Are you running a desktop environment on your machine? (i.e., GNOME or KDE
or XFCE or something similar). The issue is that pmemd.cuda may not be the
only thing running on the GPU, especially if you have other graphical stuff
going at the same time.
You can kill the X-server (and get out of your desktop environment) and see
if that helps. There are instructions for killing the desktop environment
here: http://jswails.wikidot.com/installing-amber12-and-ambertools-12#toc17--
depending on your OS you may need to use "xdm" or "kdm" instead of
"gdm"
in that example.
HTH,
Jason
On Tue, May 8, 2012 at 1:01 PM, Albert <mailmd2011.gmail.com> wrote:
> HI Ross:
> thank you very much for your such detailed reply.
> I've got two GTX590 in a workstation and all jobs are running
> locally. I compared the same system with series and MPI running:
>
> single series running can up to 27.4 ns/day, but it is unstable and it
> can change to 13 or even 10 ns/day from time to time....
> the MPI running I use mpirun -np 2 +..... for running, and it is around
> 18 ns/day.
>
> I am not sure whether series running jobs on GTX590 can distribute to
> both logical cores. I only submit one job when I did each test.
>
> at present, it only works under GCC compiled Amber12 and it doesn't work
> for intel compiled version. However the CPU version works fine for both
> GCC and ifor/icc version.
>
> thank you very much
>
> best
> Albert
>
> On 05/08/2012 06:50 PM, Ross Walker wrote:
> > Hi Albert,
> >
> > The GTX590s are pretty terrible cards - at least for running jobs in
> > parallel so I really wouldn't bother. Only run serial jobs on them, two
> > total, one on each of the GPUs reported by devicequery.
> >
> > If you are seeing variation in performance for serial GPU runs then it is
> > likely that other stuff is running on your machine that is interfering.
> Are
> > other users on the machine? Do you have other jobs running? Does the
> machine
> > run X windows? - If you are running multiple jobs on the same physical
> GPU
> > (not card but GPU - a GTX590 card has 2 GPUs) then you will likely see
> > massive performance degredation.
> >
> > You might also have issues if you are writing to a remote file system
> (over
> > NFS for example) but this is normally not so large a variation. The
> factors
> > of '2' in the numbers you quote (27 ca= 28, 14, 7) suggests to me that
> you
> > are running more than one process on a single GPU.
> >
> > Unfortunately without more specific details of how your machine is setup,
> > what is running, some output from top etc it is difficult to offer more
> > suggestions.
> >
> > All the best
> > Ross
> >
> >> -----Original Message-----
> >> From: Albert [mailto:mailmd2011.gmail.com]
> >> Sent: Tuesday, May 08, 2012 1:36 AM
> >> To: AMBER Mailing List
> >> Subject: [AMBER] strange CUDA benchmark
> >>
> >> Dear:
> >>
> >> I am testing a CUDA benchemark based on Amber 12+GNU, I use the same
> >> system under GTX590 for testing. Some times I can get as high as 27
> >> ns/day, sometimes it 14 ns/day, sometimes it is even only 7 ns/day.....
> >>
> >> I do the test in the same directory and don't modify any parameters. I
> >> am just wondering what happen?
> >>
> >> thank you very much
> >> Albert
> >>
> >> _______________________________________________
> >> AMBER mailing list
> >> AMBER.ambermd.org
> >> http://lists.ambermd.org/mailman/listinfo/amber
> >
> > _______________________________________________
> > AMBER mailing list
> > AMBER.ambermd.org
> > http://lists.ambermd.org/mailman/listinfo/amber
>
>
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
>
--
Jason M. Swails
Quantum Theory Project,
University of Florida
Ph.D. Candidate
352-392-4032
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Tue May 08 2012 - 10:30:04 PDT
