Incentive Approach

• Main Author: Kaune, Sebastian
• Format: Others
• Language: English; en
• Published: 2011
• Online Access: https://tuprints.ulb.tu-darmstadt.de/2492/1/Phd-Thesis-Kaune___2011.pdf; Kaune, Sebastian <http://tuprints.ulb.tu-darmstadt.de/view/person/Kaune=3ASebastian=3A=3A.html> (2011): Performance and Availability in Peer-to-Peer Content Distribution Systems: A Case for a Multilateral Incentive Approach.Darmstadt, Technische Universität, [Ph.D. Thesis]

The peer-to-peer paradigm offers the potential to address many of the challenges related to large-scale content distribution over the Internet (e.g., scalability, costs, etc.). However, currently, many approaches fail to offer the quality of service supported by over-provisioned client-server distribution systems such as RapidShare and YouTube. The reason for this is that peer-to-peer systems must operate under the constraints placed on them by contributing users. Consequently, selfish, strategic, or malicious users often lower the overall system utility. To address this, incentive mechanisms are employed to motivate such users to cooperate with the system (e.g., those that contribute more, receive more in return). Many users, however, observe that popular systems such as BitTorrent (employing tit-for-tat as incentive mechanism), are often ineffective at fulfilling a set of key content distribution requirements, namely performance and availability. The overarching goal of this thesis is therefore to extend existing incentive mechanisms to better fulfil these core requirements. At first, this thesis validates and quantifies these casual observations regarding peer-to-peer performance and content availability. To achieve this we carry out two major measurement studies in the BitTorrent system, the current de-facto standard for peer-to-peer content distribution. To this end we identify widespread performance and availability problems that, through detailed analysis, are attributed to ineffective incentive design. In particular, we find that the current popular approach of ’tit-for-tat’ fails to incite sufficient cooperation amongst users to ensure high performance and content availability. Further analysis shows that this is caused by a lack of incentives for seeding (i.e., the process of remaining as a data source after one has downloaded the entire file). Based on these findings, we subsequently study a set of intuitive solutions to overcome what we call the seeder promotion problem. The purpose of this analysis is to narrow the solution space and to shape a more sophisticated incentive design. In order to achieve this, three abstract cross-torrent incentive approaches are detailed, as well as single-torrent incentive mechanisms. Each of these approaches are then quantitatively analysed through extensive trace-based simulations. These analysis further confirm our finding that bilateral incentive strategies (e.g., tit-for-tat) are insufficient at providing robust incentives for seeders. This is because most users (i) do not meet each other repeatedly and (ii) do not simultaneously require each other’s content. Instead, it is shown that the only way to overcome performance and availability issues is to use multilateral incentive strategies (i.e., to allow users to contribute to one user yet receive reciprocation from another). Finally, we design and evaluate a novel multilateral incentive mechanism, named FairSwarm.KOM. Unlike digital currency systems (in which contribution information is globally visible) or tit for-tat (where no propagation of credit points occurs), FairSwarm.KOM uses one-hop information of the overlay network to evaluate the cooperativeness of the peers. Through the use of extensive trace-based simulations, it is shown that FairSwarm.KOM improves the download performance of the popular BitTorrent system by more than 86%, while guaranteeing high levels of file availability (>99%). Most importantly, these two properties are achieved without harming the fairness of individual users and with an extremely low overhead.