97th NIA CFD Seminar Webcast: Development and Implementation of Reconstructed Discontinuous Galerkin Methods for Computational Fluid Dynamics on GPUs by Jialin Lou

January 19, 2018 - Leave a Response

97th NIA CFD Seminar

Topic: Development and Implementation of Reconstructed Discontinuous Galerkin Methods for Computational Fluid Dynamics on GPUs

Date: Tuesday, January 30, 2018

Time: 11am-noon (EST)

Room: NIA, Rm137

Speaker: Jialin Lou

Speaker Bio: Dr. Jialin Lou earned his B.S. degree in Engineering Mechanics at Beijing Institute of Technology and M.S. and Ph.D. degree from North Carolina State University under Dr. Hong Luo’s advice. He recently joined Old Dominion University Research Foundation as a post-doctoral research associate, working with Dr. Nail Yamaleev in Mathematics and Statistics Department. His research interest lies in high order numerical methods, hyperbolic diffusion schemes, and High Performance Computer (HPC) parallel computing in both CPU and GPU platforms.

Abstract: The objective of the effort presented in this work is to port an unstructured CFD solver, reconstructed discontinuous Galerkin flow solver (RDGFLO), onto GPU platform using OpenACC. The solver is based on a third-order hierarchical Weighted Essentially Non-Oscillatory (WENO) reconstructed DG methods. By taking advantages of the OpenACC parallel programming model, the presented scheme requires the minimum code intrusion and algorithm alteration to upgrade a legacy CFD solver without much extra time and effort in programming, resulting in a unified portable code for both CPU and GPU platforms. A number of inviscid and viscous flow problems are presented to verify the implementation of the developed schemes on the GPU. Strong scaling tests are carried out to compare the unit running time on single GPU and single CPU to obtain the speedup factor of the developed methods. Also, weak scaling tests are used for several cases to test the parallel efficiency for multi-GPU computing by comparing the unit running time with different number of GPU cards for an approximately fixed problem size per GPU card. The results of timing measurements indicate that this OpenACC-based parallel scheme is able to significantly accelerate the solving procedure for the equivalent legacy CPU code.

Additional information, including the webcast link, can be found at the NIA CFD Seminar website:

http://www.hiroakinishikawa.com/niacfds/index.html

niacfds_square

Advertisements

TODAY: 96th NIA CFD Seminar Webcast: Hamiltonian-Strand (HAMSTR) Approach Using Hybrid Meshes for Aerodynamic Flow Analysis by Yong Su Jung

November 28, 2017 - Leave a Response

96th NIA CFD Seminar

Topic: Hamiltonian-Strand (HAMSTR) Approach Using Hybrid Meshes for Aerodynamic Flow Analysis

Date: Tuesday, November 28, 2017

Time: 11am-noon (EST)

Room: NIA, Rm101

Speaker: Yong Su Jung

Speaker Bio: Yong Su Jung is a Ph.D candidate student in Aerospace Engineering department at the University of Maryland. He holds B.S (2012). and M.S (2014) in Aerospace Engineering from Korea Advanced Institute of Science and Technology. His research interests are in developing and applying Computational Fluid Dynamics methods for external flow simulations, such as rotary wing. His research has been funded by Department of Defense (DoD) HPCMP CREATE-AV program and Korea Aerospace Research Institute. He was a member of the University Maryland team received first place in the graduate category for 2016 AHS Design Competition.

Abstract: A solution framework using Hamiltonian paths and strand grids (HAMSTR) is presented for two and three-dimensional flows. The methodology can create a volume mesh starting from either an unstructured surface mesh comprised of mixed triangular-quadrilateral elements or a fully unstructured volume mesh. “Line structure” through the meshes are found in a robust manner and the flow solver uses line-implicit schemes and stencil-based discretization along these lines, similar to a structured grid flow solver. The framework has been developed mostly for rotorcraft CFD simulations, which requires robust mesh generation around complex geometry and efficient numerical method for large scale problems.
HAMSTR is a 3D compressible finite volume solver that can operate across multiple processors using MPI. Hybrid RANS/LES turbulence modeling based on the Spalart-Allmaras turbulence model and γ-〖Re〗_θ-SA laminar/turbulent transition model of Medida-Baeder are integrated into the solver for better predictions of the boundary layer and resulting flowfield as compared with a fully turbulent RANS simulation. Furthermore, deformable meshes can also be handled for elastic body simulations such as rotor blades. An overset technique (using TIOGA) allows for a hybrid mesh system, which consists of a near-body Hamiltonian/Strand grid and off-body Cartesian nested meshes. The integration framework between the various components of the code is performed using Python to allow for ease of integration to other codes in the research group. The current infrastructure is used to explore various cases ranging from simple representative geometries, such as 2D airfoil, to complex geometries such as rotating rotor hub and a full wind turbine. Some, CFD/CSD predictions for a slowed rotor are also studied.

Additional information, including the webcast link, can be found at the NIA CFD Seminar website:

http://www.hiroakinishikawa.com/niacfds/index.html

niacfds_square