Mesh refinement on GPU for lattice Boltzmann methods: Application to subsonic and supersonic flow simulations
Please login to view abstract download link
We revisit the topic of mesh refinement on graphical processing units (GPUs) in the lattice Boltzmann (LB) framework via C++17 Parallel Algorithms. This leads to a non-invasive, hardware- and vendor-independent GPU-accelerated LB solver that can efficiently run on non-uniform grids whatever the velocity discretization. More precisely, the LB solver relies on the concept of numerical equilibrium/collision which can enforce physical constraints and positivity of the solution through a root-finding solver. This allows for the efficient, robust and accurate simulation of supersonic flows on GPUs with relatively small lattices. The grid refinement algorithm is based on the volumetric approach by Rhode et al., and which ensures data locality, efficiency and conservation rules at the interface between coarse and fine grids, whatever the lattice considered. The versatility of our approach is demonstrated through a series of 2D simulations spanning from subsonic to supersonic regimes. These simulations utilize both standard and multispeed LB solvers, showcasing a significant speedup in performance due to GPU acceleration.
