Large-Scale Molecular Dynamics Simulations on Modular Supercomputer Architecture with Gromacs

摘要

Modular Supercomputer Architecture is an innovative idea to bring advantages of different HPC architectures currently available together and make it possible for the users to benefit from it by running different parts of their applications on the most appropriate hardware to reach the best performance with high energy efficiency. DEEP-EST project aims not just at building MSA prototype, but rather co-design it by a collaboration of hardware and application software experts. The HPC applications encompass scientific fields like neuroscience, molecular dynamics, radio astronomy, space weather, data analytics in earth science and high energy physics. Here, we present our research on running large-scale molecular dynamics simulations with one the worlds' fastest MD software - Gromacs on two modules (Booster - KNL based HPC system and Cluster - Haswell based HPC system) of JURECA supercomputing system at Juelich Supercomputer Centre. We tested the performance scalability of the simulation of 20 million atoms biomolecular system running Gromacs in Cluster-Booster configuration. The long-range electrostatics calculations were conducted on the JURECA Cluster nodes, while all of the rest calculations - on JURECA Booster. Through source code profiling and Gromacs internal performance counters analysis, we investigated the influence of the performance scalability the different application's parts on the overall application performance scalability. We optimized the job parameters like Cluster to Booster number of cores ration and the corresponding number of module-interconnecting gateways. Our results prove the applicability of the MSA concept in the field of Molecular Dynamics.