Hawkeye: Leveraging Belady’s Algorithm for Improved Cache Replacement

This paper evaluates the Hawkeye cache replacement policy on the Cache Replacement Championship framework. The solution departs from that of the original paper by distinguishing prefetches from demand fetches, so that redundant prefetches can be identified and cached appropriately. Evaluation on SPEC2006 shows that in the absence of prefetching, Hawkeye provides a speedup of 4.5% over LRU (vs. 3.4% for SHiP) on the single-core configuration and a speedup of 9.2% (vs. 7.1% for SHiP) on the four-core configuration. In the presence of prefetching, Hawkeye’s performance is marginally better than SHiP’s (2.25% speedup vs 2.09% speedup over LRU).