@device(postscript) @libraryfile(Mathematics10) @libraryfile(Accents) @style(fontfamily=timesroman,fontscale=11) @pagefooting(immediate, left "@c", center "@c", right "@c") @heading(Storage Strategies for Fault-Tolerance Video Servers) @heading(CMU-CS-96-144) @center(@b(Elmootzabellah N. Elnozahy)) @center(August 1996) @center(FTP: CMU-CS-96-144.ps) @blankspace(1) @begin(text) We consider the problem of providing high availability in cluster-based video servers. The cluster acts as a parallel processor that provides the aggregate I/O and network bandwidths of the component machines. In such an environment, the failure of one server may affect the availability of the video service or its quality. Existing approaches to this problem fall into two categories. On one hand there are RAID-like schemes that store error correcting code (ECC) in addition to the video data. Should a failure occur, the unavailable data can be computed on the fly using the ECC and the service continues at the same quality. In cluster-based systems, however, the video data are distributed over several servers and there is no convenient point to reconstruct the missing blocks except at the client. Relying on the client for this task is not desirable as it may not have the necessary buffering or processing capacity On the other hand, some argue that the system could just continue operation without the need for the information on the unavailable host. The underlying assumption is that the human perception may tolerate the reduction in the frame rate that results from the temporary loss of data. This strategy, however, does not work well for compressed video. In this paper, we argue that substantial improvements can be obtained by taking into account the nature of the compressed video stream. We present two approaches for storing MPEG-1 video that helps in tolerating partial failures in server clusters at the expense of marginal degradation in the quality of service. The first approach restricts ECC storage and computation to reference frames, ignoring low-information frames. The second approach uses the structure of the compressed video stream to substitute unavailable high-information frames with equivalent low-information frames. @blankspace(2line) @begin(transparent,size=10) @b(Keywords:@ )@c @end(transparent) @blankspace(1line) @end(text) @flushright(@b[(15 pages)])