The ABINIT Users Mailing List ( CLOSED )

[Xavier Gonze <gonze@pcpm.ucl.ac.be>]

Dear Ron,

Apparently, you are testing the k point and spin parallelization, and  
you took
the file that was indeed suggested for this test, from the  ~abinit/ 
tests/paral/Input
directory. As mentioned in the ~abinit/tests/paral/README file, see  
the end of this file,
this is the only test for a proper speed testing of this  
parallelization, for a number
of processor bigger than 10 . Still, the maximum number of processor  
that
you can really use with the proposed file is 364, as also mentioned  
in the README file.
So there is no surprise that you only get 30% improvement going from  
256 to 512
processors.

The same file contains suggestions to test larger number of  
processors. You might
try to modify  the value of  ngkpt in ~abinit/tests/paral/Input/t_kpt 
+spin.in ,
i.e. set it to  16 16 16 . This would allow a maximal use of  
512*2=1024 processors.
But this is a bit less realistic, as with a 12 12 12 k point grid,  
the physical quantities
of interest are converged.
Other suggestions are proposed in the README file.

On the other hand, there is another parallelization, that has been  
made available
in ABINITv5.2, and that will be further improved in ABINITv5.3 . This  
parallelization
is much more communication-intensive than the k point and spin  
parallelization, but
applicable to systems where the number of k points is one (or only a  
few). A speed-up
of about 300 was recently achieved with this parallelization, on 384  
processors.
You might look in the ABINIT forum archives, or directly contact the  
person developing
that parallelization, François Bottin.

Best wishes,
Xavier Gonze

On 08 Dec 2006, at 13:58, Ron Lieberman wrote:

> Hi
>
> I have a cluster with up to 512 cores available to me.
> I would like to find a dataset for abinit that scales at 128,256,512
> processes.
> The one I am using now seems to scale at 70% improvement with each  
> doubeling
> of processors
>  upto about 256. From 256 to 512, it acheives a 30% improvement.
>
> Would anyone have a dataset they would like to share that might  
> offer more
> scalability?
> Or do you think its an inherent limit of the abinit application?
>
>
> The dataset I am currently using has these files
> 26fe.pspnc  ab.files  t_kpt+spin.in
>
> And the first part of 26fe... Looks like:
>  Troullier-Martins psp for element  Fe        Thu Oct 27 17:35:05  
> EDT 1994
>   26.00000   8.00000    940714                zatom, zion, pspdat
>     1    1    2    0      2001    .00000
> pspcod,pspxc,lmax,lloc,mmax,r2well
>     0   4.333  10.868    0   2.2918558         
> l,e99.0,e99.9,nproj,rcpsp
>     .00000000    .0000000000    .0000000000    .00000000
> rms,ekb1,ekb2,epsatm
>     1   1.213   4.197    1   2.8345121         
> l,e99.0,e99.9,nproj,rcpsp
>     .00000000    .0000000000    .0000000000    .00000000
> rms,ekb1,ekb2,epsatm
>     2  18.664  23.972    1   2.2918558         
> l,e99.0,e99.9,nproj,rcpsp
>     .00000000    .0000000000    .0000000000    .00000000
> rms,ekb1,ekb2,epsatm
>     1.56404770202776    2.06158206779471    6.88331421535388
> rchrg,fchrg,qchrg
>     0 =l for Troullier-Martins pseudopotential
>
> Thanks you in advance for any advice.
>
> Ron Lieberman
> HP-MPI    972-497-4248