BC-AMAT

Abstract

The wall problem requires not only the use of increasingly aggressive techniques designed to reduce the latency of memory system, but also the raise of more accurate memory metrics. C-AMAT, an extension of AMAT that considers both locality and concurrency of memory accesses, can evaluate the performance of modern memory system more accurately. However, C-AMAT only involves those cycles consumed by memory accesses, ignoring the blocked time caused by some techniques like hardware prefetch, which may result in inaccurate evaluation. In this paper, we propose a more comprehensive memory metric called Blocked C-AMAT (BC-AMAT). It extends C-AMAT to take the blocked cycles into consideration. Experimental results show that BC-AMAT correlates much better with IPC than C-AMAT does when a few prefetch strategies are applied both in single-core mode and multi-core mode. In addition, a case study is provided in which BC-AMAT is used to adjust prefetching degree dynamically. The result shows that BC-AMAT achieves higher performance improvement than C-AMAT, demonstrating its usefulness in system optimization. BC-AMAT is more accurate and comprehensive than C-AMAT in evaluating modern memory systems, meanwhile, provides more insight for architecture design.