On the Effects of Parallelization on the Flow Prediction around a Fastback DrivAer Model at Different Attitudes

When Computational Fluid Dynamics (CFD) is used in the development of road vehicles for passenger and performance use, the fidelity and numerical accuracy of the simulation are paramount as manufacturers strive to optimize the vehicle down to the single aerodynamic count. While much research has been performed on how the choices of simulation model or grid size affects the simulation results, very little has been done to investigate how the spatial decomposition of the domain amongst different nodes of a high-performance computing unit (HPC) influences the results of the simulation. As simulations grow larger, more nodes are required to reduce the simulation time, however in most commercial software this introduces a new form of error due to the accumulation of round-off errors created in the intra-node communication schemes used during iterations. This study investigates this issue by testing the repeatability and accuracy of a generic passenger car simulation using a state-of-the-art HPC along with a commercially available CFD software. This was accomplished using a readily available generic passenger vehicle body known as the DrivAer model along with the commercial software StarCCM+ which uses a Message Passing Interface (MPI) to compute solutions across many different cores in a parallel load balancing fashion. A methodology was created which tested the simulation’s ability to create accurate and reproducible predictions in complex pitch and yaw configurations as the number of computing cores varied.