Dynamic Load Balanced Simulations based on Hierarchical Cartesian Meshes#
Matthias Meinke (RWTH Aachan Univ.)
Abstract#
This paper presents technical applications, in which various solution methods are coupled to predict computational fluid dynamics (CFD) related multiphysics problems. The main focus is on the efficient implementation of the coupling strategy of the methods, which is realized in the unified solver framework m-AIA to be published as open source very soon. The solvers are based on hierarchical Cartesian meshes, where a coarse base level mesh is shared between the solution methods to allow an efficient domain decompositioning for parallel execution and dynamic load balancing also for solution adaptive mesh refinement. Various aspects of the coupled methods will be discussed for technical applications. The first comes from the important technical field of the mitigation of aerodynamically generated sound. It requires an accurate prediction of the turbulent flow and a computational aeroacoustics (CAA) method to determine the noise generation and the propagating sound field. A direct-hybrid simulation method is used, in which a solver for the flow field is directly coupled to a solution of the acoustic perturbation equations. It uses an interleaved execution pattern such that an efficient parallelization is achieved also on high-performance computing (HPC) systems. In the application of this volume coupled CFD/CAA solver for the prediction of flow generated noise, the benefit of a dynamic load balancing method is demonstrated even for static meshes. An interference between the CPU timers used for the rebalancing of the computational load and changing clock cycles of the CPU is shown to possibly occur in such cases. Another technical example discussed, is the mixing of liquid fuels with air using sprays, which is predicted using a Lagrangian particle tracking method for the spray droplets. Due to the strongly varying density of the spray particles during the injection phase, a dynamic load balancing is necessary to obtain a good parallel efficiency. It is shown that the use of a shared Cartesian background mesh also for the particle tracking allows an efficient domain repartitioning. Results presented will include the simulation of the mixing process in internal combustion engines. Strong scaling tests for up 4000 nodes of the HAWK system of HLRS Stuttgart are used to demonstrate the high efficieny of the implemented methods.