Potentials and limitations of fault-based Markov prefetching for virtual memory pages

Fault-based Markov prefetching for virtual memory pages (VMP) is examined. Markov prediction based only on the sequence of program-issued faults are shown to achieve reasonably high levels of accuracy for some scientific applications. Using the fault sequence to predict future faults makes speculative prefetching of VMPs feasible to implement. Further, despite high prediction accuracy, it is difficult in practice to obtain speedup from one-fault ahead speculative prefetching as this approach requires fault rate and fault latency sufficiently low that I/O overhead is not prohibitive in the first place.

Duke Scholars

Author Jeffrey S. Chase Computer Science