Implementation and Analysis of Parallelization Algorithms for Molecular Dynamics Simulations
Sabrina Krallmann · mediaTUM – the media and publications repository of the Technical University Munich (Technical University Munich) · 2020
Within the field of molecular dynamics, force and distance calculations are computationally heavy and the main cause of a long run-time. To decrease this run-time, there are several neighborhood-based particle search algorithms and parallelization strategies. The library AutoPas addresses this challenge and uses auto-tuning to dynamically choose the best strategy during run-time. The parallelization strategies are implemented in so-called traversals. Those are based on containers which originated from neighborhood-based particle search algorithms. To choose the best option, all traversals have to be tested by linearly scanning all options. In order to optimize this search, one would need to know which parameters are influencing the outcome of searching the currently optimal traversal. The present thesis analyzes the traversals regarding their behavior depending on different factors. Those factors are a high number of particles, a large domain size, different densities, as well as in-homogeneous and homogeneous scenarios. The scenarios were tested on a Coffee Lake platform and a Haswell platform. Both platforms were also compared. For a high density, an increasing number of particles leads to the traversals sorting by container, which leads to the assumption that for those scenarios the particle search algorithm is more important than the parallelization strategy. Overall, a high number of particles favored a Verlet Lists approach, while scenarios with a large domain size performed best for Verlet Cluster Lists based traversals. In general, an increasing standard deviation of the homogeneity showed an increasing run-time. This is especially disadvantageous for traversals whose parallelization is sliced-based.