数据与计算发展前沿 ›› 2024, Vol. 6 ›› Issue (5): 148-158.

CSTR: 32002.14.jfdc.CN10-1649/TP.2024.05.014

doi: 10.11871/jfdc.issn.2096-742X.2024.05.014

• • 上一篇    下一篇

物质点法模拟的大规模并行算法

田少博1,2(),李佳霖1,2,张鉴1,*()   

  1. 1.中国科学院计算机网格信息中心，北京 100083
    2.中国科学院大学，北京 100049
  • 收稿日期:2023-09-07 出版日期:2024-10-20 发布日期:2024-10-21
  • 通讯作者: * 张鉴（E-mail: zhangjian@sccas.cn）
  • 作者简介:田少博， 中国科学院计算机网络信息中心，中国科学院大学，在读博士研究生，主要研究方向为高性能计算。
    本文中主要承担工作为物质点法的并行设计与实现，应用测试。
    Tian Shaobo is a PhD student at the Computer Network Information Center of the Chinese Academy of Sciences. He is also with the University of Chinese Academy of Sciences. His main research direction is High Performance Computing.
    In this paper, he is mainly responsible for the parallel design and implementation of the material point method and application testing.
    E-mail: tianshaobo@cnic.cn|张鉴， 中国科学院计算机网络信息中心，研究员，博士生导师，主要研究方向为高性能计算、偏微分方程数值算法和并行算法研究。
    本文承担工作为：物质点法并行的整体框架设计，研究指导。
    Zhang Jian is a research fellow and PhD supervisor at the Computer Network Information Center of the Chinese Academy of Sciences. His main research fields are High Performance Computing, PDE numerical algorithm， and parallel algorithm.
    In this paper, he is mainly responsible for the overall framework design of the parallel algorithm for the material point method and providing research guidance.
    E-mail: zhangjian@sccas.cn
  • 基金资助:
    国家重点研发计划(2021YFB0300203)

Large Scale Parallel Algorithm for Material Point Method Simulation

TIAN Shaobo1,2(),LI jialin1,2,ZHANG Jian1,*()   

  1. 1. Computer Network Information Center, Chinese Academy of Sciences, Beijing 100083, China
    2. University of Chinese Academy of Sciences, Beijing 100049, China
  • Received:2023-09-07 Online:2024-10-20 Published:2024-10-21

RichHTML

15

PDF

172

摘要：

【目的】物质点法是一种无网格法，常被用于求解碰撞、侵彻和大变形问题。一方面，为了获得更真实的模拟效果，实际应用场景涉及数亿的物质点和网格，这对存储资源和计算能力提出要求。另一方面，物质点和网格之间频繁的插值，需要综合考虑两者实现任务的划分，而物质点的分布相对于背景网格是不均匀的，设计灵活的划分方式实现任务负载均衡成为问题的关键。基于此，本文设计并实现了物质点法的大规模并行模拟算法。【方法】为了使得进程间的任务量相对均衡，对物质点实现自适应划分设计，然后对网格点上的数据依赖和进程间移动的物质点进行通信设计，最后实现了物质点和网格点耦合的并行。【结果】针对物质点法求解侵彻问题，强扩展性获得80%以上的并行效率。【局限】由于物质点不断在空间移动，对物质点进行动态负载均衡设计可能会获得更好地加速效果。【结论】本文实现了物质点法的三维自适应划分并行设计，获得了良好加速效果，相关的数据依赖分析为之后的动态负载平衡的设计和优化提供参考。

关键词: 物质点法, 负载均衡, 三维并行, 大规模

Abstract:

[Objective] As a meshless method, the material point method (MPM) is commonly used to solve collision, penetration, and large deformation problems. On the one hand, in order to accomplish more realistic simulation effects, actual application scenarios involve hundreds of millions of material points and grids. On the other hand, frequent interpolation occurs between the material points and grid nodes. Therefore, a comprehensive consideration of both is necessary to achieve task division. Moreover, since material points are inhomogeneous with relation to the background grid, a flexible division method needs to be designed to achieve workload balancing. Based on it, we design and implement the parallel algorithm to achieve large-scale simulation. [Methods] An adaptive partitioning design is used for MPM to achieve a relatively balanced workload between processes. Then, the communication design is carried out for data dependencies on grid points and material points moving between processes. Finally, the parallel coupling of material points and grid points is implemented. [Results] For solving the penetration problem, its parallel efficiency is more than 80% in the strong scalability testing. [Limitations] Due to the continuous movement of material points in space, dynamic load balancing of material points may get better acceleration effects. [Conclusions] We design a parallel algorithm of 3D adaptive partitioning for MPM, which achieves good acceleration effects. The data dependency analysis provides a reference for the design and optimization of dynamic load-balancing strategies in the future.

Key words: material point method, load balancing, 3D parallel, large scale