国内外不同GPU用于洪水模拟时加速性能对比分析

Comparison of acceleration performance of different graphic processing units used in flood simulation at home and abroad

由于GPU硬件技术快速发展,同时具有性价比高,可方便实现高密度并行的优点,其已成为当下实现二维洪水高性能加速模拟的主要手段。鉴于国外先进GPU芯片进口遇阻的现状,及时开展基于国产GPU的二维洪水动力学模型的计算性能研究工作是非常必要的。对已有的与国外NVIDIA公司GPU硬件适配的二维洪水动力学模型代码进行改造,使其能够适配国产的GPU,同时选取了NVIDIA公司的两款GPU(P100和V100)和一款国产GPU(Z100)进行了性能测试。选取三个典型洪水案例进行三款GPU性能的对比分析,结果表明:(1)国产GPU(Z100)可实现二维洪水动力学模型的高性能加速计算,其计算性能与NVIDIA公司的P100性能整体相当,但是计算性能弱于V100;(2)随着计算网格数的增大,国产Z100相较于P100表现出更优越的加速性能;(3)随着GPU数量增加,单节点条件下国产GPU的加速比略低于NVIDIA的两款产品。研究成果可为洪水动力学模型的国产化环境移植提供参考。

Due to the rapid technology development of Graphic Processing Unit(GPU)hardware,GPU has become the main means to achieve high-performance accelerated simulation of two-dimensional floods,which has characteristics of high cost-effectiveness,and convenient implementation of high-density parallelism.Given the obstacles in the import of advanced GPU chips from abroad at present,it is necessary to conduct research immediately on the computational performance of two-dimensional flood dynamics models based on domestic GPUs.We have modified the existing 2D flood dynamics model code that is compatible with the GPUs hardware produced by NVIDIA to enable them to adapt to domestic GPUs.At the same time,two GPUs(P100 and V100)produced by NVIDIA company and one domestic GPU(Z100)were selected for performance testing.To compare the performance of the above-mentioned three GPUs,three typical flood cases were selected for test.The results show that:i)High performance on accelerated calculation of two-dimensional flood dynamics models can be achieved based on domestic GPUs(Z100),with basically the same computational performance as NVIDIA’s P100,but weaker than V100;ii)As the number of computing grids increases,the acceleration performance of the domestically produced Z100 is better than that of the P100;and iii)As the number of GPUs increases,the acceleration ratio of domestic GPUs under single node conditions is slightly lower than that of NVIDIA’s two products.The findings can provide references for the transplantation of domestic software and hardware environments for flood dynamics models.