Research On Data Dissemination For Peer-to-Peer Video-on-Demand Systems

Video-on-Demand (VoD) has emerged as a popular application over the Internet, while being evolving as the main technology to delivery rich content, such as news, sports, and movies. Driven by this requirement, it is becoming more and more important to provide highly scalable VoD streaming services to users at the lowest cost. Taking advantage of the peer-to-peer (P2P) technology for VoD is a potential approach of achieving this goal. This dissertation performs a detailed study on the design, deployment and evaluation of P2P VoD, with a special focus on its data dissemination in terms of peer organization, scheduling algorithms, and data distribution strategies.First of all, a novel ring-assisted overlay network, namely RINDY, is presented to well organize joined peers for efficient content sharing. In RINDY, each peer self-manages a set of concentric rings with power law radii and places all neighbors on these rings according to their playhead distances. More specifically, near neighbors with overlapped buffer windows with current node are placed on the innermost ring as backup data suppliers; while some randomly sampled remote neighbors are placed on the outer rings as routers to nodes with playheads specified by VCR operations. Under this scheme, a peer only needs one hop to join the overlay and at most log (T/w) hops to identify a new group of peers close to the target playback positions specified by random seeks, where T is the total time length of the video and w is the buffer window size of peers.Existing systems and proposals largely utilize scheduling mechanisms that are periodical in nature, referred as time-driven scheduling. This dissertation argues that time-driven scheduling is not efficient for P2P VoD, due to its inadequate resource utilization and inability in quickly adapting to the potential dynamic in such systems. To this end, we propose a priority-aware event-driven scheduling algorithm for P2P VoD systems, in which the scheduler is triggered by events such as peer churn or/and random seeks and the sender turns to first serve those requests with highest priorities through an optimal bandwidth allocation algorithm. The experimental results obtained from extensive simulations show that the priority-aware event-driven scheduling outperforms the time-driven scheduling in terms of both user experience and system scalability, especially during peer churn or when the source server is under-provisioned.An anchor prefetching algorithm is designed to further reduce seek latency. It differs from existing prefetching algorithms in that it adjusts the playback position to the closest anchor, avoiding the difficulty in predicting next playback positions. In the anchor prefetching, seeks can be satisfied instantly if the anchor has been already downloaded. The experimental results from the deployed systems show that the anchor prefetching significantly decreases seek latency. More specifically, forward seeks and backward seeks with less than 2 seconds latency are increased from 35% to 80%, from 65% to 78%, respectively.Based on the trace collected from a live P2P VoD system, namely GridCast, this dissertation analyzes how different caching policies can improve both scalability and user experience, evaluates their limitations, and presents some useful insights for future design and further optimizations. For example, we find that higher concurrencies delivery better content sharing and user experience, that caching leads to globally uneven load distribution, and that chunk misses caused by peer departure remains a major issue and keeps the biggest room for future improvements.From system traces, we identify that departure misses are the major cause of server load. Motivated by this finding, we examine how to use replication to decrease departure misses and thereby further reduce server load. We propose and evaluate a framework for lazy replication in P2P VoD. Lazy replication postpones data replication, trying to make efficient use of bandwidth. In this framework, two predictors are plugged in to create the working replication algorithm. Lazy replication with several predictors is compared with a naive eager replication algorithm. We find that lazy replication is more efficient than eager replication, even when using two simple predictors. With these two simple predictors, lazy replication can decrease server load by 15% from multi-video caching with only a minor increase in network traffic.