Lattice Boltzmann Performance Engineering in OpenLB

Lattice Boltzmann Methods (LBM) are an established mesoscopic approach for simulating a wide variety of transport phenomena. While they are uniquely suited to HPC usage, obtaining maximum performance across different targets and applications is an active topic of research. The open source C++ framework OpenLB is one of the major freely available LBM research codes, being in active development since 2007. In recent years, a concerted and successful effort was undertaken to improve the performance on state-of-the-art heterogeneous HPC clusters by adding e.g. support for GPUs and vectorized execution on CPUs alongside general code improvements. This talk will summarize the chosen refactoring approach alongside recent scalability benchmarks on the HoreKa supercomputer at KIT as well as some research results on optimizing single node GPU and SIMD CPU performance