为了减少片上多核处理器(Clip multi-processor,CMP)末级共享Cache中的干扰,根据应用程序的存储访问频率特性,提出了一种基于替换算法的可重用数据预测机制。当末级共享Cache的数据将被替换时,先检测此数据的历史访问信息,根据历史访问信息过滤出会被重复使用的数据,并将其保存在片上专用存储器中。仿真结果表明:本文的可重用数据预测机制将IPC(Instruction per clock)平均提高了2.9%,平均减少了应用程序中22.69%的有害替换,有效地减少了Cache抖动。%To reduce interference of the shared LLC on a chip Multi-Processor(CMP) architecture,we propose a Replacement policy based Reusable Data Predicting mechanism(RRDP) to select possible reused data from replaced data according to the memory access frequency characteristics of application.When some data will be replaced in LLC Cache,we filter out predicted reusable data by checking their history reference information,and store them into a dedicated on-chip memory.Our simulation results show that the proposed mechanism increases the IPC of programs by 2.9% on average and reduces the harmful replacements by 22.69%,so it can effectively reduce Cache thrashing.
展开▼