GTC Silicon Valley-2019: Unstructured Grid CFD Algorithms for NVIDIA GPUs
GTC Silicon Valley-2019 ID:S9369:Unstructured Grid CFD Algorithms for NVIDIA GPUs
Eric Nielsen(NASA Langley Research Center),Aaron Walden(NASA Langley Research Center)
We'll describe the transition of an unstructured-grid computational fluid dynamics (CFD) code from a dense MPI model to shared memory models suitable for a many-core landscape. Unstructured grid approaches are often used in CFD to solve Navier-Stokes equations because they accommodate geometric complexity. Turbulent flows encountered in aerospace applications generally require highly anisotropic meshes, driving the need for implicit solution methodologies to efficiently solve discrete equations. We'll discuss node-level studies of computationally intense CFD kernels on traditional x86 architectures and on NVIDIA GPUs. We'll also talk about scaling studies performed at several large supercomputing facilities.