[CP2K:3216] real space grid decomposition

Laino Teodoro teodor... at gmail.com
Sat Apr 23 09:12:55 CEST 2011


It's only pw_grids which are distributed like that.

To get a better load balance one should move from pw_grids to rs_grids.

Teo

---------------------------------------------
Teodoro Laino
Zurich Switzerland

Contact info:
Tel.:     http://www.jajah.com/Teo
E-mail: teo... at laino.eu
             teodor... at gmail.com
---------------------------------------------

On 22 Apr 2011, at 22:47, Noam Bernstein <no... at bollweevil.gdbg.org> wrote:

> Am I correct in surmising that real space grids are decomposed over nodes only
> along the 1st dimension?
> 
> I'm reaching a parallelization bottleneck for QM/MM.  If I do atom
> based parallelization,
> I get pw_restrict_s3 and corresponding routines that take a lot of
> time and don't
> speed up.  If I do grid based parallelization I get no speedup once
> the number of
> parallel processes is comparable to the size of the 1st dimension of the real
> space grid (-18..18 = 39 in my case).  With 32 processes I get some
> parallelization
> but poor load balance, because some processes do grid x value and some do 2.
> For 64 processes some processes are actually doing no work in
> qmmm_forces_with_gaussian_LG(), for example
> 
> Is there anything I'm missing?
> 
>       thanks,
>       Noam
> 
> -- 
> You received this message because you are subscribed to the Google Groups "cp2k" group.
> To post to this group, send email to cp... at googlegroups.com.
> To unsubscribe from this group, send email to cp2k+uns... at googlegroups.com.
> For more options, visit this group at http://groups.google.com/group/cp2k?hl=en.
> 



More information about the CP2K-user mailing list