Adaptive Data Prefetching for High Performance Processors

Projet ANR PersyvalInternational audienceNowadays, one of the main limiting factor in processordevelopment is the increasing speed gap between efficient processingelements and slow main memories. To reduce this limitation, prefetchingmechanisms, implemented in memory hierarchy, attempt to predict thefuture data needed in local memory. However, classical proposedsolutions are designed for specific access sequences but lack ofefficiency considering irregular and heterogeneous access patterns.Adaptive data prefetching aims to tackle the limitation of existingprefetch solution by self-reconfiguring its parameters according to thedynamic variations in the current access sequence. We propose to use amodeling of the main existing access patterns and an on-line monitoringof the access sequence to detect and trigger the data prefetching