高级检索
    刘智翔, 方勇, 宋安平, 徐磊, 王晓伟, 周丽萍, 张武. 基于MRT-LBM方法的大规模可扩展并行计算研究[J]. 计算机研究与发展, 2016, 53(5): 1156-1165. DOI: 10.7544/issn1000-1239.2016.20148441
    引用本文: 刘智翔, 方勇, 宋安平, 徐磊, 王晓伟, 周丽萍, 张武. 基于MRT-LBM方法的大规模可扩展并行计算研究[J]. 计算机研究与发展, 2016, 53(5): 1156-1165. DOI: 10.7544/issn1000-1239.2016.20148441
    Liu Zhixiang, Fang Yong, Song Anping, Xu Lei, Wang Xiaowei, Zhou Liping, Zhang Wu. Large-Scale Scalable Parallel Computing Based on LBM with Multiple-Relaxation-Time Model[J]. Journal of Computer Research and Development, 2016, 53(5): 1156-1165. DOI: 10.7544/issn1000-1239.2016.20148441
    Citation: Liu Zhixiang, Fang Yong, Song Anping, Xu Lei, Wang Xiaowei, Zhou Liping, Zhang Wu. Large-Scale Scalable Parallel Computing Based on LBM with Multiple-Relaxation-Time Model[J]. Journal of Computer Research and Development, 2016, 53(5): 1156-1165. DOI: 10.7544/issn1000-1239.2016.20148441

    基于MRT-LBM方法的大规模可扩展并行计算研究

    Large-Scale Scalable Parallel Computing Based on LBM with Multiple-Relaxation-Time Model

    • 摘要: 在大规模三维复杂流动的数值模拟中,针对具有良好数值稳定性的多弛豫时间模型格子Boltzmann方法(MRT-LBM),并结合大涡模拟湍流模型和曲面边界插值格式,分析了在D3Q19离散速度模型下的网格生成、流场信息初始化和迭代计算3部分的可并行性.采用MPI编程模型,从分布式集群的特点和计算量负载均衡的角度出发,分别提出了适合于大规模分布式集群的网格生成、流场信息初始化和迭代计算的并行算法.该并行算法也能有效适用于D3Q15和D3Q27离散速度模型.通过在国产神威蓝光超级计算机上的测试,分别针对求解问题总体计算规模固定和保持每个计算核中计算量一致的2种情况的并行性能分析,验证了该并行算法在十万计算核的量级下仍具有良好的加速比和可扩展性.

       

      Abstract: In the large-scale numerical simulation of three-dimensional complex flows, the multiple-relaxation-time model (MRT) of lattice Boltzmann method (LBM) has better property of numerical stability than single-relaxation-time model. Based on the turbulence model of large eddy simulation (LES) and the interpolation scheme of surface boundary, three iteration calculations of grid generation, initialization of flow information and parallelism property are analyzed respectively under the discrete velocity model D3Q19. Distributed architecture and the communication between different compute nodes using message passing interface (MPI) are often used by current high performance computing clusters. By considering both the features of distributed clusters and the load balance of calculation and using MPI programming model, the grid generation, initialization of flow information and the parallel algorithm of iteration calculation suitable for large-scale distributed cluster are studied, respectively. The proposed parallel algorithm also can be suitable for D3Q15 discrete velocity model and D3Q27 discrete velocity model. Two different cases, solving problem with fixed total calculation and solving problem with fixed calculate amount in every computing cores, are considered in the process of numerical simulation. The performances of parallelism are analyzed for these two cases, respectively. Experimental results on Sunway Blue Light supercomputer show that the proposed parallel algorithm still has good speedup and scalability on the order of hundreds of thousands of computing cores.

       

    /

    返回文章
    返回