Performance Analyses of a Parallel Verlet Neighbor List Algorithm for GPU-Optimized MD Simulations

Abstract

Molecular dynamics (MD) simulations provide a molecular-resolution physical description of the folding and assembly processes, but the size and the timescales of simulations are limited because the underlying algorithm is computationally demanding. We recently introduced a parallel neighbor list algorithm that was specifically optimized for MD simulations on GPUs. In our present study, we analyze the performance of the algorithm in our MD simulation software, and we observe that the major of the overall execution time is spent performing the force calculations and the evaluation of the neighbor list and pair lists. The overall speedup of the GPU-optimized MD simulations as compared to the CPU-optimized version is N-dependent and ~30x for the full 70s ribosome (10,219 beads). The pair and neighbor list evaluations have performance speedups of ~25x and ~55x, respectively. We then make direct How biomolecules fold and assemble into well-defined structures that correspond to cellular functions is a fundamental problem in biophysics with direct biomedical application because some functions lead to diseases such as Alzheimer's, Parkinson's, and cancer. Molecular dynamics (MD) simulations provide a molecular-resolution physical description of the folding and assembly processes, but the computational demands of the algorithms restrict the size and the timescales one can simulate. In a recent study, we introduced a parallel neighbor list algorithm that was specifically optimized for MD simulations on GPUs. We now analyze the performance of our MD simulation code that incorporates the algorithm, and we observe that the force calculations and the evaluation of the neighbor list and pair lists constitutes a majority of the overall execution time. The overall speedup of the GPU-optimized MD simulations as compared to the CPU-optimized version is N-dependent and ~30x for the full 70s ribosome (10,219 beads). The pair and neighbor list evaluations have performance speedups of ~25x and ~55x, respectively. We then make direct comparisons with the performance of our MD simulation code with that of the SOP model implemented in the simulation code of HOOMD, a leading general particle dynamics simulation package that is specifically optimized for GPUs.