[lionelr]

Dear all,

I'm working on optimizing OpenFOAM execution speed using GPU. For now I use Paralution (see http://www.paralution.com/) and it works fine. The OpenFOAM plugin example they provide uses a Laplacian (fvm::laplacian(DT, T)) algorithm and allows a speedup up to 7x comparing to the CPU only execution time on our testbed.

When I test Paralution with the SIMPLE algorithm on a backward facing step test case, we don't have any speedup comparing to the CPU execution.

I suspect an I/O overhead with the SIMPLE algorithm. In both cases, the solver is a PCG.

My questions are :
(1) Have you some explanations of that phenomena ?
(2) Have you some tips on how to get ride of this problem ?

Thanks and best regards,
Lionel

[luckycfd]

Hi,
I compiled paralution-0.5.0 correctly and my openfoam is 2.1.x.
To test the solver, I use the OpenFOAM plugin example, Laplace. but I don't have any speedup by openmp backend!
I attached the my laplace folder, please guide me how can I configure this example for getting speed up bye openmp threads?
I set number of openmp threads in the shell by:

export OMP_NUM_THREADS=2

is it correct for setting the number of openmp threads for paralution?
my system have 8 x six-core (NUMA) = 48 core no hyper threading.

regards
Hassan