Analyzing Molecular Simulations Trajectories by Utilizing CUDA on GPU Architecture

TL;DR

This paper demonstrates how GPU-accelerated computing using CUDA significantly speeds up the analysis of molecular simulation trajectories, achieving 10-80x faster results for static and dynamic properties.

Contribution

It introduces GPU-based kernels for molecular data analysis, providing substantial speedups over CPU codes for static and dynamic property calculations.

Findings

Order metrics show 10-50x speedup over CPU codes.

Three-particle correlation calculations are 10-45x faster.

Viscosity and mean square displacement computations are 10-80x faster.

Abstract

With the advent of high-performance computing techniques, the data for analysis has grown significantly. Here, graphic processing unit (GPU) based program kernels are discussed to exploit parallelism in the analysis codes specific to molecular simulations trajectories and data, hence reducing the time consumption. Commonly computed properties of systems which utilize static and dynamic correlations are considered. In static properties, order metrics based on two-particle correlations are shown to exhibit 10-50x speedups relative to conventional serial CPU codes. Efficiency in finding three-particle correlations, which are relatively more time consuming calculations, are also shown to be 10-45x faster depending on system size. In the case of dynamic properties, the viscosity is computed using Green-Kubo formalism at a 10-25x faster rate depending on the correlation time and total trajectory length. Similarly, for mean square displacement calculations, the speedup of approximately 40-80x is achieved for all system sizes and simulation times. Particularly for these embarrassingly parallel computationally intensive problems, the compute unified device architecture application programming interface affords maximum scale up with minimum implementation time. It is shown that all the GPU codes render 1-2 order of magnitude reduction in the analyses times. Hence these codes, in conjunction with GPU accelerated molecular simulations, can lead to an overall improved and efficient performance.