SCCS Colloquium - Oct 14, 2020

From Sccswiki
Jump to navigation Jump to search
Date: Oct 14, 2020
Room: Online (BBB)
Time: 15:00 - 16:00

Sabrina Krallmann: Implementation and Analysis of Parallelization Algorithms for Molecular Dynamics Simulations

Bachelor's thesis submission talk. Sabrina is advised by Fabio Gratl.

Within the field of molecular dynamics, force and distance calculations are causing a long run-time. To decrease it, neighborhood-based particle search algorithms and parallelization strategies are used. The library AutoPas applies auto-tuning to dynamically choose the best strategy. Therefore, all options are tested linearly. To optimize this search, one needs to know which parameters are influencing the currently best strategy. This thesis analyzes the strategies regarding their behavior depending on different factors. For a high density an increasing number of particles leads to the parallelizations sorting by particle container type, leading to the assumption that for those scenarios the particle search algorithm is more important than the parallelization strategy. A high number of particles favored a Verlet Lists approach, while scenarios with a large domain performed best for Verlet Cluster Lists strategies. With growing in-homogeneity the run-time increased.

Keywords: Molecular Dynamics, Parallelization, Auto-Tuning, Automatic Algorithm Selection

Vincent Fischer: Implementation and Analysis of Load Balancing Options for AutoPas' Sliced Traversal

Bachelor's thesis submission talk. Vincent is advised by Fabio Gratl.

AutoPas is a free and open-source library written in C++ providing multiple efficient algorithms for particle simulations. In order to fully exploit modern processors, it is necessary to distribute the computational load evenly among multiple threads. One of the shared-memory parallelization approaches implemented in AutoPas cuts the simulation domain into multiple equally large slices and assigns one slice to each thread. However, for inhomogeneous particle distributions, this does not distribute the computational load equally among the threads.

In this thesis, multiple variants of sliced traversal for different particle containers were implemented, which aid in balancing the computational load among the threads. The new traversals were benchmarked using two different particle simulations, which were run on the Haswell architecture based CoolMUC-2 cluster. The results show that each of the new traversal options outperforms the original unbalanced sliced traversal for the same container for at least one scenario. As AutoPas uses autotuning to determine the best traversal for the scenario, this results in an absolute improvement.

Keywords: AutoPas, Load Balancing, Molecular Dynamics