Here we present the results from our benchmark tests on predictive prefetching and how they affected the design of our implementation. We ran our tests on a Pentium based machine with a SCSI I/O subsystem and 256 megabytes of RAM. To evaluate our implementation we selected four application based benchmarks that provide a variety of workloads. In our test we saw predictive prefetching reduce the time spent waiting for I/O by 31% to 90%. While read latencies saw reductions from 33% to 92%, the reductions in elapsed time, ranged from 11% to 16%
Our test machine had a Pentium Pro 200 CPU, with 256 megabytes of RAM, an Adaptec AHA-2940 Ultra Wide SCSI controller and a Seagate Barracuda (ST34371W) disk. All kernels were compiled without symmetric multi-processor (SMP) support. This machine had Gnu ld version 126.96.36.199.19, gcc version 188.8.131.52 and Glimpse version 4.1.
For these tests, we focused primarily on two measures--the read latency and total I/O latency. We determined read latency from instrumentation of the read system call. Since this did not include I/O latencies from page faults, open events, and exec calls, we also considered the total I/O latency. We bound total I/O latency by taking the difference between the elapsed time and the amount of time the benchmark was computing (time in the running state or system time plus user time). This gives us the amount of time that the benchmark spent in a state other than running, which served as an upper bound on the amount of time spent waiting on I/O. Since our test machines had only the bare minimum of daemon processes, this measure is a close approximation of the total I/O latency of that benchmark.
Each test consisted of 3 warm up runs that eliminated initial transient noise and allowed the models time to learn. Then 20 runs of the test benchmark provided enough samples for us to generate meaningful confidence intervals assuming a normal distribution and statistically significant measurements. Unless otherwise stated, the I/O caches were cleared between each run of the benchmark.