This work addresses the task of dramatically accelerating the computation of 3-D unsteady aeroelastic and store separation flows via the use of commodity Graphics Processing Units (GPUs). Particular emphasis is given to the end-to-end GPU acceleration of a complete application, viz. ANDSolver, which solves the unsteady compressible Euler equations (Phase I) and the Reynolds-averaged Navier-Stokes equations (Phase II) on dynamic overset unstructured meshes. Novel data structures, graph coloring, and thread memory sharing are employed so that the high performance memory hierarchy on GPUs is efficiently utilized even for computations arising from unstructured meshes. Preliminary single GPU acceleration experiments for the numerical flux computation in ANDSolver show a 20x performance speedup over a single high-end COTS CPU. From projected improvements in GPU hardware, single GPU performance speedups of 30x to 50x are expected during Phase I & II.
Keywords: Hardware Accelerators, Aeroelastic Analysis, Store Separation, Computational Fluid Dynamics, Gpu Computing
