[Pw_forum] Trying to find reasonable parallelization parameters

[Axel Kohlmeyer]

On Thu, 22 Jan 2009, J. J. Ramsey wrote:

JJR> ----- Original Message ----
JJR> 
JJR> > From: Axel Kohlmeyer <akohlmey at cmm.chem.upenn.edu>
JJR> > To: J. J. Ramsey <jjr19 at uakron.edu>
JJR> > Cc: PWSCF <pw_forum at pwscf.org>
JJR> > Sent: Thursday, January 22, 2009 1:37:00 PM
JJR> > Subject: Re: [Pw_forum] Trying to find reasonable parallelization parameters
JJR> 
JJR> > very simple. -npools can be maximally the actual number of k-points that 
JJR> > you are using, and that should be exploited to the maximum. a second 
JJR> > option to pay attention to, is the disk_io flag. setting this to 'low' 
JJR> > can have a very positive effect on scaling at the expense of additional 
JJR> > memory use.
JJR> 

JJR> Well, I'm trying to *reduce* memory usage per processor, so I'm 
JJR> leery of lowering the disk-IO. What I'd like to do is try to 
JJR> distribute the memory for the job as much as possible.

perhaps i didn't explain it too well, but reducing disk i/o
will _significantly_ enhance scaling and that would mean
that you can use many more processors for the g-space 
parallelization, which in turn will significantly reduce
the per MPI-task memory requirements. basically, you want 
to get the most performance out of your machine, right?
and this is how you get it.

another option that frequently helps, particularly with 
current intel quad-core cpus, is to use only half the available 
cpus. due to the architecture of that intel platform, this 
reduces the load on the memory bus and effectively doubles 
the cpu cache (and per MPI-task memory, of course). the effect 
can be quite drastoc, as shown, e.g. in this scaling test 
on our local machine (using a different DFT code that is
much more demanding in terms of memory bandwidth than Q-E).

http://cp2k.googlegroups.com/web/cp2k-water-bench-cmm.png

so less is more in this case, and the only way to find this
out is to run systematic tests (this example is gamma only).

cheers,
   axel.


JJR> 
JJR> 
JJR> 
JJR>       
JJR> _______________________________________________
JJR> Pw_forum mailing list
JJR> Pw_forum at pwscf.org
JJR> http://www.democritos.it/mailman/listinfo/pw_forum
JJR> 

-- 
=======================================================================
Axel Kohlmeyer   akohlmey at cmm.chem.upenn.edu   http://www.cmm.upenn.edu
   Center for Molecular Modeling   --   University of Pennsylvania
Department of Chemistry, 231 S.34th Street, Philadelphia, PA 19104-6323
tel: 1-215-898-1582,  fax: 1-215-573-6233,  office-tel: 1-215-898-5425
=======================================================================
If you make something idiot-proof, the universe creates a better idiot.