Parallel computing on high-performance accelerators (GPUs)
Processor architectures have seen a paradigm shift towards on-chip parallelisation, which is perfected in many-core accelerators (e.g., GPUs). Such processors are the key to energy-efficient computing, yet their programming is challenged by the hierarchical design and by the diversity of hardware-specific compilers and libraries. Therefore, we have developed the software HAL's MD package [1, 2], which is an efficient, well-tested, and versatile tool for high-precision simulations of complex particle models for research in statistical physics, materials science, and agent-based modelling. It delivers a seemless workflow to compute advanced observables from particle models. It uses a small I/O footprint and sensible data management [3], thereby also serving as a reference implementation of algorithms and fostering good practices in education. The code base consists of about 700 files, it is free and open source and supports open standards. Current development goals are tutorial material for showcase examples, optimised geometric integrators, simulation of reactive flows, interoperability with material models using machine-learned interaction potentials, and vendor-independent support of accelerator hardware [4].
[1] Highly Accelerated Large-scale Molecular Dynamics package (2007–2025), https://halmd.org.
[2] P. H. Colberg and F. Höfling, Comput. Phys. Commun. 182, 1120 (2011).
[3] P. de Buyl, P. H. Colberg, and F. Höfling, Comput. Phys. Commun. 185, 1546 (2014).
[4] V. Skoblin, F. Höfling, and S. Christgau, in 29th PARS Workshop (2023), vol. 36 of Mitteilungen - Gesellschaft für Informatik e.V., Parallel-Algorithmen und Rechnerstrukturen.