[gpaw-users] Stress tensor calculations (optimizations) in PW mode fail if run with parallellization over states, "Atoms objects on different processors are not identical"
Rasmus Karlsson
rasmusk at kth.se
Tue Mar 12 20:27:13 CET 2013
Hello Jens,
after applying your patch, I have been able to run a calculation on 40
processors for 10 hours (and several optimization steps) without having
the same crash. I'm currently running two stress tensor calculation jobs
on more processors than that without problems, after several
optimization steps.
It seems the code fixed the problem!
Thanks,
Rasmus
On 03/11/2013 11:27 AM, Jens Jørgen Mortensen wrote:
> Den 07-03-2013 16:02, Rasmus Karlsson skrev
>> Hi!
>>
>> I have noticed that my stress tensor optimizations (using
>> UnitCellFilter with PW mode) fail if I run them with parallelization
>> over states. The first step of the optimization succeeds, but when
>> the next step is going to start, GPAW fails with the message "Atoms
>> objects on different processors are not identical".
>>
>> I found some messages from this list that suggested that it might be
>> related to the MKL library, which I am using. Does anyone have any
>> suggestions for what I could do to prevent GPAW from failing in this
>> situation?
>>
>> I am using the trunk version of GPAW and ASE, and I'm running on
>> Abisko ( http://www.hpc2n.umu.se/resources/abisko).
>
> Could you try this patch:
>
> Index: gpaw/gpaw/stress.py
> ===================================================================
> --- gpaw/gpaw/stress.py (revision 10042)
> +++ gpaw/gpaw/stress.py (working copy)
> @@ -70,6 +70,7 @@
> np.dot(U_cc, cell_cv)).T
> sigma_vv += np.dot(np.dot(M_vv.T, s_vv), M_vv)
> sigma_vv /= len(wfs.symmetry.op_scc)
> + wfs.world.broadcast(sigma_vv, 0)
>
> calc.text('Stress tensor:')
> for sigma_v in sigma_vv:
>
> and tell us if that helps?
>
> Jens Jørgen
>
>>
>> Thanks,
>> Rasmus Karlsson
>>
>>
>> _______________________________________________
>> gpaw-users mailing list
>> gpaw-users at listserv.fysik.dtu.dk
>> https://listserv.fysik.dtu.dk/mailman/listinfo/gpaw-users
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://listserv.fysik.dtu.dk/pipermail/gpaw-users/attachments/20130312/907c4a51/attachment.html
More information about the gpaw-users
mailing list