Scaling Poisson Solvers on Many Cores via MMEwald
Ming-Chuan Wu, Yangjun Wu, Honghui Shang, Ying Liu, Huimin Cui, Fang Li, Xiaohui Duan, Yunquan Zhang, Xiaobing Feng · IEEE Transactions on Parallel and Distributed Systems · 2021
The Poisson solver for the calculation of the electrostatic potential is an essential primitive in quantum mechanics calculations. In this article, we adopt the Ewald method and propose a highly-optimized and scalable framework for Poisson solver, MMEwald, on the new generation Sunway supercomputer, capable of utilizing the collection of 390-core accelerators it uses. The MMEwald is based on a grid adapted cut-plane approach to partition the points into batches and distribute the batch to the processors. Furthermore, we propose a set of architecture-specific optimizations to efficiently utilize the memory bandwidth and computation capacity of the supercomputer. Experimental results demonstrate the efficiency of the MMEwald in providing strong and weak scaling performance.