对多内核共享L2 cache时的cache污染问题进行了研究,认为内核猜测执行预测路径上的访存指令导致的cache污染会严重影响处理器性能.提出了一种基于置信度评估的cache污染过滤技术FCPC,该技术置信度评估机制对条件分支进行动态评估,并为每个cache数据行新增两个标志位--置信度评估标志(CET)和访问指示标志(AHT),根据评估结果将预测路径上的访存指令及其访存数据分别标识为高置信度和低置信度.在cache替换操作时根据CET和AHT就能优先将污染数据替换出去,增加了cache空间的有效利用率,从而提升IPC和cache命中率.实验结果表明,FCPC策略能够将IPC性能提升0.18%～4.86%,平均提升1.91%;同时能够将降低L2 cache缺失率降低0.65%～5.76%,平均降低2.57%.%This paper studies the cache pollution issue that affects the performance of a memory system in the shared L2 cache multi-core architecture, and believes that cache pollution caused by the speculative execution of memory reference instructions in the predictive path may affect the processor's performance seriously. Then the paper proposes a cache pollution filtration technique based on confidence estimation, called FCPC. FCPC proceeds the dynamic appraisal of each condition branch by using the confidence estimation mechanism, and filters the polluted cache data preferentially through adding two tags for each cache line as the ancillary flags, the confidence estimation tag (CET) and the accessing hint tag (AHT). According to the appraisal result, the memory reference instructions in the predictive path and their returned data in L2 cache are marked as high or low confidence separately. Then, when cache replacing operations are carried on,the polluted data in shared L2 cache can be swept out preferentially according to CET and AHT. As a result, the efficiency of cache space is increased. The simulation result indicates that, with dual-core configuration, the FCPC technique can accomplish 0.18％ ～4.86％, average 1.91％ promotions in IPC performance; and 0.65％ ～ 5.76％, average 2.57％ reductions in L2 cache miss rate can be reduced.
展开▼
