Reliability and Performance of Parallel Disks

Recently, parallel disk systems have emerged as a potential solution for achieving ultra-high capacity, performance and reliability at a reasonably low cost. This emerging technology is likely to have a great impact on the market of data-storage devices for mainframe computers and large database systems [Zachary:89]. The basic idea is to disperse (or stripe data blocks over a number of small and cheap physical disks. Accessing these disks in parallel guarantees a linear speedup in the transfer rates. Moreover, introducing storage redundancy promises an ultra-high level of availability that is impossible to achieve using usual approaches. This added redundancy can be exploited to further improve the system performance by reducing the seek and latency delays of the overall system. In this paper, we investigate a number of issues related to the design of systems of parallel disks. We start with a detailed performance analysis of these systems. In this respect, we obtain expressions for the di erent delays incurred; namely the queuing, seek, latency, transfer and overhead delays. Next, we consider the reliability requirements of these systems. We show that by adding a small amount of redundancy, the reliability of parallel disks can be made very high. Next, we present a number of potential hardware architectures; namely, bus oriented, sliced memory, and switching network organizations. Finally, we discuss a number of design decisions concerning redundancy, synchronization and grain size. This work was done while the author was with AT&T Bell Laboratories, August 1989.