Smoothed particle hydrodynamics method for free surface flow based on MPI parallel computing
Smoothed particle hydrodynamics method for free surface flow based on MPI parallel computing
复制标题
基于MPI并行计算的自由表面流光滑粒子流体动力学方法
DOI：
10.3389/fphy.2023.1141972
复制
发表时间：
2023-03
期刊：
影响因子：
--
通讯作者：
Sifan Long;K. Wong;Xiaokang Fan;Xiaowei Guo;Canqun Yang
中科院分区：
文献类型：
--
作者：
Sifan Long;K. Wong;Xiaokang Fan;Xiaowei Guo;Canqun Yang
在计算流体力学（CFD）领域，光滑粒子流体力学（SPH）非常适合模拟大变形、自由表面流动等各类流动场景问题。然而，传统的光滑粒子流体动力学方法存在计算复杂度高的问题，这限制了它们在有精度要求的场景中的应用。DualSPHysics是学术界提出的一个优秀的光滑粒子流体动力学软件。基于该工具，本文提出了一个大规模的并行光滑粒子流体动力学框架：parallelDualSPHysics，它可以解决大规模自由表面流动的模拟。首先，提出了一种有效的区域分解算法。并在并行框架下对DualSPHysics的数据结构进行了改造。其次，对并行粒子交互和粒子更新模块提出了计算和通信重叠的策略，大大提高了光滑粒子流体动力学方法的并行效率。最后，我们还添加了预处理和后处理模块，使parallelDualSPHysics能够在现代高性能计算机上运行。此外，经过充分的评估表明，测试的300万到1.2亿个粒子仍然可以保持90%以上的计算效率，这表明并行策略可以实现上级并行效率。
In the field of computational fluid dynamics (CFD), smoothed particle hydrodynamics (SPH) is very suitable for simulating problems with large deformation, free surface flow and other types of flow scenarios. However, traditional smoothed particle hydrodynamics methods suffer from the problem of high computation complexity, which constrains their application in scenarios with accuracy requirements. DualSPHysics is an excellent smoothed particle hydrodynamics software proposed in academia. Based on this tool, this paper presents a largescale parallel smoothed particle hydrodynamics framework: parallelDualSPHysics, which can solve the simulation of large-scale free surface flow. First, an efficient domain decomposition algorithm is proposed. And the data structure of DualSPHysics in a parallel framework is reshaped. Secondly, we proposed a strategy of overlapping computation and communication to the parallel particle interaction and particle update module, which greatly improves the parallel efficiency of the smoothed particle hydrodynamics method. Finally, we also added the pre-processing and post-processing modules to enable parallelDualSPHysics to run in modern high performance computers. In addition, a thorough evaluation shows that the 3 to 120 million particles tested can still maintain more than 90% computing efficiency, which demonstrates that the parallel strategy can achieve superior parallel efficiency.