数据与计算发展前沿 ›› 2026, Vol. 8 ›› Issue (4): 123-135.

CSTR: 32002.14.jfdc.CN10-1649/TP.2026.04.009

doi: 10.11871/jfdc.issn.2096-742X.2026.04.009

• • 上一篇    下一篇

分子动力学模拟软件FIREBALL的异构加速优化

靳润锋1,2(),梁文昊1,2,马英晋1,任鹏举3,JAMES P Lewis4,*(),金钟1,*()   

  1. 1 中国科学院计算机网络信息中心， 北京 100083
    2 中国科学院大学， 北京 100049
    3 中科合成油技术股份有限公司， 北京 101407
    4 华南师范大学， 广东 广州 510006
  • 收稿日期:2025-11-19 出版日期:2026-08-20 发布日期:2026-08-21
  • 通讯作者: JAMES P Lewis（E-mail: james.p.lewis.phd@gmail.com）；金钟（E-mail: zjin@sccas.cn）
  • 作者简介:靳润锋，中国科学院计算机网络信息中心，中国科学院大学，博士研究生，主要研究方向为计算化学中的高性能计算。
    本文承担的工作为：算法设计、代码实现、性能测试。
    JIN Runfeng is a Ph.D. candidate at the Computer Network Information Center, Chinese Academy of Sciences, and the University of Chinese Academy of Sciences. His main research focuses on high-performance computing in computational chemistry.
    In this paper, he is responsible for algorithm design, code implementation, and performance evaluation.
    E-mail:jinrunfeng@cnic.cn|JAMES P Lewis，华南师范大学，教授，主要研究方向是量子计算与机器学习在材料与化学中的计算模拟与算法研究。
    本文承担的工作为：指导算法设计、程序开发和性能测试。
    JAMES P Lewis is a professor at South China Normal University. His main research focuses on computational simulations and algorithm development applying quantum computing and machine learning to materials science and chemistry.
    In this paper, he is responsible for supervising the algorithm design, code development, and performance testing.
    E-mail: james.p.lewis.phd@gmail.com|金钟，中国科学院计算机网络信息中心，研究员，主要研究方向为高性能计算软件与数据应用服务平台构建技术、计算化学与计算生物学算法与应用软件。
    本文承担的工作为：指导研究方向和算法设计。
    JIN Zhong is a researcher at the Computer Network Information Center, Chinese Academy of Sciences. His main research interests include the construction of high-performance computing software and data application service platforms, as well as algorithms and application software for computational chemistry and computational biology.
    In this paper, he is responsible for guiding the research direction and algorithm design.
    E-mail: zjin@sccas.cn
  • 基金资助:
    国家重点研发计划(2023YFB3001900);中国科学院先导项目(XDB0500101);国家自然科学基金(22173114);国家自然科学基金(2233300);青促会专项基金(2022168);赣江创新研究院项目(E455F001)

Heterogeneous Acceleration and Optimization of the Molecular-Dynamics Simulation Software FIREBALL

JIN Runfeng1,2(),LIANG Wenhao1,2,MA Yingjin1,REN Pengju3,JAMES P Lewis4,*(),JIN Zhong1,*()   

  1. 1 Computer Network Information Center, Chinese Academy of Sciences, Beijing 100083, China
    2 University of Chinese Academy of Sciences, Beijing 100049, China
    3 Synfuels China Technology Co., Ltd., Beijing 101407, China
    4 South China Normal University, Guangzhou, Guangdong 510006, China
  • Received:2025-11-19 Online:2026-08-20 Published:2026-08-21

RichHTML

6

PDF

41

摘要：

【目的】 为解决FIREBALL在国产异构超算上做大规模AIMD模拟时计算开销高、扩展性不足的问题。【方法】 本文提出的多层次异构并行方案为：通过Kokkos+FLCL跨语言框架与内核拆分策略实现Ewald求和的GPU/DCU加速；采用“主从解耦”MPI架构并行化矩阵对角化模块；并以OpenMP优化复杂模块来平衡性能与成本。【结果】 性能测试表明，核心的Ewald模块、哈密顿矩阵构建和对角化模块分别获得了高达130倍、7倍和6倍的性能提升，在混合并行配置下，软件整体性能提升了4倍。【局限】 目前优化与验证主要针对有限体系规模和特定异构平台，GPU/DCU加速尚集中于Ewald模块。【结论】 该多层次并行优化方案成功解决了FIREBALL的多个性能瓶颈，为探究材料缺陷演化、界面反应等关键物理过程提供了强有力的工具。

关键词: FIREBALL, 异构并行计算, GPU加速, 高性能计算, 第一性原理

Abstract:

[Objective] It is of great significance to address the high computational cost and limited scalability of FIREBALL for large-scale ab initio molecular dynamics (AIMD) simulations on domestic heterogeneous supercomputers. [Methods] This paper proposes a multi-level heterogeneous parallelization scheme: GPU/DCU acceleration of the Ewald summation via a Kokkos+ FLCL cross-language framework combined with a kernel-splitting strategy; parallelization of the matrix diagonalization module using a decoupled master-worker MPI architecture; and OpenMP optimization of complex modules to balance performance and cost. [Results] Performance evaluation shows speedups of up to 130×, 7×, and 6× for the core Ewald module, Hamiltonian construction, and diagonalization modules, respectively. Under a hybrid parallel configuration, the overall software performance is improved by a factor of 4. [Limitations] At present, the optimization and validation are mainly conducted for limited system sizes and specific heterogeneous platforms, and the GPU/DCU acceleration is still concentrated on the Ewald module. [Conclusions] This multi-level parallel optimization scheme successfully alleviates several performance bottlenecks of FIREBALL and provides a powerful tool for investigating key physical processes such as defect evolution and interfacial reactions in materials.

Key words: FIREBALL, heterogeneous parallel computing, GPU acceleration, high performance computing, first principles