On Thu, Nov 29, 2012 at 1:43 AM, Fabian Glaser <fglaser.technion.ac.il>wrote:
> Hi David,
>
> Thanks a lot for your answer.
>
> mpirun -np 84 sander ....
>
> Runs perfectly OK (although from what you say seems not very effective)
> produces output etc. I just thought it runs very slowly, is 1/2 ns x day a
> a good or a bad pace?
>
1/2 ns x day is the speed you get running in serial. What Dave said was
that you are running 84 identical serial jobs (i.e., they are all doing the
*exact* same thing, and overwriting each others' files, which could cause a
'weird' error at some point if they intersect with each other badly).
mpirun should only be used with MPI-enabled programs, like sander.MPI.
> On the other hand tests I have done with sander.MPI job looks like it is
> running, but it does not produce any output, so I tried your test
> suggestion and .... it worked.
>
> So what is the right way to do it?
> I need to fill the following information in my script:
>
> #PBS -l select=4:ncpus=12:mpiprocs=12
>
Try fewer processors. Unless you're using replica exchange (which you're
not if sander works fine), then 84 processors is way too many for sander.
Try fewer (try, for instance, only a single node), to see if that works.
So something like:
#PBS -l select=1:ncpus=12:mpiprocs=12
And use the $PBS_NODEFILE that PBS provides to you in your mpirun
execution. Depending on the version of MPI you have, it would look
something like this:
mpirun -hostfile $PBS_NODEFILE sander.MPI
(you can use mpirun --help on your machine and look for the flag that
allows you to specify a machine file or host file or something related to
that). This will automatically launch the 'right' number of threads
exactly where they should be launched.
FWIW, I second Dan's suggestions, although the benchmarks he links to are
only valid for pmemd (and I *strongly* suggest using pmemd if your
application permits, since it provides exactly the same results as sander,
but is more efficient).
Good luck,
Jason
--
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 Thu Nov 29 2012 - 04:00:02 PST
