Improvements in Efficiency and Scalability of Wall-Distance Computations for Turbulence Modeling

The distance from a grid point to the closest wall surface, wall distance, is a fundamental quantity in turbulence modeling. An accurate, efficient, and scalable wall-distance method for turbulence modeling is critically important as the size of computational grids has significantly increased in recent years. A search-based wall-distance method that is suitable for general unstructured grids and tailored to requirements specific for turbulence models was previously developed and implemented in a high-performance computational fluid dynamics (CFD) solver. The method is referred to as SWIFT-0 (Search-based Wall-distance Implementation For Turbulence). While SWIFT-0 provided accurate wall distances and a significant reduction of the computational cost relative to state-of-the-art wall-distance methods, scalability still was lacking in parallel computations using the Message Passing Interface (MPI) because the entire wall surface was gathered on each MPI rank. This paper describes SWIFT-1 with a new implementation of the data-preparation algorithm in which each MPI rank gathers only selected wall vertices and faces with the potential of being closest to the grid points on the partition assigned to the rank. The selected vertices and faces are identified and exchanged between ranks in parallel using carefully designed data structures and algorithms. In wall-distance computations on the same grid, SWIFT-1 provides a significant speedup and computes the wall distance that is identical to the wall distance computed with SWIFT-0. However, further improvements are needed to achieve fully scalable performance. This paper documents SWIFT-1 and assesses the wall-distance calculations performed on unstructured grids generated for the Fifth AIAA CFD High-Lift Prediction Workshop and on a dynamic, overset grid system generated for a Lift+Cruise urban air mobility (UAM) concept aircraft.