Results of File System Workloads

Figure 14: Comparison of the performance of different array alternatives under two file system workloads: (a) Cello disk 2, housing ``/users'', and (b) Cello disk 6, housing ``/var/spool/news''. Each data point represents the performance of the best configuration based on a given array alternative. The rectangular labels show the degree of mirroring (or replication) used in the configurations. The unlabeled configurations in the second figure have identical degrees of mirroring (or replication) as their counterparts in the first figure.

Although the target workload of an EW-Array is TPCC-like transaction processing applications, it is natural to ask whether it works for other workloads. Figure 14 shows the performance results of two file system workloads that are selected from the HP ``Cello'' trace. Cello is a two month trace of a file server supporting simulations, compilations, editing, and reading mail and news. We use the traces of two disks during the week from 5/30/92 to 6/6/92. Disk 2 houses user home directories and disk 6 houses a news archive. To compensate for the relatively lower I/O rates of the trace, we speed up the trace playing rate by a factor of four. The difference between Figure 14(a) and (b) is due to the different locality and the different read/write ratio of the two workloads. Cello disk 2 exhibits a higher degree of locality: the average seek distance on this disk is about half of that of disk 6. Cello disk 6, on the other hand, experiences a higher write ratio: 63.2% on disk 6 versus 45.2% on disk 2. Therefore, an SR-Array performs best for disk 2 by preserving locality and balancing the reduction of seek and rotational delays; while an EW-Array excels for disk 6 by aggressively optimizing write latency.