gromacs does not uses cuda

Christian Hübner · Post by **Christian Hübner** » Sat May 23, 2009 12:55 am

Dear Peter,

thank you for keeping up with support! Do I have to remove PR2 before installing PR3? What is actually meant with "Install OpenMM". I just unpacked the tar. Do I have to do this in a special place?
Regarding the size of the system: 1483 atoms, and VRAM: 256 MB. Do you know if I have to activate the CUDA kernel extension or do I just need the library? And how can I tell CUDA which of the two GPUs on my system to use?

Best

Christian

marco dabramo · Post by **marco dabramo** » Mon May 25, 2009 6:55 am

Hi,
I have the same problem: gromacs does not uses gpu (mdrun-openmm prints out OpenMM Platform: Reference).

I have followed the readme (Gromacs-OpenMM Readme.txt) step by step.
I have checked that the OpenMM lib and cuda libs are in LD_LIBRARY_PATH and that OPENMM_PLUGIN_DIR points to the OpenMM plugin directory (I'm using a linux machine).

Could you give some help?

Thanks in advance

Marco

Peter Eastman · Post by **Peter Eastman** » Mon May 25, 2009 8:14 am

Can you tell me exactly how you are setting LD_LIBRARY_PATH and OPENMM_PLUGIN_DIR? Also, what directories do you have OpenMM, Cuda, and Gromacs installed in?

Peter

marco dabramo · Post by **marco dabramo** » Mon May 25, 2009 8:36 am

Yes!
[Xalcantaj@sherwood11 marco]$ echo $OPENMM_PLUGIN_DIR
/home_nfs/Xalcantaj/soft/openmm/lib/plugins

[Xalcantaj@sherwood11 marco]$ ls /home_nfs/Xalcantaj/soft/openmm/lib/plugins
libOpenMMCuda.so

[Xalcantaj@sherwood11 marco] echo $LD_LIBRARY_PATH
/opt/cuda//lib:/opt/cuda/sdk/2.0//lib:/opt/cuda/sdk/2.0//common/lib/linux/:/home_nfs/Xalcantaj/soft/gromacs-4.0.5/lib:/home_nfs/Xalcantaj/soft/lib:/home_nfs/Xalcantaj/soft/fftw-64bits/lib:/home_nfs/Xalcantaj/soft/fftw-3.1.2-32bits/lib

Gromacs was installed in /home_nfs/Xalcantaj/soft/gromacs-4.0.5/
Cuda in /opt/cuda
openmm-pr3 in /home_nfs/Xalcantaj/soft/openmm-pr3

My path variable is:
[Xalcantaj@sherwood11 marco]$ echo $PATH
/opt/cuda//bin:/home_nfs/Xalcantaj/soft/gromacs-4.0.5/bin:/home_nfs/Xalcantaj/soft/bin:/home_nfs/Xalcantaj/soft/sbin:/usr/kerberos/bin:/bin:/usr/bin:/home_nfs/Xalcantaj/bin:/home_nfs/Xalcantaj/soft/openmm-pr3

Hope this helps

Thanks

Peter Eastman · Post by **Peter Eastman** » Mon May 25, 2009 1:15 pm

You don't have OpenMM in your library path. Add

/home_nfs/Xalcantaj/soft/openmm-pr3/lib

to LD_LIBRARY_PATH. What's puzzling is how you could run Gromacs at all without that. Make sure that you don't have a copy of an old OpenMM release hanging about somewhere, still it your path.

Also, can you successfully run the examples from the CUDA SDK with your path set the way it is?

Finally, note that the precompiled binaries we provide are for 32 bit Linux. Do you have 32 or 64 bit? (I noticed that you included both 32 and 64 bit versions of fftw in your path.)

Peter

marco dabramo · Post by **marco dabramo** » Tue May 26, 2009 3:39 am

Hi Peter,
thanks for your help.
However, I still have the same problem (I have added openmm in the LD_LIBRARY_PATH as you suggested).
We have a 64bit linux environment; is it possible to run the mdrun-openmm binary on 64bit machines?

Now, we are checking that CUDA SDK works...

Thank you

Peter Eastman · Post by **Peter Eastman** » Tue May 26, 2009 7:27 am

We've only tested on 32 bit Linux. In theory it ought to be possible to get it working on 64 bit, but I don't know for sure.

All of our precompiled binaries are 32 bit ones. I *think* that 32 bit binaries can theoretically be run on 64 bit Linux, but I don't have any direct experience. But either way, it's essential that all binaries you use together must be consistent with each other - CUDA, OpenMM, and Gromacs. If some of those are 32 bit and others are 64 bit, I definitely wouldn't expect that to work.

Peter

Christian Hübner · Post by **Christian Hübner** » Tue May 26, 2009 8:26 am

It works, finally! I just installed all from scratch. Testing the performance, I found that when I am using explicit solvent, CUDA is slower than my CPU, whilst it is faster for implicit solvent. Is this reasonable?

Best

Christian

Peter Eastman · Post by **Peter Eastman** » Tue May 26, 2009 9:59 am

I'm glad to hear you've got it working! What OS are you using?

The relative speed of CPU vs. GPU will depend on a lot of factors, including the speed of your GPU (you mentioned before that you have a 9600M GT, which is a fairly low end processor), the speed of your CPU (I don't think you said what that is?), and the system you're simulating. At one point you mentioned a system with 1483 atoms. Is that the one? If so, you'll probably get better performance on a larger system.

Peter

Christian Hübner · Post by **Christian Hübner** » Wed May 27, 2009 6:02 am

Dear Peter,

thanks again for your interest. I am using gromacs under Leopard on a macbook pro:
Modellname: MacBook Pro
Modell-Identifizierung: MacBookPro5,1
Prozessortyp: Intel Core 2 Duo
Prozessorgeschwindigkeit: 2.4 GHz
Anzahl der Prozessoren: 1
Gesamtzahl der Kerne: 2
L2-Cache: 3 MB
Speicher: 4 GB
Busgeschwindigkeit: 1.07 GHz

Two things are not clear to me: Can I only use the amber99 force field? Or only if i am using the OBC GBSA solvent model? In the latter case, what do I have to do with the stuff in the gbsaobc tarball? Or do I have to add just the mentioned lines in the mdp file?

Best

Christian

gromacs does not uses cuda

RE: gromacs does not uses cuda

RE: gromacs does not uses cuda

RE: gromacs does not uses cuda

RE: gromacs does not uses cuda

RE: gromacs does not uses cuda

RE: gromacs does not uses cuda

RE: gromacs does not uses cuda

RE: gromacs does not uses cuda

RE: gromacs does not uses cuda

RE: gromacs does not uses cuda