Distilling Superior Peers in Large-Scale P2P Streaming Systems

In large-scale peer-to-peer (P2P) live streaming systemswith a limited supply of server bandwidth, increasing theamount of upload bandwidth supplied by peers becomes criticallyimportant to the well being of streaming sessions in livechannels. Intuitively, two types of peers are preferred to be keptup in a live session: peers that contribute a higher percentage oftheir upload capacities, and peers that are stable for a long periodof time. The fundamental challenge is to identify, and satisfy theneeds of, these types of superior peers in a live session, andto achieve this goal with minimum disruption to the traditionalpull-based protocols that real-world live streaming protocols use.In this paper, we conduct a comprehensive and in-depthstatistical analysis based on more than 130 GB worth of runtimetraces from hundreds of streaming channels in a largescalereal-world live streaming system, UUSee (among the topthree commercial systems in popularity in mainland China). Ourobjective is to discover critical factors that may influence thelongevity and bandwidth contribution ratio of peers, using survivalanalysis techniques such as the Cox proportional hazards modeland the Mantel-Haenszel test. Once these influential factors arefound, they can be used to form a superiority index to distillsuperior peers from the general peer population. The index can beused in any way to favor superior peers, and we simulate the useof a simple ranking mechanism in a natural selection algorithm toshow the effectiveness of the index, based on a replay of real-worldtraces from UUSee.