基于多图形处理单元加速的各向异性弹性波正演模拟

摘要

Comparative analysis study of the computational efficiency of anisotropic elastic wave modeling using CPU serial calculation, graphics processing unit(CPU) 16 threads parallel computing and multiple GPU parallel computing were performed on models of different size. The results show that compare to CPU serial calculation and CPU 16 threads parallel computing,multiple GPU parallel computing achieves a speedup of 30 and 156 times,re-spectively,on a model mesh of 2563. Multiple GPU parallel computing can significantly speed up the modeling, and the speedup becomes more significant when the mesh is larger. Multiple GPU parallel computing is a proper so-lution for anisotropic elastic wave modeling on single node computing environment.%比较分析了在不同网格大小介质模型情况下,分别采用串行计算、CPU 16个线程并行计算和4块图形处理单元(GPU)并行计算进行各向异性弹性波动方程正演模拟的执行时间差异.发现在网格点为2563的大模型上,用4块GPU的并行模拟计算相对16线程并行计算与串行计算的加速分别为30倍与156倍.表明多GPU并行算法可以显著缩短数值模拟时间,而且模型网格越大,加速效果越显著.因此,在单机环境下进行大尺度模型的各向异性弹性波正演模拟,采用多GPU并行计算方式是一个合适的加速选择.