Parts | Days | Selection | Search | Downloads | Help
MM: Fachverband Metall- und Materialphysik
MM 20: Poster Session I
MM 20.19: Poster
Tuesday, March 24, 2009, 14:45–16:30, P4
Performance analysis of a parallel simulator for phase-field models — •Alexander Vondrous, Michael Selzer, Britta Nestler, and Marcus Jainta — Institute of Computational Engineering (ICE), Karlsruhe, Germany
Parallel algorithms for the numerical solution of phase field models are used to improve calculation times. To simulate a large section of a complex microstructure and to consider effects such as fluid flow on the morphology evolution, large computational domains have to be considered. In such cases parallelization is needed to significantly increase the speed of the computation. To classify the performance of parallel computations, infomation about the used hardware is important. Performance characteristics are, among others, the speedup and the efficiency. The results of the presented performance analysis are used to optimize parallel algorithms and hardware utilization. The numerical scheme avoids calculations in bulk phase regions of the domain leading to an inhomogeneous load balance and an undesirable higher idle time of some of the CPUs in the network. To minimize this effect, dynamic domain decomposition during run time is applied. Simulations of dendritic growth in a flow field serve to evaluate the performance for different numbers of CPUs.