Playing the Trace at Original Speed

Figure 10: Comparison of TPC-C I/O response time on several disk array configuration alternatives as we vary the number of disks in the array. The SR-Arrays and EW-Arrays are labeled with their configuration parameters: an ``RaSb'' label denotes a $D_s \times D_r = b \times a$ SR-Array configuration, and an ``MaSbDc'' label denotes a $D_m \times D_s \times D_d = a \times b \times c$ EW-Array configuration.

We play the TPC-C trace at original speed in the experiments reflected in Figure 10. It compares I/O response times of the optimally configured EW-Arrays against those of the RAID-10s and the optimally configured SR-Arrays as we increase the number of disks. In these experiments, the second and subsequent replicas, if any, are propagated in the background. An SR-Array generally outperforms RAID-10 because its combination of striping and rotational replication balances the reduction of seek and rotational delays better. An EW-Array outperforms both because of its substantially lower write latency, which also enables a higher degree of replication, which in turn lowers read latency. As we increase the number of disks, the performance benefit derived by an SR-Array from an increasing number of disks is larger than that of an EW-Array because the aggressive rotational delay reduction of the former benefits both reads and writes, while the write latency on an EW-Array is small to begin with and further improvements are marginal. In general, far fewer disks are necessary to achieve a specific latency goal on an EW-Array.