SCCS Colloquium - Mar 18, 2020

From Sccswiki
Jump to navigation Jump to search
Date: March 18, 2020
Room: Online
Time: 15:00 - 16:00

Due to regulations regarding the COVID-19 situation, we are currently looking into online alternatives for the colloquium. In our first attempt, we are trying a videoconference with LRZ Meet. You only need a browser and your LRZ credentials.

Anna Mittermair: Efficient Time Integration of ODEs Using Different Parallel Architectures, Programming Models and Integration Methods [postponed for technical reasons]

IDP submission talk. Anna is advised by Martin Schreiber.

Inferring the parameters of a system of differential equations is a nonlinear optimization problem, which often requires solving several independent systems in parallel. Especially for high-dimensional parameter spaces this is computationally demanding and requires the use of high-performance computer systems. Writing code manually for each combination of ODE, target architecture and time discretization method is extremely time-consuming and error-prone. This can be solved by using a domain-specific language that allows automatic code generation and the use of different backends. In this project, automatic code generation was implemented for several parallel programming models and time discretization methods. This was used to evaluate the performance of different combinations of architecture, programming model and discretization method. The results show that all these components have significant impact, but the optimal combination varies depending on the specific optimization problem.

Keywords: OpenCL, OpenMP, CUDA, time integration, parameter optimization

Jakob Englhauser: Application and Evaluation of Auto-Tuning Tools in Molecular Dynamics Simulations

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

The simulation of molecular dynamics with high numbers of particles is a very computationally intensive task. To keep resource and time consumption minimal, it is, therefore, necessary to use algorithms that are well suited for the specific simulation environment. One library that can be used by N-body simulation programs is AutoPas, which can automatically select optimal combinations of algorithms during runtime through different tuning strategies. In this thesis, external auto-tuning tools, namely Active Harmony, were integrated into AutoPas and evaluated by simulating a common physical mechanism, the spinodal decomposition. In this simulation, Active Harmony was able to achieve a lower runtime than previous, in AutoPas implemented tuning strategies by optimizing configuration selection during the tuning process.

Keywords: AutoPas, Active Harmony, Auto-Tuning