The GROMACS simulation engine aims to provide flexibility and performance enabling researchers to make efficient use of compute hardware from workstations to supercomputers. GPU accelerators democratized molecular dynamics and allow extremely efficient trajectory generation. While performance has been steadily improving, hardware trends have meant that harvesting this performance has been increasingly challenging, and has required significant algorithmic and parallelization design changes in the GROMACS simulation engine. A key functionality introduced and refined in recent years is the GPU-resident MD loop scheme. This allows prioritizing keeping the GPU busy, and it has enabled further advances in improving both intra- and multi-node strong scaling. At the same time, the GROMACS heterogeneous-first design has been maintained, harnessing the CPU for flexibility and to support the broad range of the versatile GROMACS feature set (e.g. pulling, AWH). GROMACS has also made efforts in supporting new emerging GPU platforms and doing so using open standards. Join this seminar to learn about these advances that enable the performance, flexibility, and portability of the GROMACS and how to harness them in real-world simulations.
Szilárd Páll is an HPC researcher at the PCD Center for High Performance Computing at KTH Royal Institute of Technology in Stockholm. He has a background in computer science and computational biophysics and has worked with GPU accelerators for scientific computing since 2008. He helped reformulate key parallel algorithms in molecular dynamics for modern processor architectures and co-authored the first heterogeneous CPU-GPU parallelization of GROMACS. His recent focus is on efficient asynchronous task scheduling and strong scaling MD on exascale heterogeneous architectures.