Exascale-ready molecular dynamics simulations with efficient algorithms for extreme core counts

RWTH Publications (RWTH Aachen)

Abstract

Biomolecular simulations, realized by molecular dynamics (MD) and enhanced-sampling approaches, are very powerful tools for studying the structural dynamics, kinetics, and energetics of biological systems. In combination with high-performance computing (HPC), increasingly larger systems and longer timescales can be simulated. However, the sequential nature of MD’s time evolution imposes a hard parallel limit, resulting in reduced scalability and hence under-utilization of HPC systems. As a result, standard MD does not reach the typical timescale (millisecond and beyond) required to study many biological processes. Enhanced-sampling techniques, such as umbrella sampling, metadynamics, and replica-exchange MD, do simulate these long processes but often require various techniques to retrieve kinetic and thermodynamic properties and do not scale well. The arrival of exascale computers has made the need for highly scalable algorithms for MD simulations even more urgent. This doctoral thesis reports on my efforts to address these important issues via algorithmic optimization, design and development. The first was an attempt to speed up MD simulations by means of an alternative parallel 3D discrete Fourier transform (3D DFT) algorithm, which was implemented and benchmarked on the JUWELS Cluster, showing comparable scaling performance to the state-of-the-art. In the second, the software apparatus required for a massively parallel MD strategy was constructed within the highly popular GROMACS code and the PLUMED library. The implementation is capable of both multi-CPU and multi-GPU parallelism and was optimized and benchmarked on the JUWELS Booster. The results revealed its multi-modal scalability in that simulations using it can be both efficiently sped up and also greatly extended for a small increase in runtime. The implementation was shown to scale up to 94% of the JUWELS Booster (3,500 GPUs and 42,000 CPUs) with excellent parallel efficiency. Moreover, a plateauing of parallel efficiency was observed at 50% of the JUWELS Booster, which hints at its ability to scale to even higher node counts. This software has the potential to accelerate MD and thereby enable the study of more complex biological processes than was previously practicable.

Authors 1

1. Nitin Malapally corresponding Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen