ELF: shared Cache management through eliminating dead blocks and filtering less reused lines

Modern CMP processors usually employ shared last-level cache (LLC) based on LRU replacement policy and its approximations. However, as the LLC grows in capacity and associativity, the performance gap between the LRU and the theoretical optimal replacement algorithms has widened. Various alternative cache management technologies have been proposed to resolve this problem, but they only cover a single type of memory access behavior, and exploit little frequency information of cache accesses, thus have limited performance benefits. In this paper, we propose a unified cache management policy ELF which can cover a variety of memory behaviors and exploit both recency and frequency information of a program simultaneously. Motivatedbythe observation that cache blocks often exhibit a small number of uses during their life time in the LLC, ELF is designed to (1) predict dead lines through a counter-based mechanism and evict them early, (2) filter less reused blocks through dynamic insertion and promotion policies. Thereby, the potentially live blocks are retained and most of the working set keeps undisturbed in the ELF managed L2 cache. Our evaluation on 4-way CMPs shows that ELF improves the overall performance by 14.5% on average over the LRU policy, and the performance benefit of ELF is 1.06x compared to PIPP and 1.09x compared to TADIP.