CRP: Context-directed Replacement Policy to Improve Cache Performance for Coarse-Grained Reconfigurable Arrays

Abstract

This paper proposed a context-directed replacement policy (CRP) to improve cache hit rate for Coarse-grained reconfigurable arrays (CGRA). CRP updates the replacement priority according to the usage rate of configuration contexts. Once finding two candidates of configuration contexts with the highest replacement priority, CRP reserves the larger context in cache and evict the smaller one. Using five different access patterns as benchmark, experimental results show that CRP outperforms LRU and RRIP replacement policies, especially for scan and mixed access patterns. Under the test of random access sequence, CRP can averagely improve cache hit rate by 36% and 20%, compared to RRIP and LRU, respectively.