MimdRAID: Low Latency Secondary Storage

Abstract:

Disk arrays can provide scalable bandwidth with increasing number of
disks, but they do not address latency. In this paper, we study how to
improve secondary storage read latency by systematically increasing
the ratio between disk heads and usable capacity. The most direct
technique is to build ``faster" disks by altering disk geometry. In
addition, as an alternative to building such disks, we also describe
how we can emulate these faster disks using multiple conventional
disks. We show that all these techniques are governed by a common
principle that the overhead-independent part of the latency improves
by a factor of the square root of the amount of extra resources. We
evaluate some of these latency reduction techniques on a prototype
implementation which we call MimdRAID. On a 12-disk system, for random
reads, we achieve a 2x improvement in latency and 21x improvement in
throughput compared to a single drive. This is 80% to 95% better than
the throughput achieved on a conventional mirrored system.