SC²S Colloquium - August 27, 2013

From Sccswiki
Jump to navigation Jump to search
Date: Aug 27, 2013
Room: 02.07.023
Time: 3:00 pm, s.t.

Ludmilla Kreschenska: Modellierung städtischer Infrastrukturnetzwerke als gekoppelte nicht-lineare Dynamische Systeme in Kooperation mit Siemens

Für eine nachhaltige Stadtentwicklung ist eine Modellierung und Simulation der Stadt interessant. In dieser Arbeit wird die Stadt als ein gekoppeltes dynamisches System modelliert, bestehend aus einem Verkehrs- und Energienetzwerk. Dabei entsteht ein sehr großes komplexes diskretes nicht-lineares dynamisches System. Die Berechnung der Lösung für dieses System ist auch für leistungsstarke Rechner sehr ungenau und die Rechenzeit ist zu groß. Um diese Probleme zu vermeiden sind Verfahren der Modellordnungsreduktion notwendig. Für die MOR kann die Discrete Empirical Interpolation Method (DEIM) in Kombination mit der POD-Galerkin Projektion verwendet werden.

Robert Guder: Parallelization and Optimization of a Random ODE Benchmark Software

The basis of this Bachelor’s thesis is the model of ground motion by Kanai and Tajimi, which is used to implement a simulation of several multi-story buildings moving due to the effect of earthquake-induced ground motions. It is solved numerically using averaged Euler schemes. Starting from an existing C-implementation, the aim of this thesis is the optimization for single-core architectures and the parallelization for shared memory systems and distributed memory systems.

The basis of the optimization of the software forms a thorough performance analysis which helps us to identify its bottlenecks. We introduce a performance model as a general way to determine the maximum performance of an implementation on a given system. Therefore, an important part is to determine the memory bandwidth on all systems investigated.

Starting with an overview of available parallel computing and optimization techniques we identify the most important methods which improve the performance of our given software. These methods are vectorization, memory management and the usage of Intrinsics. Moreover, some routines of the Intel Math Kernel Library are benchmarked and examined for further usage in the Random ODE Benchmark Software. Applying all the techniques presented we achieved an improvement of runtime-performance of more than 330 percent on a single core and a good parallel scaling efficiency. Furthermore, the general memory requirements are reduced of approximately two-thirds for sufficiently small time steps. With this we form a good basis for the further investigation and optimization of more complex random ODE problems.