SC²S Colloquium - May 30, 2018

From Sccswiki
Jump to navigation Jump to search
Date: May 30, 2018
Room: 02.07.023
Time: 15:00 - 16:00

Felix Späth: Optimizing communication strategies for a massively parallel multi-phase multi-resolution framework

This is an external Master's Thesis completion talk advised by Michael Bader

Complex flow phenomena play a significant role in future biomedical treatments. For example in Histotripsy, shock-bubble interactions, induced by ultrasound, might allow the destruction of malicious tissue or drug inducement into cells as in Sonoporation. In order to optimize those treatments, large scale 3D simulations are necessary. Efficient communication patterns are crucial, in order to simulate more than one billion degrees of freedom. Multiple measures have been taken and are presented in this thesis in order to increase scalability from 32 to 512 nodes of the SuperMUC. The theoretical limit in simulation size induced by a limited amount of MPI tags is solved by a counter based tagging strategy. Simulation accuracy is increased by a change of the halo update pattern, from indirect to direct diagonal exchange. Furthermore, a bucket based shared memory parallelization approach is presented that has a weak scaling efficiency of 60% for a scaling factor of 512.

Keywords: MPI, massively parallel, SuperMUC

David Damerow: Coupling general purpose PDE solvers with a Combination Technique Framework

Bachelor's Thesis completion talk advised by Michael Obersteiner

This bachelor’s thesis deals with the coupling of the combination technique and a general purpose PDE solver in Python. For the FEniCS project which was chosen as the PDE solver in this thesis and all other solvers the curse of dimensionality limits the number of dimensions that can be simulated. Solving PDEs with the finite element method in high dimensions quickly leads to memory problems. The combination technique provides an approximation of high quality for these large-scale problems. It uses the solutions of coarse grids to approximate the solution on fine-mesh grids. However, it is not commonly agreed upon for which PDEs the combination technique does or does not work. The coupling of a general purpose PDE solver and the combination technique makes it possible to investigate this matter further. This thesis implements a basis for this coupling. In this thesis no PDE was found for which the combination technique did not work properly. However, only a few of them - the Poisson, diffusion, linear elasiticy and the Navier-Stokes equations - were tested. The combination technique found the expected solutions for all of them with a small exception of a nonlinear Poisson equation and the Navier-Stokes equation. For both equations the results of the component grids of the combination technique were more precise than the result of the combination technique itself.

Keywords: Sparse Grid Combination Technique, FEniCS, partial differential equations