A Communication-Avoiding Algorithm for Molecular Dynamics Simulation

Abstract

Molecular dynamics and many similar time-dependent computing tasks are defined as simple state updates over multiple time steps. In recent years, modern supercomputing clusters have enjoyed fast-growing compute capability and moderate-growing memory bandwidth, but their improvement of network bandwidth/latency is limited. In this paper, we propose a new communication-avoiding algorithmic model based on asynchronous communications which, unlike BSP, records and handles multiple iterative states together. The basic idea is to let computation run in small regular time steps while communications over longer dynamic time steps. Computation keeps checking inaccuracies so that the intervals between communications are small in volatile scenarios but longer when dynamics is smooth. This helps reduce the number of data exchanges via network communication and hence improve the overall performance when communication is the bottleneck. We test MD simulation of condensed covalent materials on the Sunway TaihuLight. For best time-to-solution, the general-purpose supercomputer Sunway TaihuLight performs 11.8 K steps/s for a system with 2.1 million silicon atoms and 5.1 K steps/s for 50.4 million silicon atoms. This time-to-solution performance is close to those of state-of-art hardware solution. A software solution using general-purpose supercomputers makes the technology more accessible to the general scientific users.