Riassunto analitico
In recent years, high-order numerical methods have been proposed for Computational Fluid Dynamics simulations. These are particularly appealing because of its high order accuracy and competitive computational cost. Furthermore, the architecture of new computers has changed, using graphics processing unit (GPU) accelerators for CFD simulations. This can significantly reduce simulations times, and we must adapt with new solvers. In particular, one of the more interesting and recent solvers -- which was used in the present project -- is PyFR. This implements high-order numerical methods and runs on both traditional CPUs and modern GPUs. This is an open-source Python based framework which can solve different governing systems on mixed unstructured grids, using the Flux Reconstruction approach of Huynh. It was tested with a fourth-order ILES simulation of a flow around a cylinder in subcritical regime, with Re=3900, running on a cluster of NVIDIA Volta V100 GPUs. This is a difficult test case because it includes a variety of fluid dynamics phenomena and because, having no edges, it does not give information on the separation point. A detailed post-processing of the simulation was made, and the results that have been found are in agreement with experimental data. A performance comparison between CPUs and GPUs performance was conducted.
|
