Benchmarking MD systems simulations on the Graphics Processing Unit and Multi-Core Systems

TL;DR

This paper compares molecular dynamics simulations on GPUs and multi-core systems, demonstrating that multi-core systems can outperform GPUs significantly in simulation speed while maintaining accurate temperature control.

Contribution

It introduces a comparative benchmarking of MD simulations on GPUs and multi-core systems using novel thermostats and demonstrates the superior performance of multi-core parallelization.

Findings

Multi-core systems achieve up to 33 times faster performance than GPUs.

GPU simulations with CUDA reach optimal speed at four processors, 3.67 times faster than a single processor.

Both systems maintain system temperature close to reference values.

Abstract

Molecular dynamics facilitates the simulation of a complex system to be analyzed at molecular and atomic levels. Simulations can last a long period of time, even months. Due to this cause the graphics processing units (GPUs) and multi-core systems are used as solutions to overcome this impediment. The current paper describes a comparison done between these two kinds of systems. The first system used implies the graphics processing unit, respectively CUDA with the OpenMM molecular dynamics package and OpenCL that allows the kernels to run on the GPU. This simulation is done on a new thermostat which mixes the Berendsen thermostat with the Langevin dynamics. The second comprises the molecular dynamics simulation and energy minimization package GROMACS which is based on a parallelization through MPI (Message Passing Interface) on multi-core systems. The second simulation uses another new thermostat algorithm related respectively, dissipative particle dynamics - isotropic type (DPD-ISO). Both thermostats are innovative, based on a new theory developed by us. Results show that parallelization on multi-core systems has a performance up to 33 times greater than the one performed on the graphics processing unit. In both cases temperature of the system was maintained close to the one taken as reference. For the simulation using the CUDA GPU, the faster runtime was obtained when the number of processors was equal to four, the simulation speed being 3.67 times faster compared to the case of only one processor.