摘要:随着异构系统成为建造超级计算机的重要选择，如何让CPU与加速器协调工作以充分发挥异构系统的计算性能具有重要意义.HPL是高性能计算领域最重要的基准测试程序，传统面向纯CPU系统的HPL算法通过加速器加速矩阵乘法的做法已经无法取得很好的性能.针对这一问题，提出了基于国产处理器-国产加速器异构系统的HPL性能模型和多线程细粒度流水HPL算法.完成了一个轻量级跨平台异构加速框架HPCX，以实现跨平台的HPL算法.该性能模型能够准确地预测类似异构系统的HPL性能.该HPL算法在NVIDIA GPU平台上性能超过了NVIDIA官方闭源nvhpl程序9%.在国产处理器-国产加速器平台512个节点的规模上，优化的HPL算法实现了2.3 PFLOPS实测峰值性能和71.1%的浮点效率.



Abstract:As heterogeneous system becomes one of the most important choices to build super computers, how to orchestrate CPU and accelerator to leverage the great computability of heterogeneous systems is of great significance. HPL is the most important benchmark in HPC field, traditional HPL algorithm targeting at CPU-only systems cannot achieve high performance by only offloading matrix multiplication workload to accelerators. To solve this problem, this work proposes a HPL performance model and a multithread fine-grained pipelining algorithm for domestic-processor-domestic-accelerator heterogeneous system. Meanwhile, a light weight cross-platform heterogeneous framework is implemented to carry out a cross-platform HPL algorithm. The proposed performance model predicts HPL performance accurately on similar heterogeneous systems. On NVIDIA platform, the proposed HPL algorithm outperforms the NVIDIA proprietary counterparts by 9%. On domestic-processor-domestic-accelerator platform, the finally optimized Linpack program achieves 2.3 PFLOPS on 512 nodes, with floating-point efficiency 71.1%.



PDF全文下载地址:

http://jos.org.cn/jos/article/pdf/6004