B.S. White and K. Skadron
Tech Report CS-2000-06, Univ. of Virginia Dept. of Computer Science, Feb., 2000.
Prefetching is a well-known technique for mitigating the von Neumann bottleneck. In its most rudimentary form, prefetching simplifies to sequential lookahead. Unfortunately, large classes of applications exhibit file access patterns that are not amenable to sequential prefetching. More general purpose approaches often use models to develop an appropriate prefetching strategy. Such models tend to be large, thus preventing a kernel implementation which would lead to user transparency and more efficient execution. This work applies the target cache approach of path-based branch target prediction to file system prefetching to combat these deficiencies. The feasibility and worth of such a design are evaluated against a number of parallel applications popular in the scientific community.