数万核级与异构架构有限元程序可扩展性测试与分析.pdfVIP

数万核级与异构架构有限元程序可扩展性测试与分析 † 任江勇 王朝尉 王迎瑞 田荣 ( 中国科学院计算技术研究所 高性能计算机研究中心，北京 100190) Scalability Tests of a Finite Element Code on Tens of Thousands Cores and Heterogeneous Architecture* REN Jiang-yong, WANG Chao-wei, WANG Ying-rui, TIAN Rong† (High Performance Computer Research Center, Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100190, China) †Corresponding author: Phn: +86-010 E-mail: rongtian@ Abstract : A multi-scale finite element method, msFEM, is tested . The flat MPI version is scaled from 20K up to 200K CPU cores on Jaguar, delivering parallel efficiency of 71% at 200K cores with a finite element model of eight millions of degrees of freedom. GPU versions in both double precision and mixed precision , which are coded through MPI+OpenMP+CUDA hybrid programming , deliver linear speedup up to 900 (on Jaguar) and 1500 (on + Nebulae) GPU nodes, achieving remarkable 90 % parallel efficiency. The mixed-precision GPU version delivers further 1.5 times of speedup over the fully double precision version with no significantly extra coding cost. The large-scale tests show that msFEM bears very good scalability and is highly potential for domain applications at extreme-scale. Key words:Tens of Thousands Cores, Scalability, GPU, Mixed Precision Algorithm, Finite Element Method 摘 要: 本文对多尺度有限元程序 msFEM 的 CPU 、GPU 双精度以及 GPU 混合精度三个版本进行了测试。 msFEM_CPU 在美国橡树岭国家实验室的Jaguar 上实现了从2 万到20 万核的可扩展性测试，在80 亿自由度的测 试规模下，20 万核的并行效率为7 1%。msFEM_GPU 采用MPI+OpenMP+CUDA 的混合并行模式实现了GPU 全 双精度和混合精度两个版本。GPU 版在 Jaguar 和华南超算中心（深圳）的曙光6000/Nebulae 分区上分别实现了 900 和1500 个GPU 节点的全系统可扩展性测试，实现近线性加速，GPU 节点规模最大时并行效率为90%以上。 混合精度算版本相对全双精度算法有 1.5 倍加速。测试表明：msFEM 具有良好的可扩展性，为大规模问题求解 奠定了基础。 关键词: 万核级，GPU ，可扩展性