AutoPas: Dynamic Algorithm Selection in Molecular Dynamics for Optimal Time and Energy

The large computational cost of force calculations within Molecular Dynamics has led to the development of specialist algorithms such as Linked Cells or Verlet Lists, each with multiple parallelisation schemes. There is, however, no single best algorithm for all scenarios and architectures, and the best algorithm can vary across the domain of a simulation and can change over time. In this work, we will present AutoPas: a black-box particle simulation library that aims to select the fastest or most energy efficient algorithm for a given simulation, including using different algorithmic choices for different regions of the domain and dynamically updating the algorithmic choices as the simulation changes [F. Gratl et al, 2022, N ways to simulate short-range particle systems: Automated algorithm selection with the node-level library AutoPas]. In particular, we will discuss the impact of simulation specific properties, such as density and homogeneity, on the performance of the algorithms and how machine learning and expert knowledge can make use of such properties to ensure the optimal algorithm selections with minimal overhead.     «

The large computational cost of force calculations within Molecular Dynamics has led to the development of specialist algorithms such as Linked Cells or Verlet Lists, each with multiple parallelisation schemes. There is, however, no single best algorithm for all scenarios and architectures, and the best algorithm can vary across the domain of a simulation and can change over time. In this work, we will present AutoPas: a black-box particle simulation library that aims to select the fastest or...     »