Designing an I/O Scheduler for Continuous Media

A growing set of multi-media applications require continuous ow of high volume data, e.g. video and audio, that can wholly reside only in secondary (and possibly tertiary) storage devices. Additionally, data accessed by such applications have temporal relationships that must be maintained at storage and retrieval time. In order to provide such high volume data in real-time large investment in secondary storage devices and mechanisms is necessary. In order to provide economical access to such data specialized storage servers called Continuous Media (CM) servers must be shared by a large community of users. In this paper we discuss the design of an I/O scheduler that schedules synchronized access to continuous media in real-time from secondary storage devices. To schedule multiple accesses to the storage server, we describe a control system model for a class of cyclical schedulers for continuous media that guarantee data rate. This model is then used to explore diierent trade-oos like the buuer space and slack time. Connguration issues like main memory requirements and control policies like admission control are then quantitatively analyzed. In addition to simultaneously accessing data, users will execute operations like Play and FastPlay that require data at varying rates, and FastForward and FastReverse operations that require data in a sequence that is diierent from the original recording. In order to implement such VCR-like operations without aaecting data rate guarantees requires a gradual build up of data until a time when it is safe enough to eeect them. We describe this phenomenon and the design equation that limits this rate of data build up. Problems involved in feasible implementation of this design are presented. In summary, a set of six design equations are presented that provide the designer of such a scheduler means to quantitatively decide admission control of new users, primary memory connguration, schedule computation, balancing real-time and non real-time workload, supporting stream operations sought by users, and to derive feasible implementations to realize such schedulers.