Personal tools

SC²S Colloquium - Jan 24, 2018

From Sccswiki

Revision as of 14:15, 10 January 2019 by Makis (Talk | contribs)

Jump to: navigation, search
Date: Jan 24, 2018
Room: 00.08.053
Time: 15:00 - 16:15

Sebastian Burreiner: Comparison of classic and task-based scheduling for the local time-stepping in SeisSol using OpenMP

This is a Master's thesis introduction talk advised by Carsten Uphoff.

The earthquake simulation software SeisSol uses local time-stepping to efficiently calculate large-scale simulations. The multi-threaded work distribution is done by OpenMP and currently a prototype exists that uses the OpenMP task construct. This thesis looks into different scheduling options and compares approaches possible through OpenMP tasks with more classic methods like list-based scheduling as well as the current implementation. The goal is to minimize load imbalances and thus improve the overall runtime.

Keywords: SeisSol, HPC, load balancing, local time-stepping, list-based scheduling

Language: English Eng small.jpg

Lukas Krenz: Cloud Simulation with the ExaHyPE-Engine

This is a Master's thesis submission talk advised by Leonhard Rannabauer.

ExaHyPE is an engine for hyperbolic PDEs that uses the ADER-DG scheme.

We use the compressible Navier-Stokes (NS) equations with heat conduction to simulate the dynamics of clouds. Due to the diffusive components, the NS-Equations are not hyperbolic. We thus need to modify parts of the numerical scheme.

Furthermore, we extend the equation set to the reactive compressible Navier-Stokes equations which include a source term that models chemical reactions.

Keywords: ExaHyPE, ADER-DG, CFD

Language: English Eng small.jpg

Jan Nguyen: Shared-Memory Parallelization of Verlet Lists

This is a Bachelor's thesis submission talk advised by Steffen Seckler.

This thesis deals with different algorithms for molecular dynamics simulations. We focus on variants of the Verlet lists. The presented algorithms are implemented in the C++ library AutoPas. For acceleration, the library uses shared-memory parallelization. We ran performance tests of the various implemented methods on the CoolMUC-2 system of the Leibniz Supercomputing Centre. From which we learned that performance, unfortunately, does not grow linearly with the number of threads used. How well we can scale depends on various factors and the algorithm used.

Keywords: Molecular Dynamics, Verlet lists, AutoPas, HPC

Language: German Deu small.jpg