Check out the new USENIX Web site. next up previous
Next: Characterizing the Workload Up: Evaluating Predictive Prefetching Previous: Measuring Predictive Prefetching

Andrew Benchmark

Phase five of the Andrew benchmark [7] features a basic build of a C program. Although this benchmark is quite dated, to our knowledge it is the only existing file system benchmark that has been widely used and accurately portrays the predictive relationship between files. For these reasons our first benchmark was the build from phase five of the Andrew benchmark [7].

Initially predictive prefetching kernels were able to reduce the total I/O latency for this benchmark by 26%. From these tests we observed that to achieve greater reductions in I/O latency, our models would need to predict further ahead than merely the next event. So we modified PCM to create Extended Partitioned Context Modeling. Prefetching based on EPCM, was able reduce the total elapsed time by 12%, and remove almost all (90%) of the I/O latency from this benchmark.


Tom M. Kroeger