Accelerated CFD computations on multi-GPU using OpenMP and OpenACC

With the demand for increased computing precision and a large-scale domain in many computational fluid dynamics problems, the computational load on the processor is getting heavier than ever. Graphical Processing Units (GPU) are an excellent computing platform for high-precision floating-points on huge computational loads. There is no direct provisioning for parallelizing code across multiple GPUs using OpenACC, a directive-based programming model. Hence, a hybrid type of programming is required to tackle this problem. In this present work, the hybrid-type (CPU+GPU) parallelization of the Poisson solver on multi-GPU was demonstrated using directive-based programming models (OpenMP and OpenACC), which reduced the computational time by 61x on multi-GPU as compared to a single CPU. We have further analysed a turbulent 3D lid-driven cavity flow by direct numerical simulation (DNS) using this multi-GPU solver. The numerical computations and experimental results were in good agreement.