Optimization of single node load balancing for lattice Boltzmann method on heterogeneous high performance computers

Abstract

Lattice Boltzmann Methods (LBM) are particularly suited for highly parallel computational fluid dynamics simulations on heterogeneous HPC systems combining CPUs and GPUs. However, the computationally dominant collide-and-stream loops commonly utilize only GPUs, leaving CPU resources underutilized. To overcome this limitation, this article proposes a novel load balancing strategy based on a genetic algorithm for bottom-up, cost-aware optimization of spatial domain decompositions. This approach generates subdomains and rank assignments inherently suited for cooperative execution on both CPUs and GPUs. Implemented in the open source framework OpenLB, the strategy is applied to turbulent flow reference cases, including a multi-physics reactive mixer. A detailed evaluation on heterogeneous HPC nodes demonstrates significant performance gains, achieving speedups of up to 87% compared to traditional GPU-only execution. This work therefore establishes cost-aware, bottom-up decomposition as a suitable strategy for exploiting the native heterogeneity of modern compute nodes.