Towards the Smarter Tuning of Molecular Dynamics Simulations

The large computational cost of pairwise force calculations within Molecular Dynamics requires the use of specialist algorithms, such as Linked Cells or Verlet Lists, as well as efficient ways of parallelising such algorithms. There is, however, no ``silver bullet'' best algorithm for all simulations, and the best algorithm can change over the course of a simulation. AutoPas is a node-level particle simulation library that aims to dynamically select the most optimal algorithm, vectorisation strategy, and shared memory parallelism for a given metric, such as time for force calculation [F. Gratl et al, N ways to simulate short-range particle systems: Automated algorithm selection with the node-level library AutoPas, 2022]. In multi-node HPC systems, each node has their own AutoPas container, making it's own tuning decisions. Practically, this autotuning requires trialling algorithms during the course of the simulation, however trialling slow algorithms can provide significant overhead, and so smart tuning strategies must be developed that can select optimal, or close to optimal, performance, with minimal overhead. In this poster, we will discuss how statistical techniques such as Bayesian Optimisation, Gaussian Process Models, and Reinforcement Learning can be adapted into smart tuning strategies within AutoPas. To support our claims, we present results using our smart tuning strategies applied to the field of Molecular Dynamics.     «

The large computational cost of pairwise force calculations within Molecular Dynamics requires the use of specialist algorithms, such as Linked Cells or Verlet Lists, as well as efficient ways of parallelising such algorithms. There is, however, no ``silver bullet'' best algorithm for all simulations, and the best algorithm can change over the course of a simulation. AutoPas is a node-level particle simulation library that aims to dynamically select the most optimal algorithm, vectorisation str...     »