The performance of MOD servers is highly constrained by the stringent requirements of multimedia data. Specifically, multimedia data must be presented continuously in time so that they can convey meaningful information. They also require high transfer rates. Without careful management of the limited system resources, both the available bandwidth of the storage subsystem and the available bandwidth to the network will run out rather quickly after supporting only a relatively small number of customers. Resource sharing strategies save precious resources by using the same set of resources in servicing multiple requests for the same multimedia content. Taking advantage of the high locality of reference in the access to multimedia data, resource sharing can significantly increase the numbers of customers that can be serviced concurrently.
The exploited degrees of resource sharing depend greatly on how MOD servers schedule the waiting requests. By scheduling the requests intelligently, a server can support more concurrent customers, can reduce their waiting times for service, and/or can meet some other performance objectives. Choosing an appropriate scheduling policy is very complicated by the presence of several design tradeoffs and the dependence of performance on customer waiting tolerance and server load.
This thesis addresses the scheduling problem in MOD servers. The main application of interest is video-on-demand (VOD). A VOD server maintains a waiting queue for every movie, selects an appropriate queue for service whenever the required set of resources becomes available, and services all requests in that queue together using only one stream. Scheduling policies for VOD servers include First Come First Serve (FCFS), Maximum Queue Length (MQL), and Maximum Factored Queue Length (MFQL). FCFS schedules the waiting requests based on their waiting times, whereas MQL and MFQL schedule them based on the queue lengths.
The main contributions of this thesis are providing a detailed investigation of scheduling policies and proposing two new policies, called Quantized First Come First Serve (QFCFS) and Enhanced Minimum Idling Maximum Loss (IML+). QFCFS combines the benefits of FCFS and MQL/MFQL, whereas IML+ improves the overall performance by exploiting minimum request waiting times. This study compares, mainly through simulation, various policies in terms of several performance metrics and examines the impacts of customer waiting tolerance and server load. The results show that the proposed policies suite different patterns of waiting tolerance and provide significant performance benefits.