Live Streaming Performance of Peer-to-Peer Systems

Sammanfattning: In peer-to-peer (P2P) live streaming systems peers organize themselves in an overlay and contributewith their resources to help diffuse live content to all peers in a timely manner. The performanceof such systems is usually characterized by the delay-loss curve, which quantifies theplayback delay required for achieving a certain streaming quality, expressed as the chunk missingratio at the peers. The streaming quality is determined by the overlay construction algorithm, theforwarding algorithm, the loss process in the underlying network, the number of peers in the overlayand their bandwidth distribution, the willingness of the peers to contribute with their resourcesand the viewing behavior of the peers (churn). The overlay construction and forwarding algorithmsare inherent characteristics of a P2P protocol, while the remaining factors are artifacts of thedeployment of the P2P system over a best-effort network such as the Internet, as well as the factthat peers act as independent agents. The current thesis addresses the problem of evaluating andimproving the performance of P2P streaming protocols based on models of the network and of thepeers' behavior. The first part of the thesis is devoted to the performance evaluation of P2P overlay constructionand forwarding algorithms and offers three contributions. First, we study the efficiency of datadistribution in multiple tree-based overlays employing forward error correction. We deriveanalytical expressions for the average packet possession probability as well as its asymptoticbounds and verify our results through simulations. Second, we evaluate the performance of astreaming system in the presence of free-riders. We define two admission control policies and studythe streaming feasibility using an analytical model and via simulations. Third, we present ananalytic framework for the evaluation of forwarding algorithms in mesh-based systems. We validate itvia simulations and use it to evaluate and to compare four push-based forwarding algorithms in termsof their delay-loss curves. The second part of the thesis investigates potential improvements to the operation of P2P streamingsystems and offers three contributions in that area. First, we study the impact of selfish peerbehavior on streaming quality in overlays where a fraction of peers has limited contribution due tophysical constraints. We show that selfish peer behavior results in suboptimal streaming quality andwe propose an incentive mechanism that increases the streaming quality by using the server uploadcapacity to reward high contributing peers. Second, we study the problem of building network aware P2P streaming overlays, taking into accountrecent measurement results that indicate that the AS-level topology of the Internet is flattening.Through extensive simulations on regular and measured topologies we show that it is possible tocreate better than random overlays relying on information about the underlying topology. Finally, westudy the problem of playout adaptation in P2P streaming systems under churn. We propose andevaluate two algorithms that tune the playback delay of the peers in such a way that the streamingquality of the peers is maintained within predetermined limits. We use simulations to show thecorrectness of the proposed algorithms and the benefits from their use.