College of Computer Science and Technology, Shandong University of Science and Technology
College of Computer, Qufu Normal University

宇宙学模拟是典型的N体问题,是高性能计算中具有代表性和挑战性的问题之一。本研究在神威·太湖之光平台上对天文N体模拟软件PhotoNs-2中的计算主体———快速多极子方法(fastmultipolemethod,FMM)进行移植和性能优化。针对目前研究中存在的计算效率不高、通信开销大问题,结合神威·太湖之光SW26010处理器架构特点,通过数据重整、超越函数计算重构、设计双缓冲和消息传递接口通信时合并发送树进行优化。相较于优化前,优化后的PhotoNs-2在3个不同算例规模下均取得约24倍的加速效果。提出的优化方案可以为其他高性能应用在神威·太湖之光平台上的移植与优化提供参考。

Cosmological simulation is a typical N-body problem and one of the representative and challenging problems in high-performance computing. This paper ports and optimizes the performance of the fast multipole method (FMM), the main part of the astronomical N-body simulation software PhotoNs-2 on the Sunway TaihuLight platform. Aiming to increase the computational efficiency and reduce the communication expense in the current research, this paper optimizes the fast multipole method by reorganizing data, reconstructing the transcendental function computation and designing double buffering and merging the sending tree during message passing interface (MPI) communication based on the characteristics of the architecture of the SW26010 processor of Sunway TaihuLight. Compared with the algorithm before opitimization, the optimized PhotoNs-2 achieves a speedup of about times under three different arithmetic cases. The proposed optimization scheme can provide a reference for the porting and optimization of other high-performance applications on the Sunway TaihuLight plat- form.

山东省自然科学基金项目(ZR2022MF274;ZR2021LZH004;ZR2023LZH009)