New paper published: Dynamic load balancing for direct-coupled multiphysics simulations

High parallel efficiency for large-scale coupled multiphysics simulations requires the computational load to be evenly distributed among all compute cores. For complex applications and massively parallel computations, even minor load imbalances can have a severe impact on the overall performance and resource usage. Exemplarily for a volume-coupled multiphysics simulation, a direct-hybrid method is considered, in which a CFD and a CAA simulation are performed concurrently on the same parallel subdomains. For differing load compositions on each subdomain, accurate computational weights for CFD and CAA cells must be known to determine an efficient domain decomposition. Therefore, a dynamic load balancing scheme is presented, which allows to increase the efficiency of complex coupled simulations with non-trivial domain decompositions. A fully-coupled three-dimensional jet simulation with approximately 300 million degrees of freedom demonstrates the effectiveness of the approach to reduce load imbalances. A detailed performance analysis substantiates the necessity of dynamic load balancing. Furthermore, the results of a strong scaling experiment show the benefit of load balancing to be proportional to the degree of parallelism. In addition, it is shown that the approach allows to attenuate imbalances also for parallel computations on heterogeneous computing hardware. The acoustic field of a chevron nozzle will also be discussed.