Communication Optimization Strategy for Molecular Dynamics Simulation on Sunway TaihuLight

Abstract

Molecular dynamics simulation defined as simple state updates over multiple time steps is the main approach to the description of the chemical, mechanical and electrical processes in many real-world application. Currently, most optimization methods focus on simplifying forces or computing parallelization. However, modern general-purpose supercomputers have succeeded on compute-intensive or memory/bandwidth-intensive applications, but the improvement of network bandwidth/latency is fairly limited. The communication overhead is critical, and severely degrades the overall performance and scalability. In this paper, we propose a communication optimization strategy to reduce inter-nodes communication overheads, including a ghost communication mode to reduce the total amount of message, a shift communication algorithm to reduce the total number of messages and a zero-copy RDMA (Remote Direct Memory Access) communication method to reduce inter-nodes memory copy overheads. We implement our ghost communication, shift communication and zero-copy RDMA communication on the third highest performance supercomputer Sunway TaihuLight in the world (before June 2020). We test molecular dynamics simulation of condensed covalent materials and scale the simulation up to 8,519,680 cores to simulate more than 50.4 million silicon atoms, where the parallel efficiency is over 80% on the whole machine. Results show that the communication optimization strategy reduces the communication time by nearly 75% compared with traditional inter-nodes MPI (Message Passing Interface) communication with memory copy and the dimension of the simulated system significantly exceeds the experimentally measurable range. Our simulation enables virtual experiments on real-world applications and makes the technology more accessible to the general scientific users by using general-purpose supercomputers.