[gmx-developers] benchmark Lys errors?

[David]

On Tue, 2004-03-30 at 04:06, Josh Hursey wrote:

> >>>Yes, your protein has now more atoms than what fits on one processor.
> >>>Although it maybe shouldn't crash, you are already way beyond the
> >>>maximum performancewise. More than 4 processors is not worthwhile for
> >>>such small systems...
> >>
> >>Is there a rough metric for determing when one has allocated too many processors for a given system?
> >>'
> > 
> > Depends on the communication speed. For cut-off sims you need a few
> > thousand atoms per processor, for PME sims more than that.
> 
> 
> So when looking at a General Protein, what determines the scaling limit? I assume it involves an evaluation of the number of 
> atoms in the system and the type of calculation involved.
In the end it's merely the system size, but it depends on your
network/CPU combination as well. For typical systems, such as the Lys
benchmark it can easily be measured. Just run the cut-off and the PME
benchmarks on 1,2,4,6,8 processors and make an informed decision.

I have run a 200,000 atom system on 16 processors (with PME) but in the
end I moved it to 8 processors, which was somewaht slower but gave me
more throughput since I had a limited CPU time on that cluster.

As soon as there are more jobs that CPUs it is worthwhile to reduce the
number of CPUs per job.
-- 
David.
________________________________________________________________________
David van der Spoel, PhD, Assist. 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
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++