Modeling and Analysis of Peer-to-Peer (P2P) Live Video Streaming

Peer-to-Peer (P2P) live streaming has become a popular means of distributing real-time online video contents. The distributed nature of the system provides great flexibility, scalability and robustness. The central theme of this thesis revolves around the following general questions that concern the modeling, performance, and stability of P2P live-streaming systems: 1. What models are effective in describing and analyzing the dynamics of a P2P streaming network and what are the trade-offs in different modeling methods? We examine three stochastic model candidates (Chapter 2) that depict the network at various levels of granularity [1, 2]. While we do not present any explicit results related to the models, they serve not only as conceptual frameworks for understanding the critical characteristics of the network, but also important analytical tools on which later results are based. 2. What are the decision variables involved in the process, and how can we utilize them to optimize the peers' viewing experiences? In particular, we identify the average downloading rate achieved by the network as one of the most important performance measures, as it assesses the video playback continuity experienced by an average peer. We present heuristic criteria for optimizing permutation-based downloading policies, which outperforms the best Mixed policy in [1]. 3. Is the P2P video streaming network stable under certain downloading policies and incentive strategies? Even if the model suggests the existence of a healthy steady state with a desirable throughput, is there still a chance for the network to become stuck in a state with poor performance? We show uniqueness of marginal distribution for several typical permutation-based downloading policies, indicating the steady-state marginal chunk distributions for these policies are stable. We also show that there exists a bistability of fixed point of marginal distribution when the tit-for-tat incentive requirement is enforced, leading to drastically different continuity performances depending on the initial state of the network. 4. Lastly, we ask what downloading policies a selfish peer would choose in order to maximize its own playback continuity, given that it has complete information of the empirical distribution of other peers. This sheds light on the robustness of downloading policies against malicious peers. We present a necessary condition for any optimal downloading policy.unpublishednot peer reviewedU of I OnlyUndergraduate senior thesis not recommended for open acces