A narrow band-based dynamic load balancing scheme for the level-set ghost-fluid method

Abstract

We present a dynamic load balancing scheme for compressible two-phase flows simulations using a high-order level-set ghost-fluid method. The load imbalance arises from introducing an element masking that applies the costly interface-tracking algorithm only to the grid cells near the phase interface. The load balancing scheme is based on a static domain decomposition by the Hilbert space-filling curve and employs an efficient heuristic for the dynamic repartitioning. The current workload distribution is determined through element-local wall time measurements, exploiting the masking approach for an efficient code instrumentation. The dynamic repartitioning effectively carries over the single-core performance gain through the element masking to massively parallelized simulations. We investigate the strong scaling behavior for up to 16384 cores, revealing near optimal parallel efficiency and a performance gain of factor five on average compared to previous, unbalanced simulations without element masking. The load balancing scheme is applied to a well-studied two- and three-dimensional shock-drop interaction in the Rayleigh–Taylor piercing regime, providing an overall runtime reduction of approximately 65%.