[Q-e-developers] [Pw_forum] Increasing the FFT grid spacing/ using wavefunction to provide initial charge density

Carlo Cavazzoni c.cavazzoni at cineca.it
Fri Sep 30 11:40:23 CEST 2016


regarding this question I don't have a clear idea yet, even if, in 
general, I see in other fields
that successive refinements are a quick and simple way to precondition 
the solution of a set of equations...

BUT, I would like to point out that the recent revision of the FFT 
(included in QE6.0), allow to instantiate and allocate as many
FFT grids with different cut-off as you want, being all grids coherent 
on the data distribution (in g-space).
As long as I can understand, this feature, should simplify a lot 
implementation of this kind of algorithms

carlo


Il 30/09/2016 11:22, Nicola Marzari ha scritto:
> On 29/09/2016 22:45, Paolo Giannozzi wrote:
>> In my opinion it is neither simple not very useful. What might be useful
>> in my opinion is the possibility to
>> 1. restart from lower-cutoff charge density AND wavefunctions, and
>> 2. interpolate wavefunctions to a denser grid of k-points, adding the
>> missing bands if needed.
>> The charge density alone often typically doesn't make a big difference.
>> Moreover the charge density, now written in real space, should be
>> written in G-space, like wavefunctions, so that one can use a similar
>> logic (or illogic).
>>
>> Paolo
>
> Ciao Paolo,
>
>
> one of the few times I beg to disagree, but mind you that I was thinking
> at a case where you do a lot of ionic steps to relax a geometry (a
> fairly common
> case in large systems where you eg put a flopyy molecule on a surface or a
> nanostructure). You could easily have ~10-20 ionic
> iterations, i.e. 50-100 electronic loops. Many pseudopotentials
> do a decent jobs at lower cutoffs (and especially at low fft cutoffs,
> e.g. a dual of 2-3 for norm-conserving, and a dual of 4-6 for ultrasoft),
> so a lot of calculations could be done rapidly initially, with some final
> refinement. (to be honest, I used to do this all the time 25 years ago or
> so, with castep)
>
> We'd be very happy to test this before Jan on 1-2 systems, if you need
> more justification.
>
> Of course one would need to do the fft interpolation - i.e. go through
> the charge density in g-space, but that can be done independently from
> the way the charge density is saved (there is an argument to have it in real
> space to facilitate post-processing. Of course, adding it in reciprocal
> sapce would only add a minor cost in space, and so maybe one should save
> both?)
>
> The more we look at the performance of QE and the more we realize
> that "other codes" are faster because the pseudopotentials are significantly
> softer, and the parameters quite loose. We could easily explore an approach
> in which in a relaxation we start loose and we perfect only at the very end.
>
> Of course, this is most relevant for those calculations where there are
> a lot
> of ionic steps to get to the minimum.
>
> 				nicola
>
>
> ----------------------------------------------------------------------
> Prof Nicola Marzari, Chair of Theory and Simulation of Materials, EPFL
> Director, National Centre for Competence in Research NCCR MARVEL, EPFL
> http://theossrv1.epfl.ch/Main/Contact http://nccr-marvel.ch/en/project
> _______________________________________________
> Q-e-developers mailing list
> Q-e-developers at qe-forge.org
> http://qe-forge.org/mailman/listinfo/q-e-developers


-- 
Ph.D. Carlo Cavazzoni
SuperComputing Applications and Innovation Department
CINECA - Via Magnanelli 6/3, 40033 Casalecchio di Reno (Bologna)
Tel: +39 051 6171411  Fax: +39 051 6132198
www.cineca.it




More information about the developers mailing list