SCCS Colloquium - May 8, 2019
|Date:||May 8, 2019|
|Time:||15:00 - 16:00|
Nico Rösel: Combigrid Based Dimensional Adaptivity for Sparse Grid Density Estimation and Classification
In the frame of this thesis the sparse grid combination technique was used to expand the density estimation based approach to classification in the SG++ framework. Thanks to the independence of the component grids, it was possible to implement much faster dimensional adaptive refinements. While the old implementation had to refit the whole grid after each refinement step, now just the additionally added components have to be fitted. The thesis contains the theoretical background, a description, and an evaluation of the implementation.
Ravil Dorozhinskii: Configuration of a linear solver for linearly implicit time integration and efficient data transfer in parallel thermo-hydraulic computations
This is a Master's thesis submission talk. Ravil is advised by Tobias Neckel.
An application of linearly implicit methods for time integration of stiff systems of ODEs results in solving sparse systems of linear equations. An optimal selection and configuration of a parallel linear solver can considerably accelerate the time integration process. A comparison of iterative and direct sparse linear solvers has shown that direct ones are the most suitable for this purpose because of their natural robustness to ill-conditioned linear systems that can occur during numerical time integration. Testing of different direct sparse solvers applied to systems generated by ATHLET software has revealed that MUMPS, an implementation of the multifrontal method, performs better than the others in terms of the overall parallel execution time.
In this study, we have mainly focused on configuring MUMPS with the aim of improving parallel performance of the solver for thermo-hydraulic computations within a single node of GRS compute-cluster. However, the overall approach, proposed in the study, may be considered as a general framework for a selection and adaptation of a linear sparse solver for solving problem-specific systems of linear equations on distributed-memory machines. Additionally, we have shown that an intelligent application of non-blocking MPI communication in some parts of the existing thermo-hydraulic simulation code, ATHLET, can additionally solve issues of inefficient data transfer preserving the current software design and implementation without drastic changes of the source code.
Keywords: direct sparse linear methods, multifrontal methods, MUMPS, BLAS, parallel performance, distributed-memory computations