Ensuring availability of data in the presence of disk failures has been recognized as a difficult problem. In addition to investigating this problem, this study investigates availability of bandwidth in support of continuous display once failures occur. This is important for continuous media (audio and video clips) servers for several reasons. First, the service time of a clip might be in the order of tens of minutes, e.g., video-on-demand. Second, continuous media servers must schedule the available disk bandwidth in a manner that displays a clip free from disruptions and delays, i.e., hiccup-free display. They almost always employ an admission control module based on the available disk bandwidth. In the presence of disk failures, the aggregate disk bandwidth is reduced. If the admission control has scheduled all available disk bandwidth then a failure would terminate one or more of the active displays. This is undesirable for certain applications, especially those that strive to sell an experience, e.g., video-on-demand.
Of course, failures are unavoidable and one must consider the costs associated with continued service. For the afromentioned applications, it is obvious that the system should strive to minimize both (a) the number of streams that are terminated due to disk failures, i.e., minimize the amount of disk bandwidth that becomes unavailable due to disk failures, and (b) the occurrence of such events. Thequestion is at what expense? This is where this study makes its greates impact by identifying the tardeoff between: Mean Time To Failure (MTTF), storage cost, and throughput of the system. It investigates alternative placement techniques for partiy blocks and scheduling algorithms during failed mode of operation. It does a through job of showing these tradeoffs at play with one another.
Copyright © 2000 by the author(s). Review published with permission.