异构内存系统全局优化的数据预取算法

Data Prefetching Algorithm for Globally Optimizing Heterogeneous Memory System

鉴于现有的数据预取算法不能满足高效能异构计算系统对动态随机存取存储器(DRAM)和非易失性存储器(NVM)相结合的新型异构存储器高效访问的要求,提出了一种模拟退火的全局优化数据预取算法(SADPA)。该算法在启发式搜索模拟退火算法的基础上,引入了随机因子,以避免局部最优,从而确定了全局优化阈值以预取NVM页面的有效数量。实验结果表明,该算法相对于静态阈值调整算法,平均访问延时降低了4%,每个时钟周期内的平均指令数(IPC)增加了10.1%;对于cactusADM应用,该算法相对于软硬件协同的动态阈值调整算法,系统能耗降低了3.4%。

Due to the existing data prefetching algorithms can ’t meet the requirements of the novel heterogeneous memory system combining the dynamic random access memory (DRAM) with the nonvolatile memory (NVM) in high energy-efficiency heterogeneous computing systems,a simulated annealing data prefetching algorithm (SADPA) was proposed.It was a heuristic search inspired simulated annealing algorithm,in which a random factor was introduced to confirm the global optimal threshold and the valid number of prefetching NVM pages.The results show that the average accessing latency of SADPA is 4% lower than that of the static threshold adjustment algorithm,and the average instruction per cycle (IPC) of the SADPA is 10.1% greater than that of the static threshold adjustment algorithm.Besides,the systemic power supported by SADPA,as for the cactusADM,is reduced by 3.4% compared with the cooperative hardware/software dynamic threshold adjustment algorithm.