Data caching and data allocation for multimedia servers.

摘要

In this dissertation, we investigate methods for efficiently managing two important system resources, storage and bandwidth, to obtain better system performance in the distributed multimedia system. The storage devices are usually organized as hierarchical storage systems in multimedia servers. We propose a novel Space Efficient Pipelining (SEP) technique to reduce the user request latencies in the hierarchical storage system. A unified disk space is used not only to cache frequently referenced objects, but also to provide staging buffers for playing back objects. We also develop three buffer replacement policies to enhance storage and bandwidth management. The simulation results indicate that the data staging method used by SEP is very efficient, and each of the enhancement features offers significant additional improvement. In a computer system, disk I/O bandwidth is limited and caching data in the memory buffer can reduce the disk I/O bandwidth requirement. However, the traditional LRU memory page replacement algorithms are not suitable for continuous video objects. We propose a Bi-directional Fragmental Pipelining (BFP) technique and its variable buffer size data-caching scheme BFP_V to reduce the disk I/O bandwidth requirement for multimedia servers. Our mathematical analysis shows that the BFP technique is superior to the traditional unidirectional pipelining technique in terms of memory buffer space requirement. We further demonstrate that the memory buffer management using BFP_V is better than that of using the fixed buffer size approach BFP_F. Our simulation results quantitatively confirm our analysis. For distributed video servers, system performance depends on data allocation across servers in addition to the efficient management of storage and I/O bandwidth. We propose two data allocation algorithms, Bandwidth Weighted Partition (BWP) algorithm and Popularity Based (PB) algorithm, based on the bandwidth and storage capacity limits of the distributed multimedia servers. We compare those two algorithms with the traditional Round Robin (RR) data allocation algorithm. The analysis and simulation studies show that the BWP algorithm provides optimal system performance in the ideal situation. On the other hand, the PB algorithm is a simple and practical alternative for distributed video servers. It provides near optimal system performance under any system condition.