Parallelization Strategies and Efficiency of CFD Computations in Complex Geometries Using Lattice Boltzmann Methods on High-Performance Computers
A frequently stated property of the Lattice Boltzmann (LB) method is, that it is easy to implement and that the generation of computational grids is trivial even for three-dimensional problems. This is mainly due to the usually chosen approach of using full matrices to store the primary variables of the scheme. However this kind of implementation has severe disadvantages for simulations, where the volume of the bounding box of the flow domain is large compared to the actual volume of the flow domain. Thus the authors developed data structures which allow to discretize only the fluid volume including boundary conditions to minimize memory requirements, while retaining the excellent performance with respect to vectorization of standard LB-implementations on supercomputers. Due to extensive communication hiding using asynchronous non-blocking message transfer an almost linear parallel speedup is achieved
KeywordsLattice Boltzmann Method Memory Consumption Flow Domain Lattice Boltzmann Collision Term
Unable to display preview. Download preview PDF.
- 2.Krafczyk, M. (2001) Gitter-Boltzmann-Methoden - von der Theorie zur Anwendung. Professorial Thesis LS Bauinformatik TU MünchenGoogle Scholar
- 3.Karypis, G., Kumar, V. (1998) Multilevel Algorithms for Multi-Constraint Graph Partitioninghttp://www-users.cs.umn.edu/~karypis/publications/partitioning.html/~karypis/publications/partitioning.htmlGoogle Scholar
- 5.Engeln-Müllges, G., Reutter, F. (1996) Numerikalgorithmen. VDI Verlag 1996, ISBN 978–3–18–401539–4Google Scholar