异构系统结构力学计算GPU加速性能分析与应用

Analysis and application of structural mechanical simulation accelerated by GPU with heterogeneous architecture

高性能计算领域利用GPU加速计算已逐渐发展成为主流应用的普遍功能之一。主流结构力学应用ABAQUS支持GPU通用计算,充分发挥GPU的高性能浮点运算能力与访存带宽,提高软件求解效率。介绍了ABAQUS软件对GPU加速应用的发展历程,在上海超级计算中心三种不同的异构环境下,针对实际算例的GPU加速性能进行了测试,建立计算效率分析方法,分析GPU加速对求解时间、系统资源调用、软件并行效率的影响,并针对采用隐式方法求解百万量级自由度问题时资源的合理使用提出建议。

General computation accelerated by Graphic Processing Unit （GPU） has become popular in today＇s High Performance Computing （HPC） area. As one of the leading software solving structural mechanical problems, ABAQUS now offers general purpose GPU （GPGPU） simulation module. Benefited from high floating point computational ability and memory bandwidth, GPU is able to accelerate Computer Aided Engineering （CAE） simulation in ABAQUS. History of GPGPU simulation was firstly introduced, followed by software test regarding to engineering cases with three different heterogeneous hardware systems provided in Shanghai Supereomputer Center. The influence on solver time, system resource usage, and parallel efficiency by GPU acceleration was addressed, while suggestions on solving multi-million degree-of-freedom cases by implicit solver were made.