Efficient coupling of fluid and acoustic interaction on massively parallel systems

Multi-scale problems like the generation of sound in a flow field and its sound wave propagation in the far field have become increasingly important in the design phase of industrial devices: One example is noise reduction of aircrafts or wind turbines. Although the generation of sound as well as its propagation can both be described by the same governing equations, wave propagation is a linear phenomenon and its equations can be simplified, which results in less computational effort. Additionally, the generation of sound in a flow field occurs at small spatial scales, while its propagation in the far field has to be observed on a large spatial scale. These large differences in scales are particularly challenging for numerical simulations. Resolving the entire domain with the high resolution that is required for the small scales of the flow domain is impossible due to the vast computational demand. In this thesis, we propose a partitioned coupling approach for the efficient simulation of such problems on massively parallel supercomputers. In partitioned coupling the physical domain is split into smaller subdomains, each covering a different physical phenomenon. Their interaction is realized by exchanging data at the joint coupling interface. Subsequently, these subdomains can be solved with numerical methods, resolutions, and equations tailored to the local physical requirements. Even different solvers can be used. However, this approach also holds numerical challenges at the coupling interface: E.g. for a consistent data exchange in case of different spatial resolutions, direct data evaluation, or efficient interpolation methods are necessary. Within this work, two different approaches of partitioned coupling are implemented and compared: A blackbox and a white-box approach. The black-box approach is characterized by a flexible choice of numerical solvers which allows for a wide range of different applications. Its generality comes with limited access to information inside each solver and, therefore, with a potential loss of performance. However, a black-box approach only acts on point data at the coupling interface and therefore requires external interpolation methods for a consistent coupling in space which is expected to be less efficient than solver-internal data mapping. In contrast, the white-box approach is fully integrated within one numerical framework. Accordingly, it can access solver-internal data mapping methods which promises better numerical results. This tight integration allows for the exploitation of knowledge about internal data structures and, therefore, yields performance benefits. On the other hand, it comes with less flexibility. Both strategies will be compared with respect to quality of data mapping at the coupling interface as well as performance on modern supercomputers. In order to achieve the best performance, the optimal load balancing strategy for a coupled setup is investigated. The benefits of the partitioned coupling approach are demonstrated on an industrial application of a 3D free-stream jet with a high Reynolds number showing that a multi-scale problem can be simulated using today’s compute resources.