[gpaw-users] Changes in the new PAW setups bundle of 26/10 ("gpaw-setups-0.9.9672")

[Ask Hjorth Larsen]

Hi

2012/11/23 Marcin Dulak <Marcin.Dulak at fysik.dtu.dk>:
> Hi,
>
>
> On 11/23/12 15:03, Ask Hjorth Larsen wrote:
>>
>> Hi
>>
>> 2012/11/23 Marcin Dulak <Marcin.Dulak at fysik.dtu.dk>:
>> (...)
>>>
>>> A negative outcome from the way new setups are generated is that most of
>>> them require now eigensolver='cg' or 'dav' to achieve convergence.
>>> The new Oxygen setup is also a bit harder, so requires finer grid.
>>>
>>> Best regards,
>>>
>>> Marcin
>>
>> This last part is rather unfortunate because cg and dav, compared to
>> rmm-diis, reduce the performance of GPAW by a large factor for systems
>> of non-trivial size.  Furthermore those solvers don't support band
>> parallelization AFAIK.  What can be done about this?
>
> i remember Jens Jørgen and your comments about that (old Ti setup case),
> can you remind what was your conclusion?
> The convergence problems seem to be related to the presence of semicore
> p-states:
> https://listserv.fysik.dtu.dk/pipermail/gpaw-developers/2012-September/003190.html

I don't think anything was concluded other than that it was nasty.

I'm not sure that I have ever seen something that could not be fixed
with rmm-diis by having a better initialization or adding more states.

Presumably you have some calculations from which it is clear that
neither of these helps.  But it'll be a major blow if cg is needed as
a default.  I think it's well below half the performance of rmm-diis
even when band parallelization is not needed.  So something should be
figured out. (Small molecules would of course be the exception because
time is spent on XC/Poisson.)

Regards
Ask