摘要：Traditional caches employ the LRU management policy to drive replacement decisions.However, previous studies have shown LRU can perform significantly worse than the theo-retical optimum, OPT [1]. To better match OPT, it is necessary to aggressively anticipatethe future memory references performed in the cache. Recently, several researchers havetried to approximate OPT management by predicting last touch references [2, 3, 4, 5].Existing last touch predictors (LTPs) either correlate last touch references with executionsignatures, like instruction traces [3, 4] or last touch history [5], or they predict cache blocklife times based on reference [2] or cycle [6] counts. On a predicted last touch, the refer-enced cache blo ck is marked for early eviction. This permits cache blocks lower in the LRUstack–but with shorter reuse distances–to remain in cache longer, resulting in additionalcache hits.This paper investigates three mechanisms to improve LTP-driven cache management.First, we propose exploiting reuse distance information to increase LTP accuracy. Specif-ically, we correlate a memory reference's last touch outcome with its global reuse distancehistory. Second, for LTPs, we also advocate selecting the most-recently-used LRU lasttouch block for eviction. We find an MRU victim selection policy evicts fewer LNO lasttouches [5] and mispredicted LRU last touches. Our results show that for an 8-way 1 MBL2 cache, a 54 KB RD-LTP which combines both mechanisms reduces the cache miss rateby 12.6% and 15.8% compared to LvP and AIP [2], two state-of-the-art last touch predic-tors, and by 9.3% compared to DIP [7], a recent insertion policy. Finally, we also proposepredicting actual reuse distance values using reuse distance predictors (RDPs). An RDPis very similar to an RD-LTP except its predictor table stores exact reuse distance valuesinstead of last touch outcomes. Because RDPs predict reuse distances, we can distinguishbetween LNO and OPT last touches more accurately. Our results show an 64 KB RDPcan improve the miss rate compared to an RD-LTP by an additional 2.7%
