A cache replacement policy based on re-reference count

Abstract

The cache replacement policy is a major factor which determines the effectiveness of memory hierarchy. The replacement policy affects both the hit rate and the access latency of the cache. It decides the cache block to be replaced to give room for the incoming block. The replacement policy has to be chosen in such a way that the cache misses are reduced. Last level cache misses causes hundreds of stall cycles due to the need for main memory access. So last level cache misses are given more priority over L1 cache misses. The traditional cache replacement policy used is Least Recently Used (LRU) policy. LRU policy favors workloads having cyclic access pattern which fit in cache, but it exhibit thrashing behavior for memory-intensive workloads that does not fit in the available cache. Hence, many replacement policies were proposed to improve the miss rate for last level caches while maintaining low hardware overhead and minimum design changes. Here a novel replacement policy which is a variation of LRU Insertion policy (LIP) based on re-reference count is proposed. The promotion policy in LIP is modified to implement the new policy which is based on the re-reference count. The proposed replacement policy was implemented and performance comparisons of the replacement policies were done on Gem5 simulator using cpu2006 benchmarks. Under this policy, the memory intensive workload mcf attains 64% improvement in L2 cache miss rate over LRU policy.