Clean-prefetcher: look-ahead prefetching without cache pollution

Abstract

In this letter, we propose a novel prefetching technique that is free from cache pollution and thus achieves high performance for multicore processors. Unlike the conventional prefetchers that cause incorrect predictions, the proposed prefetcher reads instructions in advance only in determined paths and charges dynamic random access memory (DRAM) cells storing instructions in undetermined paths via refreshing DRAM cells. The DRAM cells highly charged will be quickly accessed. Since caches served by our prefetcher always store useful instructions, as a result, they are free from cache pollution that results in lower cache hit rate. In the case that SPEC CPU2006 benchmarks run on an 8-core processor, the proposed prefetcher consumes more 3.2% DRAM power, but achieves 12% higher performance on average.