I/O performance of fully-replicated disk systems

Abstract

Mirrored disk storage is an accepted technique to enhance fault-tolerance of data through complete replication. Recent research suggested that in addition to higher reliability, mirrored disks can offer better I/O performance by either allowing parallel reads or by reducing the seek time in cases where shortest seek distance algorithms can be used. Accurate analysis of mirrored disk systems shows that there is a significant correlation between the distributions of the disk heads positions, which makes them both interdependent and non-uniform. Furthermore, after each write all disk heads move to the same position to form what the authors call a 'bundle'. Such phenomenon may seriously deteriorate system performance. Subsequent reads gradually break up the bundles. Finally, under low rates of request arrivals, the authors provide analytical expressions for optimal anticipation points for two and three mirrored disks under general read/write combinations. Their work makes the following contributions: (1) new heuristics for treating the 'bundling phenomenon' adaptively, for applications with unknown read/write ratios; (2) a new technique that improves the expected seek time and workload balancing by using 'shifted mirroring'.<>