OpenFOAM Compatible CUDA Accelerated CFD-solver for Wind Engineering – Københavns Universitet

OpenFOAM Compatible CUDA Accelerated CFD-solver for Wind Engineering

Specialeforsvar i eScience ved Jonas Kastrup  


The Danish engineering firm Rambøll requested a fast OpenFOAMa` compatible fluid dynamic (CFD) solver for urban wind analysis.

This thesis examines the possible performance gain from implementing an OpenFOAMa` compatible fluid dynamic solver on GPU hardware. The implemented solver is based on the incompressible laminar Navier-Stokes equations solved with the PISO-algorithm. Therefore it is comparable with the icoFoam application implemented in OpenFOAMa` . When comparing single time step integrations results show a 11.7 times speed-up compared to the icoFoam application. A possible increase in performance of 20-30% has been justified which will bring the total speed up to 15.

The matrix layout used in the GPU implementation is very different from the one used in OpenFOAMa` . To fit the GPU matrix format the data fields are sorted at initialization of the GPU. Results have shown that sorting and copying of data from OpenFOAMa` to the GPU is in the order of 1% of the total time usage. The speed of the implemented GPU solver can potentially change the CFD engineering process from a production to a design oriented process.

When and where? October 31, at 10:00 in Meeting room A+B at Department of Computer Science, University of Copenhagen.

Supervisor: Kenny Erleben, DIKU.

External supervisor: Jens Christian Bennetsen, Rambøll.

Censor: Niels Jørgen Christensen, IMM DTU.