摘要

The Least recently used (LRU) replacement has been widely employed by on-chip shared caches. With the increase of cache associativity, the performance gap between LRU and theoretical optimal replacement algorithm is significant, motivating us to design alternative replacement algorithms to improve cache performance. We propose a Partition-based cache replacement (PCR) algorithm to manage multi-core shared L2 caches. PCR takes into consideration recency and frequency information of cache access, and cache resource interference among multiple competing applications. PCR is effective to filter out the less frequently reused blocks through a dynamic insertion/promotion policy based on the target partitions, and ensures that some fraction of the working set is retained in the cache to contribute to more hits. Our evaluation shows that PCR improves performance of an 8-core system by an average of 11.3% over LRU-based cache.