Why not SU2 use OpenMP or MKL?

codezy · April 2, 2014, 06:16

Hello guys,
I'd like to try SU2 with OpenMP and MKL. I try changing the Gauss Elimination to MKL dgesv in LAPACK. It becoms slow. ToT
Could somebody know about why SU2 didn't use them?
And how could I run the SU2 on the GPU with CUDA?

Thanks a lot.

economon · April 3, 2014, 20:13

Hi,

Your question is well timed. It turns out that we are starting to tackle exactly this problem: we will be reimplementing many of the high-performance aspects of the code to take advantage of OpenMP (with edge coloring), linear solvers and preconditioners that exhibit better scalability, and new MPI approaches for distributed memory calculations.

Please stay tuned for more updates on this front,
Tom

p.s. at the moment, we do not have a CUDA implementation, but this is also in consideration.

April 2, 2014, 06:16	Why not SU2 use OpenMP or MKL?	#1
codezy New Member zhen yi Join Date: Mar 2014 Posts: 16 Rep Power: 12	Hello guys, I'd like to try SU2 with OpenMP and MKL. I try changing the Gauss Elimination to MKL dgesv in LAPACK. It becoms slow. ToT Could somebody know about why SU2 didn't use them? And how could I run the SU2 on the GPU with CUDA? Thanks a lot.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Official release of SU2 V3.0 and SU2 Educational V1.0	fpalacios	SU2 News & Announcements	2	January 22, 2014 05:28
Incorrect calculation of forces/moments coefficients with older versions of SU2	diwakaranant	SU2	2	August 16, 2013 04:17
SU2 suite has moved to GitHub!	fpalacios	SU2 News & Announcements	0	August 12, 2013 02:07
Welcome to the Stanford University Unstructured (SU2) forum!	economon	SU2	0	January 7, 2013 02:48
New SU2 Forum Opened at CFD Online	pete	Site News & Announcements	0	January 5, 2013 17:59

April 3, 2014, 20:13		#2
economon Super Moderator Thomas D. Economon Join Date: Jan 2013 Location: Stanford, CA Posts: 271 Rep Power: 14	Hi, Your question is well timed. It turns out that we are starting to tackle exactly this problem: we will be reimplementing many of the high-performance aspects of the code to take advantage of OpenMP (with edge coloring), linear solvers and preconditioners that exhibit better scalability, and new MPI approaches for distributed memory calculations. Please stay tuned for more updates on this front, Tom p.s. at the moment, we do not have a CUDA implementation, but this is also in consideration.