摘要
网格重排序是提升流体力学CPU和GPU并行计算效率的重要手段之一。对于非结构网格,由于其数据存储无规律,数据的间接访问会导致访存延迟,尤其是在GPU并行计算时,数据的间接访问将引起内存的非对齐访问,放大了访存延迟的影响。对此,采用Reverse Cuthill-Mckee网格重排序方法优化了非结构网格的数据局部性,并设计了一种面向编号重排序方法。算例测试表明,网格重排序不影响最终计算结果。对比分析了网格重排序对非结构求解器在CPU和GPU上的性能影响:对CPU计算,可以使部分热点函数运行时间降低约20%,整体运行时间降低15%~20%;对GPU计算,大部分热点函数运行时间可降低35%~60%,程序整体运行时间降低约40%。
Mesh renumbering or reordering is one of the important means to improve the CPU and GPU parallel computing efficiency of Computational Fluid Dynamics(CFD).For unstructured meshes,due to the irregular data storage,indirect data access will lead to large memory access delays.Especially in GPU parallel computing,indirect data access will cause non-aligned memory access,amplifying the impact of memory access latency.In this regard,the Reverse Cuthill-Mckee mesh reordering method is used to optimize the data locality of unstructured meshes,and a face renumbering method is designed.The example test shows that the mesh reordering does not affect the final calculation result.The impact of mesh reordering on the performance of unstructured solvers on CPU and GPU is compared and analyzed.For CPU computing,the running time of some hotspot functions can be reduced by about 20%,and the overall running time can be reduced by 15%~20%.For GPU computing,the running time of most hotspot functions is reduced by 35%~60%,and the overall running time of the program is reduced by about 40%.
作者
张勇
张曦
万云博
何先耀
赵钟
卢宇彤
ZHANG Yong;ZHANG Xi;WAN Yun-bo;HE Xian-yao;ZHAO Zhong;LU Yu-tong(Computational Aerodynamics Institute,China Aerodynamics Research and Development Center,Mianyang 621000;School of Computer Science and Engineering,Sun Yat-sen University,Guangzhou 510006,China)
出处
《计算机工程与科学》
CSCD
北大核心
2022年第10期1721-1729,共9页
Computer Engineering & Science
基金
国家数值风洞工程项目(NNW2019ZT6-B18)
广东省引进创新创业团队项目(2016ZT06D211)。
关键词
非结构网格
网格重排序
GPU并行计算
计算流体力学
风雷软件
unstructured mesh
mesh renumbering
GPU parallel computing
computational fluid dynamics(CFD)
PHengLEI
