a timely fashion need to be developed. Finally, since accesses to CM data is typically sequential in nature, improved buuer page replacement policies that exploit the sequentiality of access need to be devised. We h a ve addressed a number of the research issues described in the paper in the context of the Fellini multimedia storage manager being implemented at AT&T Bell Labs. A framework for the storage and retrieval of continuous media data. A case for redundant array of inexpensive disks (raid). In Proceedings of ACM SIGMOD'88, pages 109{116, June 1988. 4 reduces the additional load on disks in case of a disk failuree however, additional buuer space is required to store the pre-fetched blocks. On the other hand, buuer space overheads can be reduced by not pre-fetching blockss however, a drawback is that bandwidth for retrieving the additional blocks needs to be reserved on each disk (this bandwidth goes unused in the absence of failures). Thus, the above trade-oos must be taken into account when designing fault-tolerant s c hemes for multimedia servers. By having requests share a global pool of buuer pages, the number of I/O requests can be considerably reduced, thereby enabling a larger number of requests to be serviced 4]. For example, if two requests for a clip arrive at an interval of 5 or 10 seconds, then by c a c hing the pages accessed by the rst request, disk accesses for the second requests can be totally eliminated. An important research issue is that of buuer page replacement p o l i c y. Existing page replacement policies may be unsuitable since they exploit neither the knowledge of outstanding requests nor the fact that CM clip data is accessed predominantly sequentially. F or example, consider a server with 100 buuer pages of size d that follows the least recently used (LRU) policy (this is followed in most conventional storage servers). Let data retrieval for clips C 1 and C 2 be initiated at some round k. T h us, at round k + 51, the rst page of C 1 and C 2 are replaced to make room for the 51st page. If a request for C 1 were to arrive in round k + 52, then with LRU, pages for the second request for C 1 would need to be accessed separately from the disks (since the page to be retrieved for …