On the applicability of simple cache models for modern processors

Abstract

Cache performance estimation is the first step in assuring good cache utilization and hence application performance. However, it is difficult to create good cache models as the implementation of commercial caches is complex, constantly evolving, and, protected information. As a result many practical compilers use simple cache models such as Fully Associative LRU Cache (FALC) model. In this paper we quantify the applicability of the FALC model for three modern processors. Our investigation reveals that the applicability is both application and architecture dependent. This insight is used to develop a model for an early (i.e. no profiling required) identification of applicability: Early Picking Criterion. The Early Picking Criterion is developed using synthetic benchmarks and validated with 15 memory intensive SPEC CPU2006 benchmarks. All applications identified by the Early Picking Criterion demonstrate high applicability.