An overlay agent framework for multimedia delivery services
Zhonghang Xia, I‐Ling Yen · 2004
While rapid advances in the Internet and the World Wide Web have made many multimedia-based services a reality, it has also drastically increased the media server load and Internet delay. In this dissertation, we aim at enhancing the performance of media delivery services over the Internet. We propose a proxy assistant framework which focuses on highly distributed approaches and integrated solutions. To achieve highly distributed solutions, we move some high level services from the media server to the proxy servers. First, we propose a novel distributed admission control algorithm, where proxy servers reserve the media server's disk bandwidth according to their demands, and make admission decisions based on the granted bandwidth. To further improve the media system performance, we integrate prefix caching, batching, and patching schemes with the distributed admission control algorithm. Due to prefix caching, the admission process does not simply consider the currently available bandwidth. A request can be admitted in advance as long as the delivery session can be established before the cached content is exhausted. The prefix playback duration also improves the effectiveness of batching. More requests which requests the same media can be included in the same batch group. This reduces the number of requests to be forwarded to the server and further reduces the network traffic. Besides distributed request processing, we have also developed a new, efficient approach for multicast tree construction. The approach includes two stages and combines centralized and distributed mechanisms. First, the media server constructs an initial multicast tree for the requests arriving in a short batch window, using a centralized approach. Then, subsequent requests can join the multicast tree using a distributed approach according to the traffic load prediction. Traditionally, the routing path requires frequent revision due to the multicast group membership change. Such computation may create a significant overhead. To reduce the overhead, we develop a neural network traffic load predictor. It predicts the future traffic of neighboring proxy servers and determines the point to join the multicast tree such that the traffic load on the tree nodes remains balanced. Our approach shows significant performance improvement compared to conventional algorithms.