Targeted data prefetching

Abstract

Given the increasing gap between processors and memory, prefetching data into cache becomes an important strategy for preventing the processor from being starved of data. The success of any data prefetching scheme depends on three factors: timeliness, accuracy and overhead. In most hardware prefetching mechanism, the focus has been on accuracy - ensuring that the predicted address do turn out to be demanded in a later part of the code. In this paper, we introduce a simple hardware prefetching mechanism that targets delinquent loads, i.e. loads that account for a large proportion of the load misses in an application. Our results show that our prefetch strategy can reduce up to 45% of stall cycles of benchmarks running on a simulated out-of-order superscalar processor with an overhead of 0.0005 prefetch per CPU cycle. © Springer-Verlag Berlin Heidelberg 2005.