[gmx-users] PME scaling

[Carsten Kutzner]

Hi Rob,

a 72 k atom system should definitely scale better on GigE. At least it
should not be slower on 6 CPUs compared to 4. What kind of switch are
you using? Please look at the network settings and make shure that both
the switch and the network cards have flow control enabled.

Carsten



On Tue, 10 Jan 2006, Robert Bjornson wrote:

> Hi David,
>
> Sure, thanks for your interest, here is a bit more info:
>
> I'm running gromacs 3.3, with patch to pme.c (using 1.61.2.5 to fix
> pme-order bug)
> I'm using lam 7.1.1 under pbspro, running on a cluster of 3.2 Ghz
> EMT-64 xeons.  The interconnect is gigE.
>
> The model is 72k atoms, using PME with pme_order=6
>
> Here is the performance I'm seeing (I've rounded numbers):
>
> cpus   steps/hour   steps/hour/cpu        typical node cpu utilization
> (remember 2 cpus/node)
> 8           6200              775                   .5-.9
> 6           8500              1400                  .75-1.08
> 4           25000            6250                  2.0
> 1           8870              8870                  1.0
>
> If I run the same model with cutoff instead of PME, I see:
> 8            82000            10274             1.6-1.8
> 1            24000            24000              1.0
>
>
> So, the performance I'm seeing on more than 4 nodes with PME is pretty
> bad.  How does this compare to what you expected?  I can provide you
> with more info if you'd like.  Thanks very much for any insight you
> might have.
>
> Sincerely,
>
> Rob Bjornson
>
>
> On 1/10/06, David van der Spoel <spoel at xray.bmc.uu.se> wrote:
> > Robert Bjornson wrote:
> > > Hi,
> > >
> > > I'm experiencing very poor scaling when using PME on gromacs-3.3, and
> > > looking through the list indicates that this is a known issue with
> > > that release.  However, there was some indication that work has been
> > > done on parallelizing PME, and that the top of CVS might contain a
> > > version that is worth trying.
> >
> > In addition to Erik's answer: can you be more specific? We've seen quite
> > decent scaling on up to 16 processors, but depending strongly on the
> > size of the system and the interconnect.
> > >
> > > Has anyone tried this?  Did your PME performance improve?  If so, did
> > > you simply take the top of cvs, or is there a tag that is more likely
> > > to work successfully?
> > >
> > > thanks,
> > >
> > > Rob Bjornson
> > > _______________________________________________
> > > gmx-users mailing list
> > > gmx-users at gromacs.org
> > > http://www.gromacs.org/mailman/listinfo/gmx-users
> > > Please don't post (un)subscribe requests to the list. Use the
> > > www interface or send it to gmx-users-request at gromacs.org.
> >
> >
> > --
> > David.
> > ________________________________________________________________________
> > David van der Spoel, PhD, Assoc. Prof., Molecular Biophysics group,
> > Dept. of Cell and Molecular Biology, Uppsala University.
> > Husargatan 3, Box 596,          75124 Uppsala, Sweden
> > phone:  46 18 471 4205          fax: 46 18 511 755
> > spoel at xray.bmc.uu.se    spoel at gromacs.org   http://xray.bmc.uu.se/~spoel
> > ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
> >
> _______________________________________________
> gmx-users mailing list
> gmx-users at gromacs.org
> http://www.gromacs.org/mailman/listinfo/gmx-users
> Please don't post (un)subscribe requests to the list. Use the
> www interface or send it to gmx-users-request at gromacs.org.
>


---------------------------------------------------
Dr. Carsten Kutzner
Max Planck Institute for Biophysical Chemistry
Theoretical and Computational Biophysics Department
Am Fassberg 11
37077 Goettingen, Germany
Tel. +49-551-2012313, Fax: +49-551-2012302
eMail ckutzne at gwdg.de
http://www.gwdg.de/~ckutzne