Performance studies of digital video in a client/server environment

Abstract

Networked multimedia services cover a wide spectrum of possible applications. One system architecture of interest is that of a fileserver delivering prerecorded digital movie clips for continuous time playback to clients. This problem is similar to that of delivery of real-time live video to clients, but is more forgiving in that some amount of the movie clip may be prebuffered at both fileserver and client to smooth out variations in the system. On the other hand, such a service requires fileservers capable of maintaining continuous streams of data, or their equivalent. The “fileserver delivering movie clips” is also at the heart of many proposed applications. Although, it is unlikely that all clients would be playing movie clips at once, it is important to understand how many simultaneously active clients a given network architecture can support.

Delivery of stored digital video from a shared fileserver to remote clients is a systems architecture problem with many interacting components. Key items include the physical network, the network protocols, the architecture of both client and server, filesharing and operating system software, and the characteristics of the disk drives used in video storage. If components of the system are not well-matched, we would like to locate the bottleneck components, find how they interact in the system, and explore means of correcting the problem.

This paper presents a study of a commercial system, Apple's QuickTime^Tm version 1.1, employed in a client/server environment. The measurements studies show clearly the pitfalls of a system that is “out of balance”. Bottlenecks and upper bounds are identified, first a simple analytical model is employed to examine small improvements, then simulation is employed to further explore the space of possible system enhancements.