HPC performance study of different collision models using the Lattice Boltzmann solver Musubi

Over the past decades, the lattice Boltzmann method (LBM) has become increasingly popular thanks to its capabilities in the domain of Large-Eddy Simulations (LES). Different collision schemes have been proposed to extend the scope of application to higher Reynolds number flows. This study compares the accuracy and the performance of some of these schemes on a D3Q27 lattice, including the original Multiple Relaxation Times (MRT) model, the Hybrid Recursive Regularized Bhatnagar-Gross-Krook (HRR) operator, as well as the Projected Recursive Regularized Bhatnagar-Gross-Krook (PRR) operator and the parametrized Cumulant collision scheme. For this purpose, the above-mentioned schemes are implemented in the HPC LBM solver Musubi and tested on a well-documented test-case describing the flow past a circular cylinder at a Reynolds number of 3900. Three different subgrid scale (SGS) models are used to account for the unresolved turbulence, i.e. the Smagorinsky model, the Wall-Adapting Local Eddy-viscosity (WALE) model, and the Vreman model. The Cumulant scheme uses an Implicit LES (ILES) subgrid scale model and shows the best agreement with the experimental data followed by MRT with WALE, and HRR with Vreman. The examined collision models are able to capture the second peak at f = 3fvs of the power spectra density of the y velocity component first discovered in experiments. With respect to performance, the collision models are compared in terms of MLUPs/node and parallel efficiency for a strong scaling analysis. Again the Cumulant scheme outperforms the other collision models even when they are run on the reduced D3Q19 stencil. All the collision schemes show a strong scaling parallel efficiency above 60% on up to 16384 cores in our implementation.

Automated summary

This study compares the accuracy and performance of several collision schemes on a D3Q27 lattice. The parameterized Cumulant collision scheme shows the best agreement with experimental data in the flow past a circular cylinder test at a Reynolds number of 3900. In terms of performance, all collision schemes achieve a strong scaling parallel efficiency above 60% on up to 16384 cores.