Evaluating a Model for Cache Conflict Miss Prediction

Abstract

Cache conflict misses can cause severe degradation in application performance. Previous research has shown that for many scientific applications majority of cache misses are due to conflicts in cache. Although, conflicts in cache are a major concern for application performance it is often difficult to eliminate them completely. Eliminating conflict misses requires detailed knowledge of the cache replacement policy and the allocation of data in memory. This information is usually not available to the compiler. As such, the compiler has to resort to applying heuristics to try and minimize the occurrence of conflict misses. In this paper, we present a probabilistic method of estimating cache conflict misses for set-associative caches. We present a set of experiments evaluating the model and discuss the implications of the experimental results.