Optimizing Astrophysical Simulation Software on Sunway Heterogeneous Manycore Architecture

Tao Fang,Junshi Chen,Mingfan Li,Ziyu Zhang,Hong An,Wenting Han

DOI: https://doi.org/10.1109/hpcc-smartcity-dss50907.2020.00014

2020-01-01

Abstract:Numerical simulation technology significantly contributes to many field, such as astrophysics, molecular dynamics, vortex methods, electrostatics. PHoToNs, a widely used software implemented based on Particle-Mesh (PM) and Fast Multipole Method (FMM), studies the laws of motion and interaction between galaxy particles in astrophysics. It requires enormous computing power to simulate hundreds of billions or even trillions of particles to meet the requirements of resolution and precision, which simultaneously appeals numerous experts on high performance scientific computing. As one of the top-ranking supercomputers, Sunway TaihuLight has huge demands for performing high performance numerical simulations. This paper presents our contributions for porting the PHoToNs onto Sunway supercomputer and the heterogeneous manycore processor, SW26010. To adapt the customized chip architecture, we propose some effective techniques like data rearrange and reuse, LDM data buffer, and overlap of the communication and computation to address the issues on weak main memory bandwidth. We also enable vectorization on computing CPEs to exploit the finegrained data level parallelism. The proposed solutions for optimizing PHoToNs on Sunway TaihuLight finally achieves up to 55x speedup, as well as a good reference for other N-body problems for similar heterogeneous environment.