GPU parallelization of multigrid RANS solver for three-dimensional aerodynamic simulations on multiblock grids

In this paper, graphical processing units (GPUs) are leveraged to accelerate Bombardier's full aircraft Navier-Stokes solver, a finite-volume, cell-centered RANS flow solver for multiblock structured grids. The efficiency of different parallel smoothers on GPUs is studied, in the context of solving the RANS equations with a nonlinear full approximation storage multigrid scheme. Many variants of parallel red-black Gauss-Seidel and Jacobi solvers are investigated and their efficiency compared against sequential algorithms such as the lower-upper symmetric Gauss-Seidel solver on both CPUs and GPUs. Parametric studies on three-dimensional aircraft configurations are performed to identify the optimal smoothers and determine the optimal number of smoothing iterations on each multigrid level. Furthermore, the efficiency of different approaches to overlapping the communication and computation for the MPI-CUDA implementation in the multi-GPU code is investigated. The results show that the best runtime with the GPU code is obtained using a weaker smoother with more sweeps per multigrid level, whereas the best runtime with the CPU code is obtained using a stronger smoother with fewer sweeps per multigrid level. Despite using a weaker smoother and therefore more iterations to converge to the final solutions, the GPU-accelerated code is significantly faster than the CPU code.