基于国产加速卡的地震模拟计算性能分析与优化

Performance Analysis and Optimization of Seismic Simulation Computation Based on Domestic Accelerator Cards

摘要:

AWP‐ODC是基于有限差分数值方法来实现大规模三维地震模拟的软件。随着国外对我国高性能计算芯片的出口限制，我国急需发展自己的高性能计算芯片及其软件生态。早期的AWP‐ODC加速主要基于NVIDIA GPU软硬件架构来设计优化，近年来，多种异构计算平台迅猛发展，如何基于新的异构计算软硬件平台来加速 AWP‐ODC具有重要研究价值。为此，本文在一种国产加速卡上对AWP‐ODC进行移植。针对耗时较多的核函数 dstrqc，通过GPU访存优化和网格参数优化等方式缩短了其运行时间。最后分别在国产类GPU单卡和双卡上，利用Fréchet Kernels地震和8·3鲁甸地震数据集进行性能测试。实验结果表明，在单卡计算环境下，两个数据集的 FLOPS分别提高了30.51%和25.21%；在双卡计算环境下，两个数据集的FLOPS分别提高了9.42%和23.6%。

Abstract:

AWP-ODC is a software for large-scale 3D seismic simulation based on the finite difference numerical method. Due to foreign export restrictions on high-performance computing chips to China, there is an urgent need to develop China's own high-performance computing chips and software ecosys tem. The early acceleration of AWP-ODC was primarily designed and optimized based on the NVIDIA GPU software and hardware architecture. In recent years, various heterogeneous computing plat forms developed rapidly. How to accelerate AWP-ODC based on new heterogeneous computing soft ware and hardware platforms showed significant research value. To this end, AWP-ODC was ported to a domestic accelerator card. By optimizing GPU memory access and grid parameters, the execution time of the time-consuming kernel function dstrqc was reduced. Finally, performance tests were con ducted on a domestic GPU single-card and dual-card setup using the Fréchet Kernels seismic dataset and the 8·3 Ludian earthquake dataset. Experimental results showed that, under a single-card comput ing environment, the FLOPS for the two datasets increased by 30.51% and 25.21%, respectively. Under a dual-card computing environment, the FLOPS for the two datasets increased by 9.42% and 23.6%, respectively.