102nd NIA CFD Seminar: A GPU Accelerated Adjoint Solver for Shape Optimization
Date: Tuesday, June 20, 2018
Time: 11am-Noon (EDT)
Room: NIA, Rm 137
Speaker: Asitav Mishra, Assistant Research Scientist, University of Maryland
Abstract: A graphics processing units (GPUs) accelerated adjoint-based optimization platform is proposed in this paper. Significant speed up gains and strong linear scalability of an existing in-house developed three-dimensional structured GPU Reynolds Averaged Navier-Stokes solver is presented first. As a first step towards the proposed GPU adjoint solver, a two-dimensional structured adjoint Euler solver is developed. The adjoint solver is further utilized to set up an airfoil shape optimization framework in Python and demonstrated for an airfoil shape optimization inverse problem. The two-dimensional adjoint Euler solver is extended to incorporate GPU acceleration using Compute Unified Device Architecture (CUDA) kernels and named ADjoint-GARfield (ADGAR). The adjoint optimization platform, ADGAR, is verified to a high accuracy of 14 significant digits with the serial adjoint Euler solver. Diagonalized Alternate Direction Implicit (DADI) iterative implicit schemes for both the forward and adjoint formulations are implemented and accelerated using CUDA kernels. The GPU accelerated structured code is finally successfully utilized to perform several airfoil shape optimizations for inverse design problems. Significant speedup up to 20x is observed using ADGAR for computations on a single GPU over a single CPU core.
Speaker Bio: Asitav Mishra is an Assistant Research Scientist in the Department of Aerospace Engineering at the University of Maryland as well as at the NIA since Oct 2017. His earlier research experiences include post-doctoral scholar positions at the University of Michigan (2015-2017) and the University of Wyoming (2012-2015) following his Ph.D in Aerospace Engineering from the University of Maryland in 2012. His research interests include adjoint based coupled multi-disciplinary fixed and rotary-wing design optimization, vortex wake-lifting surface interactions as well as performance predictions in rotary wing flows, and high performance computing using heterogenous GPU/CPU computing paradigms applied to CFD problems.