Efficient parallelization for volume-coupled multiphysics simulations on hierarchical Cartesian grids

Abstract

One of the main challenges for multiphysics simulations of volume-coupled problems is data exchange, which becomes a serious bottleneck for large-scale applications executed on distributed memory computer architectures. Unlike surface coupling, the transfer of volumetric information via network communication can be too inefficient. To circumvent this problem, a new concept based on a joint hierarchical mesh is proposed, where the coupled solvers use different cells of a single shared grid. A domain decomposition method based on space-filling curves guarantees that all coupling data can be exchanged locally by in-memory operations. Furthermore, the interleaved execution of the solvers ensures that there is no additional overhead due to communication barriers. To demonstrate the efficiency of the presented method for realistic three-dimensional problems, it is used to predict the acoustic field of a turbulent jet in a direct-hybrid simulation, in which fluid dynamics and aeroacoustics are directly coupled. The results confirm the good parallel scalability of the approach and show the design of the coupling algorithm to be more efficient than a standard hybrid flow-aeroacoustics scheme coupled via disk I/O.