Dresden 2020 – wissenschaftliches Programm

Die DPG-Frühjahrstagung in Dresden musste abgesagt werden! Lesen Sie mehr ...

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

BP: Fachverband Biologische Physik

BP 17: Poster V

BP 17.24: Poster

Dienstag, 17. März 2020, 14:00–16:00, P2/1OG

MPI-based multi-GPU extension of the Lattice Boltzmann Method — •Fabian Häusl, Moritz Lehmann, and Stephan Gekle — Biofluid Simulation and Modeling, University of Bayreuth, Germany

The lattice Boltzmann method (LBM) is a highly versatile flow solver which benefits greatly from graphics processing unit (GPU) computing. However, the LBM is very memory-intensive while at the same time the on-board memory of GPUs is quite limited, which directly restricts simulation domain size. This poster presents a muli-GPU implementation based on the framework OpenCL and the Message Passing Interface (MPI) which is able to widen this limitation as well as to gain additional speedup. By using spezialized buffer types and memory layouts as well as applying the concept of templates to OpenCL-kernels in order to reduce memory access, it is precisely tailored to the requirements of GPUs and MPI. It differs from comparable implementations especially in that the domain can be subdivided along all three spatial directions. The communication scheme remains independent of the velocity set selected, can easily be adapted to the various extensions of the LBM and guarantees optimal buffer bandwidth. Communication time consumption can be hidden for the most part by overlapping it with computation, so the algorithm can reach 95% of its theoretical optimum in the weak-scaling and 13000 MLUPs using 4 Radeon VII GPUs for a cubic benchmark setup can be observed.