题目：Design Issues on P2P-based Video-on-Demand Streaming Services
摘要：In the last few years with the increased popularity of Peer-to-Peer (P2P) streaming systems such as PPlive and Youtube, P2P technology has been proved to be an effective and attractive approach for providing on-demand streaming services over the Internet. However, due to the uncertainty of frequent VCR operations, it is challenging to provide true on-demand streaming services with VCR functionality over distributed self-organized P2P overlay networks. In this talk, I will delivery our recent research efforts on exploring the key design issues including streaming overlay organization, media content discovery, and VCR operation support involved in eliminating the negative impact of VCR interactivity in P2P on-demanding streaming systems. Based on the observation that users’ viewing quality is seriously affected by their suppliers’ frequently leaving current positions for VCR-like operations, we present two graph theory motivated streaming schemes with O(logN) overhead (N is the number of sessions/clusters) by employing skip graph and derivative tree to improve the resilience of dynamic streaming overlay network. We take a further step to investigate the user personalization issue, aiming at improving user perceived Quality of Experience (QoE) for VCR operations by reducing response latency and optimizing content sharing. By employing the reinforcement learning techniques, we first transform users’ streaming service procedure into a set of abstract states and present an online prediction based prefetching scheme to proactively fetch segments in accordance with the predicted VCR behavior. Following that we introduce a collaborative filtering to find out active user’s preference and look for peers sharing the same preference, thus optimizing the membership management and minimizing the VCR response latency. Both theoretical analysis and comprehensive simulations based on the user logs collected from real deployed VoD system show that our work outperforms most existing schemes in terms of server stress, response latency, searching efficiency, and accumulated hit ratio.