Data prefetching with co-operative caching

Abstract

Recent research in data cache prefetching is found to be selective in nature: achieving high prediction accuracy over a set of selected references such as array access with constant strides. As a result, for applications where the memory latency is mainly due to data accesses in the set of non selected references of a program, they lose their effectiveness. In fact, their performance might be worse than that of the traditional, less accurate prefetch-on-miss scheme. To overcome this situation, we propose three cooperative cache techniques to assist data prefetching. They are: [1] default prefetching to increase the overall prefetch coverage; [2] block concept to perform variable distance lookahead prefetching; and [3] a spatial data buffer with load balancing to reduce the interference between spatial data and temporal data. To illustrate the potentials of these techniques, they were implemented on top of our previously proposed Instruction Opcode-Based Prefetching (IOBP) scheme (T.F. Chen, 1993). Trace driven simulation on SPEC92 showed that a 8 Kbytes data cache with a 512 bytes spatial buffer can achieve similar performance as a 32 Kbytes data cache through these techniques.