A Case for Exploiting Memory-Access Persistence

Memory access l atencies have become a major bottleneck in the performance of modern computer systems. It has been observed that the SPECint2000 benchmark suite suffers particularly high performance penalties from instruction and data cache misses. Current hardware-based and softwarebased prefetch mechanisms cannot always learn access patterns early enough to avoid compulsory data cache misses. Memory-access persistence is based on the idea that long-latency memory accesses tend to repeat themselves over subsequent executions of a program, despite changing inputs. By logging long-latency memory accesses during execution and correlating those accesses and their patterns across multiple executions driven with varying input sets, a program can be dynamically optimized to predictively prefetch these memory locations during subsequent executions. In this paper, we determine the extent to which memory-access persistence exists in modern applications.