A hardware locality-informed dynamic load balancing for a p-adaptive local discontinuous Galerkin LES solver

The current work introduces a dynamic, hardware locality-informed load balancing strategy in a p-adaptive local discontinuous Galerkin solver suitable for Large Eddy Simulation. Firstly, the problem is reformulated as a graph partitioning problem. Different graph’s weights are proposed to estimate the load and communication imbalances and the implied predictions are compared with numerical experimental measures. Then, the ultimate goal is achieved by combining a proper binding of MPI tasks to the hardware with a hierarchical partitioning strategy based on the Zoltan library. Within this context, both geometric and connectivity-based partitioning methods are considered, discussing their advantages and disadvantages. The effectiveness of the proposed implementation is demonstrated by simulating the advection of a viscous vortex and the flow past a square cylinder.