多GPU加速在高性能数值计算中的应用

摘要

Aiming at the numerical calculation of physical problems in the field of nuclear energy,based on the multiple GPU cal-culation elements of multi-core hardware platforms,the programming architecture of multiple GPU development platform on task level parallelization was proposed,which provided basic parallel programming models for the nuclear engineering design proce-dure.The difficulties for the designer and the technical staff were mitigated when they used the low-level interface with CUDA, the programming framework could promote the acceleration performance of computing tasks under multiple GPUs by means of rewriting the time-consuming calculation module and transplanting them to the GPUs.Experimental results showed that the pro-gramming architecture could effectively promote the acceleration performance of computing tasks under multiple GPUs.%针对核能领域中存在的大量数值计算问题，结合实际应用中多核硬件平台里面的多GPU （图形处理器）计算节点，提出基于CUDA （统一计算设备架构）的任务级的并行编程框架。为核电工程设计中的数值计算提供并行硬件平台下的基础GPU编程模型，将CUDA底层对多GPU的有效调度与上层使用进行分离，隔离底层的编程技术，减轻设计人员和开发人员对CUDA底层接口的使用难度；将主程序中的耗时计算模块用CUDA进行改写，再移植到GPU上执行，提升多GPU对计算任务的加速性能。实验结果表明，该编程框架能有效提升多GPU对计算任务的加速。