An Adaptive Demotion Policy with Considering Temporal Locality

Abstract

In the upper-level caches like as L1 caches, the temporal locality mainly affects the performance of the caches. Although the Adaptive Demotion Policy (ADP) can achieve the performance improvement at the upper-level caches compared to the conventional hardware suitable replacement policies, the most recently used block will be evicted when all the blocks in the accessed set have the same priority value. Hence, there is a possibility of decreasing the performance of the cache by spoiling the temporal locality. To avoid this problem, this paper proposes a cache replacement policy which focuses on preventing the eviction of the most recently used block, and the policy is called Adaptive Demotion Policy with considering Temporal Locality (ADPTL). At a cache hit, if the priority values of all the blocks in the accessed set except the hit block are the highest value, the ADPTL decreases these priority values by one. This operation prevents the eviction of the most recently used block. The performance evaluation shows that the ADPTL achieves the reduction of the cache misses and the IPC improvement at the L1 caches compared to the LRU policy and the ADP.