[Pw_forum] Trying to find reasonable parallelization parameters

[Axel Kohlmeyer]

On Fri, 23 Jan 2009, Lorenzo Paulatto wrote:

LP> On Thu, 22 Jan 2009 19:37:00 +0100, Axel Kohlmeyer
LP> <akohlmey at cmm.chem.upenn.edu> wrote:
LP> >very simple. -npools can be maximally the actual number of k-points that
LP> >you are using, and that should be exploited to the maximum.
 
LP> True, but the maximum may not be the number of kpoints; using the parallel
LP> davidson algorithm (default) the hamiltonian matrix will be distributed and
LP> diagonalized on the larger perfect square number of processor *inside* each

lorenzo,

please note, that my first and most important statement was, that
the only reliable source of what is the best solution, is to make
benchmarks and find it out. as i was mentioning before when using 
multi-core processors theses days, choices can be quite confusing,
i.e. you can run much faster by not using a significant amount of
your compute power.

there are many parameters that affect the performance of a job
and they are different from the theoretical performance of an 
algorithm.

LP> pool. If you are low on memory you may need to spread the hamiltonian over
LP> several processors in order to fit in. I.e. for a 3000 electrons job I'm
LP> trying these days, I have to use at least 100 procs per pool (and set
LP> diago_david_ndim=2) to stay (well) under ~2GB.

well, one could also see this as a manifestation that you have
to beg to get stefano to buy you a larger machine. ;-)

not an easy task these days in italy, though. :-(

with all that in mind, i still maintain that most people should
pay more attention to maximizing use of -npools, as this is not
chosen automatically, and thus quite often people will run with
much reduced efficiency. this is particularly true for the majority
of calculations that are smaller than yours and run on smaller 
machines, quite often only with GigE interconnect, which is saturated
very fast with only g-space parallelization.

with best regards,
    axel.

LP> cheers
LP> 
LP> 

-- 
=======================================================================
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.