应用多GPU的可压缩湍流并行计算

发布时间：2018-01-31 19:54

本文关键词： CUDA 图形处理器湍流并行计算计算流体力学　出处：《国防科技大学学报》2015年03期 　论文类型：期刊论文

【摘要】：利用CUDA Fortran语言发展了基于图形处理器(GPU)的计算流体力学可压缩湍流求解器。该求解器基于结构网格有限体积法,空间离散采用AUSMPW+格式,湍流模型为k-ωSST两方程模型,采用MPI实现并行计算。针对最新的GPU架构,讨论了通量计算的优化方法及GPU计算与PCIe数据传输、MPI通信重叠的多GPU并行算法。进行了超声速进气道及空天飞机等算例的数值模拟以验证GPU在大网格量情况下的加速性能。计算结果表明:相对于Intel Xeon E5-2670 CPU单一核心的计算时间,单块NVIDIA GTX Titan Black GPU可获得107~125倍的加速比。利用四块GPU实现了复杂外形1.34亿网格的快速计算,并行效率为91.6%。
[Abstract]:A computational fluid dynamics compressible turbulence solver based on graphics processor (GPU) is developed by using CUDA Fortran language, which is based on the finite volume method of structured meshes. Spatial discretization is based on AUSMPW scheme, turbulence model is k- 蠅 SST two-equation model, and MPI is used to realize parallel computation. Aiming at the latest GPU framework. The optimization method of flux calculation, GPU calculation and PCIe data transmission are discussed. Multiple GPU parallel algorithm for overlapping MPI communication. Numerical simulations of supersonic inlet and aircrafts are carried out to verify the acceleration performance of GPU in the case of large mesh quantities. The results show that:. The computational time relative to the single core of Intel Xeon E5-2670 CPU. Single block NVIDIA GTX Titan Black. GPU can get a speedup ratio of 107 ~ 125.The fast calculation of 134 million mesh with complex shape is realized by using four GPU blocks. The parallel efficiency is 91.6.
【作者单位】：国防科技大学航天科学与工程学院;
【基金】：国家自然科学基金资助项目(91016010,91216117)
【分类号】：TP338.6
【正文快照】： 随着硬件性能的提高及编程技术的改进,图形处理器(Graphical Processing Unit,GPU)加速器在高性能计算领域逐渐得到广泛的应用。在最新公布的超级计算机Top500名单中共有62套系统采用了加速器/协处理器,其中采用GPU加速器有46套,而在最新的Green500名单中前10位的超级计算机均

【共引文献】