On the Comparison of OpenFOAM on CPU & GPU

deniz · August 19, 2015, 11:01

Dear All,

Here is a beginner's question.

I see some members have tested the performance of OpenFOAM when it is run on GPU. In addition to this there are some publications on the open literature focused on similar topics.

When people state the linear solver that they implement for their test case (let's say PCG) does this mean PCG solver is used for all of the equations to be solved (velocity, pressure, turbulence etc)? So they use only one linear solver to see it's performance?

In connection to this, is there any guideline to select appropriate solvers, preconditioners and smoothers? How do we select which linear solver to be used for a specific problem?

I have tried to read some papers and wiki'ed, but these are the dark spots in my mind. Would be more than happy if someone clarified.

Thank you all,

Deniz

Lukasz · August 24, 2015, 04:25

Hi Deniz,

we decied to accelerate only pressure equation with GPU-accelerated PCG, since it was responsible for 90% of sparse algebra in simple/piso. In other solvers this ratio will be different. Nevertheless, too much code (ca. 50%) is running on CPU anyway, therefore acceleration of sparse algebra does not bring much.

From Ahmdal's Law you will immediately see that more code should be running on GPU to make GPU-acceleration reasonable. This is why we decided to implement simple/piso completely on GPU. The results, as expected, were much much better.

August 19, 2015, 11:01	On the Comparison of OpenFOAM on CPU & GPU	#1
deniz New Member Join Date: Jul 2012 Posts: 2 Rep Power: 0	Dear All, Here is a beginner's question. I see some members have tested the performance of OpenFOAM when it is run on GPU. In addition to this there are some publications on the open literature focused on similar topics. When people state the linear solver that they implement for their test case (let's say PCG) does this mean PCG solver is used for all of the equations to be solved (velocity, pressure, turbulence etc)? So they use only one linear solver to see it's performance? In connection to this, is there any guideline to select appropriate solvers, preconditioners and smoothers? How do we select which linear solver to be used for a specific problem? I have tried to read some papers and wiki'ed, but these are the dark spots in my mind. Would be more than happy if someone clarified. Thank you all, Deniz

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Frequently Asked Questions about Installing OpenFOAM	wyldckat	OpenFOAM Installation	3	November 14, 2023 12:58
RapidCFD - OpenFOAM running on GPU	newoscar	OpenFOAM Announcements from Other Sources	7	November 2, 2022 00:51
[Resolved] GPU on Fluent	Daveo643	FLUENT	4	March 7, 2018 09:02
Suggestion for a new sub-forum at OpenFOAM's Forum	wyldckat	Site Help, Feedback & Discussions	20	October 28, 2014 10:04
Cross-compiling OpenFOAM 1.7.0 on Linux for Windows 32 and 64bits with Mingw-w64	wyldckat	OpenFOAM Announcements from Other Sources	3	September 8, 2010 07:25

August 24, 2015, 04:25		#2
Lukasz Member Lukasz Miroslaw Join Date: Dec 2009 Location: Poland Posts: 66 Rep Power: 16	Hi Deniz, we decied to accelerate only pressure equation with GPU-accelerated PCG, since it was responsible for 90% of sparse algebra in simple/piso. In other solvers this ratio will be different. Nevertheless, too much code (ca. 50%) is running on CPU anyway, therefore acceleration of sparse algebra does not bring much. From Ahmdal's Law you will immediately see that more code should be running on GPU to make GPU-acceleration reasonable. This is why we decided to implement simple/piso completely on GPU. The results, as expected, were much much better.