[gpaw-users] k-point parallelisation

[Ask Hjorth Larsen]

Hello

(Re-adding list to CC)

2013/10/5 Alexander Hupfer <sonium at gmail.com>:
> Ok, it works if number of CPUs is a integer divisor of the number of 1st
> Brillouin-zone k-points.
> Domain decomposition seems to be not supported with plane waves.
> Since band parallelisation only seems seems to be effective at 1000 bands or
> more I think the calculations might take a while, or is there any more
> parallelisation once can engage?
>
> I can still do the dry run if you want.

It's probably written somewhere that band parallelization in FD mode
begins to be relevant at 1000 bands as a competitor to domain
decomposition (actually I think there are cases where it becomes a
good idea already before then).  But that is not true for PW mode
where its status as the only non-kpt/spìn parallelization mode, I'd
wager, makes it quite worthwhile even with much fewer bands.  So do
try band parallelization.  And don't hesitate to run a performance
test or two as this will depend on your supercomputer.

There are no other parallelization options to consider unless your
system is much larger.

Note that in GPAW trunk you can now do parallel={'kpt': N} to
explicitly specify a certain level of k-point parallelization.

Regards
Ask

>
> Regards,
> Alex
>
> On October 5, 2013 at 8:25:13 PM, Ask Hjorth Larsen (asklarsen at gmail.com)
> wrote:
>
> Hello
>
> It's the number of k-points in the 1st Brillouin-zone which counts.
> 6x6x6 is probably not the final number of IBZ kpts. Please attach the
> logfile from a --dry-run or anything listing the number of IBZ kpts,
> and the script while we are at it.
>
> Regards
> Ask
>
> 2013/10/5 Alexander Hupfer <sonium at gmail.com>:
>> Hi,
>> I have same trouble with k-points parallelisation. As I understood from a
>> earlier post there must be at least one k-point per CPU. However I
>> experience "load imbalance" problems with much lower number of CPU nodes
>> than this. I use the standard parallelisation options give by:
>>
>> parallel = {'domain': None,
>> 'band': 1,
>> 'stridebands': False,
>> 'sl_auto': False,
>> 'sl_default': None,
>> 'sl_diagonalize': None,
>> 'sl_inverse_cholesky': None,
>> 'sl_lcao': None,
>> 'buffer_size': None}
>>
>> for example a 6x6x6 kpoint calculation works on 27 nodes, but crashes on
>> 54
>> nodes. What can I do about this?
>>
>> - Alex
>>
>> _______________________________________________
>> gpaw-users mailing list
>> gpaw-users at listserv.fysik.dtu.dk
>> https://listserv.fysik.dtu.dk/mailman/listinfo/gpaw-users