[gpaw-users] Scalapack problem

[Jens Jørgen Mortensen]

On 21-12-2011 15:52, Jussi Enkovaara wrote:
> On 2011-12-21 16:36, Jens Jørgen Mortensen wrote:
>> On 21-12-2011 11:46, Jussi Enkovaara wrote:
>>> On 2011-12-21 10:28, Jens Jørgen Mortensen wrote:
>>>> Hi!
>>>>
>>>> The attached GPAW text output files show the results of two
>>>> calculations.  One with a BLACS 4x2 grid and 16x16 blocksize and another
>>>> one with 2x4 grid and 32x32 blocksize.  The first one converges nicely,
>>>> but the second one does not.  Is this a bug or am I doing something wrong?
>>> Hi,
>>> you seem to have different number of bands in the two calculations (700 vs. 651),
>>> could this explain the difference?
>> No, 4x2+16x16 works for both 700 and 651 bands and 2x4+64x64 fails for
>> both 700 and 651 bands.
>>
> OK, then it could very well be a bug...
> Could you share your input script,

I used GPAW's command line tool:

   gpaw bulk -k 4,4,1 -p "mode=lcao,basis=dzp,xc=PBE,nbands=700" c8.traj 
--domain-decomposition=8 --sl_default=2,4,32

You can just use one of the text files I attached to the first email 
instead of c8.traj.

Or something like this (not tested):

from ase.io import read
from gpaw import GPAW
c8 = read('c8.traj-bulk-gpaw-d8s4216.txt')
c8.calc = GPAW(mode='lcao', basis='dzp', xc='PBE', nbands=700,
     parallel={'domain': 8, 'sl_default': (2, 4, 32)})
c8.get_potential_energy()

>   I could check whether I can reproduce the error
> with different ScaLAPACK implementation? (Even thought I might have time only in
> January). Have you compared to serial LAPACK, does 4x2+16x16 give the same numbers
> as LAPACK?

Yes, and I got the same numbers.

Jens Jørgen

> Best regards,
> Jussi