一种降低末级高速缓存污染的分阶段自适应动态插入策略

A Phase-Based Adaptive Dynamic Insertion Policy for Reducing Last Level Cache Pollution

对多种末级高速缓存插入策略进行分析,并在动态插入策略DIP的基础上提出一种分阶段自我调整的动态插入策略,用于消除局部性差数据访问末级高速缓存造成的不良影响。实验结果表明,与现有LRU替换算法相比,此方法将末级高速缓存的MPKI平均降低了7.07%,即使与动态插入策略DIP相比,此方法也获得进一步的性能提升,末级高速缓存MPKI平均降低了4.36%。

The authors analyze multiple last level cache insertion policies and propose a phase-based adaptive dynamic insertion policy based on DIP （dynamic insertion policy） for eliminating the harmful effect of non-reusable data. The experiment result shows that in comparison with LRU （least recently used）, the proposed approach can reduce the last level cache misses per thousand instructions, MPKI, 7.07% on average; even in comparison with DIP policy, the new approach can also improve last level cache performance in further, and average MPKI is reduced by 4.36%.