[Pw_forum] Efficient parallelization

[Axel Kohlmeyer]

On Wed, 6 Apr 2005, Sergey Lisenkov wrote:

SL> Dear Axel,

dear sergey,
 
SL> >looks perfectly. 6 times 4 is 24. the parallelization is first
SL> >across the k-points, since those calculations are almost independent
SL> >AND THEN for EACH k-point across g/r-space. the latter parallelization
SL> >is less efficient, so the choice of the -npool parameters is very
SL> >important for efficient use of your cpu resources (if you have
SL> >multiple k-points). this is explained in the manual in more detail.
SL> 
SL> If I run the code on SMP machine, does k-parallelization efficient?

it should be. there may be a problem with memory requirements.
if you run several k-points in parallel, you need to be able to store
those wavefunctions in memory simulaneously, whereas with g-/r-space
parallelization the total amount of memory should not increase that
much when increasing the number of processors. i was referring to
the efficiency of the algrorithm itself. there are always technical
issues to be considered. under some cicumstances i found, that
it is occasionally better to not use all cpus to have more memory 
bandwidth per processor (e.g. on dual Xeon machines).

best regards,
	axel

SL> 
SL>  Best wishes,
SL>    Sergey
SL> 
SL> 

-- 

=======================================================================
Dr. Axel Kohlmeyer   e-mail: axel.kohlmeyer at theochem.ruhr-uni-bochum.de
Lehrstuhl fuer Theoretische Chemie          Phone: ++49 (0)234/32-26673
Ruhr-Universitaet Bochum - NC 03/53         Fax:   ++49 (0)234/32-14045
D-44780 Bochum  http://www.theochem.ruhr-uni-bochum.de/~axel.kohlmeyer/
=======================================================================
If you make something idiot-proof, the universe creates a better idiot.