• Research,

LHEEA's Mini-seminar 26/10/2017: "GPU accelerated Weakly-Compressible Cartesian Hydrodynamics", Xuhui Li

The LHEEA organizes mini-seminars every Thursday afternoon. The next seminar will be held in the distance-learning room on Thursday 26 October at 1.45pm. Open to all!

On October 26, 2017 from 13:45 To 14:30

On October 26, 2017 from 13:45 To 14:30

On October 26, 2017 from 13:45 To 14:30

On October 26, 2017 from 13:45 To 14:30

On October 26, 2017 from 13:45 To 14:30

On October 26, 2017 from 13:45 To 14:30

On October 26, 2017 from 13:45 To 14:30

On October 26, 2017 from 13:45 To 14:30

On October 26, 2017 from 13:45 To 14:30

On October 26, 2017 from 13:45 To 14:30

On October 26, 2017 from 13:45 To 14:30

On October 26, 2017 from 13:45 To 14:30

On October 26, 2017 from 13:45 To 14:30

On October 26, 2017 from 13:45 To 14:30

On October 26, 2017 from 13:45 To 14:30

On October 26, 2017 from 13:45 To 14:30

This week:

Xuhui Li (post-doc H2I) will present (in English) his work at 1:45 PM in D028 on "GPU accelerated Weakly-Compressible Cartesian Hydrodynamics".


Abstract:
"Unlike the conventional incompressible hydrodynamic solvers,which are usually implicity time-marched with the  time-consuming Poisson equation solver.In the solver WCCH, an explict weakly compressible finite volume method  has been used. The Cartesian mesh is adopted for the purpose of algorithmic simplicity and implemention efficiency.  To improve the numerical accuracy and reduce the calculation cost,the adaptive mesh refinement has been developed.  The high-order WENO scheme has been performed to reconstruct the variables on the interfaces of the cells.Suitable
 Riemann solver has been applied to evaluate the flux aross the interface.The fouth-order Runge-Kutta temporal scheme  is used to update the ordinary differential equation.
 
 Based on the above remarkable characteristics of WCCH, in this work, the GPU version, GWCCH, has been developed.  Since the source code of WCCH has been developed with Fortran language for around 5 years by a team, the hybrid programming strategy of Fortran and CUDA C has been proposed. The GPU porting of each section has been finished one by one.  Finally, the full CUDA C verison of the finite volume solver except the AMR module has been developed. The complicated  mesh refinement is conducted by the AMR module on CPU every 20 time steps while the solution of the high intensive algorithmic operations  of the finite volume solver (WENO reconstruction, Riemann solver) is performed on GPU, in which both the advantages of CPU and GPU  have been fully ultilized. The performance test is conducted on the GTX 1080 and K80 on ICI. The speed-up ratio of the GPU code is  two-orders of magnitude compared with the CPU counterpart."
Published on October 23, 2017 Updated on July 8, 2018