Accelerating Molecular Dynamics Simulations on Heterogeneous Architecture

Abstract

Molecular dynamics (MD) is an important computational tool used to simulate chemical and physical processes at the molecular level. MD simulations focus on the motion of the interaction of numerous molecules or atoms. Most scholars focus on accelerating MD on multicore central processing units (CPUs) or other coprocessors, such as graphics processing unit (GPU) or many integrated cores [1]. However, most researchers disregard CPU resources and merely perceive a CPU as a controller when using coprocessors. Thus, hybrid computing cannot be achieved, thereby resulting in the waste of CPU computing resources. In this study, we propose three strategies to accelerate MD simulation. The first strategy uses Compute Unified Device Architecture [2] to rewrite the MD code and to run applications on a single-core CPU-GPU platform. This strategy can achieve satisfactory performance but does not make use of CPU resources to compute for most research activities. In the second strategy, the CPU is set to compute the pair force of a small part of molecules along with the GPU after accomplishing the task of starting the GPU computation. The third strategy is applicable under the condition that the GPU is shared by numerous MPI processes, each of which uses the GPU separately. In this situation, the performance can be improved.