DL POLY_3 I/O: Analysis, alternatives, and future strategies

Abstract

The molecular dynamics (MD) method is the only tool to provide detailed information on the time evolution of a molecular system on an atomistic scale. Although novel numerical algorithms and data reorganization approches can speed up the numerical calculations, the actual science of a simulation is contained in the captured frames of the system’s state and simulation data during the evolution. Therefore, an important bottleneck in the scalability and efficiency of any MD software is the I/O speed and reliabilty as data has to be dumped and stored for postmortem analysis. This becomes increasingly more important when simulations scale to many thousands of processors and system sizes increase to many millions of particles. This study outlines the problems associated with I/O when performing large classic MD runs and shows that it is necessary to use parallel I/O methods when studying large systems.