To address the memory wall issue of the high performance hybrid computing systems, this paper proposes a novel hierarchical explicit memory access mechanism based on the analysis of hybrid computing mode and the limitations of the traditional memory access mechanism. The proposed mechanism is implemented and evaluated on a multi-core hybrid computing system Engineering and Scientific Computing Architecture (ESCA). Experimental results show that the hidden of memory access latency can occupy 56% of the total run time and achieve 1.5 times speedup with the kernel of DGEMM, which proves that the proposed memory access mechanism is beneficial to fill the gap between computing and memory, thus improving the system efficiency.%针对高性能混合计算系统中的存储墙问题,在分析其计算模式特点及传统访存机制局限性的基础上,提出适用于混合计算系统的层次化显式存储访问机制,并基于ESCA多核处理器系统进行实现和评测.实验结果显示,针对核心应用程序DGEMM,延迟隐藏能够占据整体运行时间的56％,并获得1.5倍的加速比,能弥补计算与存储访问间的速度差异,提高系统计算效率.
展开▼
