Improving computational efficiency of lattice Boltzmann methods on complex geometries

The lattice Boltzmann method has evolved over the past decade to a common technique in the field of computational fluid dynamics. The basic principle of cellular-automata allows an easy simulation of discrete surfaces and implementation of different boundary conditions. To meet the numerical requirements of continuous surfaces the enhanced Bouzidi bounce-back with interpolation was implemented. As the LBM requires a fine grained grid base, thus high memory usage, a particular emphasis was put on performance issues and performance optimization. To meet the increased obstacle to fluid ratio of unsteady surfaces, an unstructered list format was introduced, which turned out to be insensitive to the obstacle fluid ratio in terms of performance. The unstructured list algorithm with enhanced spatial data locality, finally performed nearly equal to the full matrix implementation. To exploit the growing parallelism in state-of-the-art compute nodes, a shared memory parallelization approach was implemented.