Optimal prefetching via data compression

Abstract

A form of the competitive philosophy is applied to the problem of prefetching to develop an optimal universal prefetcher in terms of fault ratio, with particular applications to large-scale databases and hypertext systems. The algorithms are novel in that they are based on data compression techniques that are both theoretically optimal and good in practice. Intuitively, in order to compress data effectively, one has to be able to predict feature data well, and thus good data compressors should be able to predict well for purposes of prefetching. It is shown for powerful models such as Markov sources and mth order Markov sources that the page fault rates incurred by the prefetching algorithms presented are optimal in the limit for almost all sequences of page accesses.<>