关键词: 访问模式, 替换算法, 多核共享末级缓存, 组竞争

Abstract:

Many proposals about the optimization of cache management policies depend on set dueling, and they choose policies according to the results of miss rates. However, they are facing challenges in multicore since they have lower accuracy and higher storage overhead. The main reason is that they only consider the results of cache replacement policies rather than the characteristic of cache access sequences (or access pattern). Therefore, an online access pattern identification method, complex stack distance weighting method was proposed, and an adaptive cache replacement policy, least recently used-C (LRU-C), was put forward. Its basic idea is to adjust the insertion position according to the result of the online access pattern identification. Experiment in the full-system Simics running 18 benchmarks from SPEC CPU2000/2006 shows that the performance of LRU-C is better than those of dynamic insertion policy,re-reference interval prediction,thread aware dynamic insertion policy and promotion/insertion pseudo-partitioning.

Key words: access pattern, replacement policy, last-level-cache, set dueling