While four-disk arrays are widely available, we also tested a single disk configuration and smaller and larger arrays. As shown in Table 8, the original, non-hinting applications are unable to derive much benefit from additional disks.
| Benchmark | Number of disks | |||
|---|---|---|---|---|
| 1 | 2 | 4 | 10 | |
| Agrep | 23.8 | 34.1 | 21.4 | 20.1 |
| Gnuld | 93.7 | 101.3 | 89.5 | 82.8 |
| XDataSlice | 303.5 | 292.0 | 324.6 | 265.7 |
As shown in Figure 5, all the benchmarks receive significantly less benefit from speculative execution when there is only one disk because prefetching can only be overlapped with computation. The performance of speculating Gnuld degrades with one disk because erroneous prefetches consume scarce bandwidth, delaying service for the application's demand requests. As we discuss in Section 5, we believe that simple mechanisms can be employed to address this problem.
One objection to our assumptions - that disk-bound applications will be running on machines that have both disk arrays and no competing tasks to run on the processor - is that more than one disk is attached to a machine only if it is a shared server. However, Rochberg has shown that the TIP system can be effectively extended to allow clients to prefetch from distributed file servers with multiple disks [Rochberg97]. It is these ``personal'' clients that will be most rich in excess processor cycles.
As shown in Figure 5, the benefit of the hinting applications increase, and their runtimes decrease, when I/O parallelism is available. The benefit obtained by the manually modified applications increases monotonically with the number of disks since these applications always issue enough hints to take advantage of the additional disks. For Agrep, the benefit of the speculating application mirrors that of the manually modified application for the 2 and 4 disk configurations. However, due to the dilation factor discussed in Section 4.4, speculative execution is not far enough ahead of normal execution to issue sufficient hints to keep 10 disks busy. For Gnuld, data dependencies limit hint generation, and therefore the degree to which the speculating application is able to utilize additional disks. For XDataSlice, however, speculative execution generates more than enough hints to take advantage of the additional I/O parallelism.
Figure 5: Performance improvement as the number of
disks is varied.