A New Dynamical Domain Decomposition Method for Parallel Molecular Dynamics Simulation on Grid

TL;DR

This paper introduces MPD^3, a dynamic domain decomposition method using Voronoi tessellation for efficient parallel molecular dynamics simulations on heterogeneous computing networks.

Contribution

It presents a novel Lagrangian particle-based decomposition algorithm that adapts dynamically to heterogeneous computing environments, improving load balancing in large-scale MD simulations.

Findings

MPD^3 achieves high adaptability in domain decomposition.

Dynamical Voronoi tessellation results in well-balanced load distribution.

The method is effective across various physical problems and computing environments.

Abstract

We develop a new Lagrangian material particle -- dynamical domain decomposition method (MPD^3) for large scale parallel molecular dynamics (MD) simulation of nonstationary heterogeneous systems on a heterogeneous computing net. MPD^3 is based on Voronoi decomposition of simulated matter. The map of Voronoi polygons is known as the Dirichlet tessellation and used for grid generation in computational fluid dynamics. From the hydrodynamics point of view the moving Voronoi polygon looks as a material particle (MP). MPs can exchange particles and information. To balance heterogeneous computing conditions the MP centers should be dependent on timing data. We propose a simple and efficient iterative algorithm which based on definition of the timing-dependent balancing displacement of MP center for next simulation step. The MPD^3 program was tested in various computing environments and physical problems. We have demonstrated that MPD^3 is a high-adaptive decomposition algorithm for MD simulation. It was shown that the well-balanced decomposition can result from dynamical Voronoi polygon tessellation. One would expect the similar approach can be successfully applied for other particle methods like Monte Carlo, particle-in-cell, and smooth-particle-hydrodynamics.