Optimizing the Management of Reference Prediction Table for Prefetching and Prepromotion

Abstract

Prefetching and prepromotion are two important techniques for hiding the memory access latency. Reference prediction tables (RPT) plays a significant role in the process of prefetching or prepromoting data with linear memory access patterns. The traditional RPT management, LRU replacement algorithm, can not manage RPT efficiently. This leads to that large RPT has to be used for the considerable performance. The cost brought from the large capacity limits the usage of RPT in real processors. This paper uses bimodal insert policy (BIP) and proposed scalar filter policy (SFP) in the RPT management. Owing to matching the using characteristics of RPT, BIP can reduce the RPT thrashing and SFP can filter the useless scalar instructions in it. After testing 8 NPB benchmarks on a full-system simulator, we find that our approaches improve the RPT hit rate by 53.81% averagely, and increases prefetching and prepromotion operations by 18.85% and 53.55% averagely over the traditional LRU management.