REPORT FOR TRANSFER OF REGISTRATION FROM MPHIL TO PHD Efficient Arbitration and Bridging Techniques for High-Performance Multimedia Servers

Advances in semiconductor integration and network technology make it possible for the first time to create sophisticated distributed multimedia applications for a large audience. The client-server paradigm of communication is a popular choice for a number of important distributed multimedia applications, such as video-on-demand. Certain forms of multimedia data, such as audio-visual data must be delivered in continuous, high-volume streams of data. Hence multimedia servers face more complex challenges than conventional file servers. This research study focuses on advanced architectural features to improve the cost/performance ratio of multimedia servers based on conventional workstation architectures. The majority of prior research in the context of multimedia servers has concentrated on issues such as data placement and access strategies for secondary storage, while other areas have not enjoyed the same level of attention. Of particular interest is the area of arbitration: to meet deadlines imposed by real-time data streaming, server components must compete for access to shared resources, such as the system bus. In such an environment, arbitration should be able to guarantee worst-case access latency, so that buffers in the I/O controllers can be dimensioned accordingly. Simulation shows that multimedia servers can behave very differently to multiprocessor machines, which are often used as reference models in research on arbitration. Small variations in the system load (i.e. the number of concurrent data streams delivered by the server) can have severe effects on the arbitration latency. The results are consistent across a range of arbitration protocols, with no clear performance advantages for any particular protocol. Hence deterministic resource scheduling with fixed latency is suggested as an alternative approach to “ad-hoc” arbitration. A second issue for traditional multimedia servers is limited scalability: centralised resources tend to become bottlenecks with growing user demands. Two novel local buffering schemes are proposed to eliminate this bottleneck and increase server scalability with little associated cost. The first technique, “cell batching”, aims at improving ATM adapter efficiency by transferring a variable number of cells in a single bus transaction. Simulation results show that cell batching can reduce bus occupancy of ATM adapters by up to one third. The second approach is complementary to cell batching, and not restricted to ATM adapters. It uses bus bridges and local DRAM to encapsulate storage and network subsystems in a module, which can be replicated to enable concurrent data streaming. Such streaming modules can be realised mainly with inexpensive off-the-shelf components and are transparent to existing infrastructures. Simulation shows this approach to improve server scalability and performance significantly. The results also apply to other environments, where high-volume continuous data streams must be supported.