Prediction oriented analysis of optimal replacement

Abstract

The optimization of memory latency is always an important bottleneck to improving the performance of computer systems. The memory system, especially the last-level cache (LLC) as the important method to solve the “Memory Wall” problem, its management has become a key factors of influencing the performance of processor. And prefetching technology can improve the overall performance of the system by reducing pipeline stalls according to the temporal and spatial locality. This article is based on the characteristics of different workloads to study the performance of state-of-art LLC management policies with prediction technology. We achieve Bimodal Insertion Policy (BIP) which can adapt to changes in the working set. In order to further reduce the cache miss rate, we use the Set Dueling mechanism to dynamically choose the best replacement policy between Static Re-Reference Interval Policy (SRRIP) and Bimodal Re-Reference Interval Policy (BRRIP) based on the historical information [13]. We take SPLASH-2 as the benchmark to test the performance of these replacement policies. Finally we give a summary on the characteristics of different kinds of policies.