Highly Parallel Complex Moving Geometries on CPU and GPU using the Partially Saturated Cells Method

Suffa P, Kemmler S, Köstler H, Rüde U (2024)

Publication Language: English

Publication Type: Conference contribution, Abstract of lecture

Publication year: 2024

Pages Range: 1

Event location: ETH Zürich Rämistrasse 101 8092 Zürich Schweiz

Abstract

For computational fluid dynamics simulations of wind power plants, cooling fans, counter-
rotating open rotors (CROR) as in Figure 1, or similar applications, fluid-structure interaction
of moving boundaries poses a challenging task. To handle the movement of complex domains
with the Lattice Boltzmann Method (LBM) efficiently on CPUs and accelerators, we employ the
partially saturated cells method (PSM) [1].
For the PSM, the classical Lattice Boltzmann equation is extended by a solid collision oper-
ator, which functions similarly to a no-slip boundary. A solid volume fraction field indicates the
presence or absence of a geometry object in a specific cell and, therefore, decides the influence
of the solid or fluid collision ratio in the PSM kernel. Due to its similarities to pure LBM and,
therefore, its highly parallel nature, as well as due to the absence of the fresh node problem, as
it is needed for some other moving boundary techniques, the PSM is very efficient and suitable
for running on accelerator cards.
We implemented the PSM method in the multiphysics framework waLBerla to be able to
run large-scale MPI simulations. We also integrated the PSM in the code generation framework
lbmpy, which generates highly optimized compute kernels for different variations of the LBM.
With this approach, we can show a performance reduction for the handling of the moving
geometry of less than 20% in terms of mega lattice updates per second (MLUPS) compared
to an LBM-only kernel on single CPU and GPU. We can further show excellent strong scaling
performance of the CROR with the PSM for ∼ 4 · 109 cells on up to 32,768 CPU cores on
the HPC system LUMI, and on up to 1024 NVIDIA A100 GPUs on JUWELS Booster, ending
up with a maximum performance of 106 MLUPS. To summarize, the proposed approach en-
ables highly parallel simulations of complex moving geometries in fluid flow on state-of-the-art
supercomputers.

Authors with CRIS profile

How to cite

APA:

Suffa, P., Kemmler, S., Köstler, H., & Rüde, U. (2024). Highly Parallel Complex Moving Geometries on CPU and GPU using the Partially Saturated Cells Method. Paper presentation at 33rd Discrete Simulation of Fluid Dynamics (DSFD) Conference, ETH Zürich Rämistrasse 101 8092 Zürich Schweiz, CH.

MLA:

Suffa, Philipp, et al. "Highly Parallel Complex Moving Geometries on CPU and GPU using the Partially Saturated Cells Method." Presented at 33rd Discrete Simulation of Fluid Dynamics (DSFD) Conference, ETH Zürich Rämistrasse 101 8092 Zürich Schweiz 2024.

BibTeX: Download