On the Performance of Peer Selection Strategies in Stochastic Peer-to-Peer Networks

Abstract

Peer-to-peer (P2P) file-sharing applications are becoming increasingly popular and account for a large portion of the Internet's bandwidth usage. Measurement studies show that a typical download session lasts from hours up to several days depending on the level of network congestion or the service capacity fluctuation. In this thesis, we first consider two major factors that have significant impact on the average download time, namely, the spatial heterogeneity of service capacities in different source peers and the temporal fluctuation in service capacity of a given single source peer. We point out that the common approach of analyzing the average download time, or more generally the performance of peer to peer networks based on average service capacity is fundamentally flawed. We rigorously prove that both spatial heterogeneity and temporal correlations in service capacity increase the average download time in P2P networks. We then analyze the impact of the interaction and resource competition between peers on the file download performance under stochastic, heterogeneous, unstructured P2P settings. We introduce the notion of system utilization tailored to a P2P network so as to capture the characteristics of the average download time in a P2P network with multiple competing downloading peers. We then derive an achievable lower bound on the average download time and propose a distributed algorithm with which peers can achieve this minimum average download time, thereby bypassing the curse of spatial heterogeneity and temporal stochastic fluctuation. Our algorithm relies on constantly changing connected source peers and selecting source peers probabilistically. The performance of different peer selection algorithms is compared under NS-2 simulations. Our results also provide theoretical explanation to the inconsistency of performance improvement by using parallel connections (parallel connection sometimes does not outperform single connection) observed in some measurement studies.