[QE-users] ?==?utf-8?q? ?==?utf-8?q? ?= Same run not accelerated when starting from converged rho and wf

Mon Aug 17 09:15:33 CEST 2020

Dear Stephano,
I reduced diago_thr_init to 1.0D-13 but it changed nothing to the first estimated scf accuracy.:

   Initial potential from superposition of free atoms
     a scf correction to at. rho is read from./calcforces_QE/step0000/initialization/initial_bassin/relax//results/dbed.relax.save/rho.in
     negative rho (up, down):  1.432E-02 0.000E+00
     Starting wfcs from file
     Checking if some PAW data can be deallocated...
       PAW data deallocated on    4 nodes for type:  1
       PAW data deallocated on   32 nodes for type:  2
       PAW data deallocated on   35 nodes for type:  3
     total cpu time spent up to now is       29.8 secs
     Self-consistent Calculation
     iteration #  1     ecut=    58.00 Ry     beta= 0.10
     Davidson diagonalization with overlap
     c_bands:  5 eigenvalues not converged
     ethr =  1.00E-13,  avg # of iterations = 40.0
     negative rho (up, down):  1.432E-02 0.000E+00
     total cpu time spent up to now is      813.3 secs
     total energy              =  -11390.75602693 Ry
     Harris-Foulkes estimate   =  -11390.75013390 Ry
     estimated scf accuracy    <       0.00021610 Ry
     iteration #  2     ecut=    58.00 Ry     beta= 0.10
     Davidson diagonalization with overlap
     ethr =  2.16E-08,  avg # of iterations =  1.0
... and the rest of the file is the same as the qe.0002.out I send.

Please note that I start from the wfcs of exactely the same configuration and parameters by using startingwfc='file',
but I do not use startingpot='file'.
Instead I have activated input_drho='rho.in'
This 'rho.in' file is a copy of the file 'rho.out' that have been created by activating the parameter input_drho='rho.out'.
This trick is done because sometimes, the atomic positions have a small variation (but not in this case).

Please note that I already tried this trick on smaller systems plenty of times and I never saw this problem,
this is why I wonder if it is a problem of the system size or of the PAW pseudo for wich for example the becsum couldbe bad performed when some PAW data are deallocated:
" Checking if some PAW data can be deallocated...
       PAW data deallocated on    4 nodes for type:  1
       PAW data deallocated on   32 nodes for type:  2
       PAW data deallocated on   35 nodes for type:  3"

Regards,

Antoine Jay
LAAS-CNRS
Toulouse, France

Le Dimanche, Août 16, 2020 23:20 CEST, Stefano de Gironcoli <degironc at sissa.it> a écrit:

6.d-9 is still too large.. it should be something like 1d-13 to aim at a smaller scf estimate.
are you really starting from the scf charge and wfcs of the same conficuration ? 
stefanoOn 16/08/20 22:05, Antoine Jay wrote:Dear Stephano,
adding diago_thr_int=1.0D-8 does not change the first conv_thr (exept the average#of iterations)
As you said, the first value 1.0D-2 is detected to  be too large and is updated to 6.0D-9 so I don't see why changing manually the first value would change something if it is already automatically changed...

Antoine Jay
LAAS-CNRS
Toulouse, France

Le Samedi, Août 15, 2020 17:10 CEST, Stefano de Gironcoli <degironc at sissa.it> a écrit:

Hi Antoine,
  don't know exactly why you get this result but one thing you can try is to set diag_thr_init ~ conv_thr/Nelec/10 so the first diagonalization is pushed tighter (if the wfcs are already very good it should not take too many iterations) and the computed dr2 estimate should be more faithful
  now diag_thr_int is 1.d-2 then updated to 6.e-9 which is consistent with conv_thr ~6.d-5...
  idk. you can try
  stefano
  On 14/08/20 17:09, Antoine Jay wrote:Dear all,

I'am doing two consecutive scf calculations with exactly the same structure and parameters by calling qe6.5 as a library (attached output files).
For the second call, I use the options:
startingwfc='file' and input_rho ='rho.in'
where these inputs are the converged wfc1.dat and charge-density.dat of the first step.
Here I face two problems:

-I expected that the initial scf accuracy is 10^-11 as obtained at the end of the first step, but it is only 10^-4.
How is it possible to explain such a decrease? I generally loose only 2 orders of magnitude by doing this.

-Even with less scf iterations, the cpu time is greater.
Is it possible that some extra memory is allocated by qe when input rho and wfc are asked, and not desallocated?

Note that until now, I have these troubles only when I use paw pseudopotentials on big systems.

Regards,

Antoine Jay
LAAS-CNRS
Toulouse France

  _______________________________________________
Quantum ESPRESSO is supported by MaX (www.max-centre.eu/quantum-espresso)
users mailing list users at lists.quantum-espresso.org
https://lists.quantum-espresso.org/mailman/listinfo/users

  _______________________________________________
Quantum ESPRESSO is supported by MaX (www.max-centre.eu/quantum-espresso)
users mailing list users at lists.quantum-espresso.org
https://lists.quantum-espresso.org/mailman/listinfo/users

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.quantum-espresso.org/pipermail/users/attachments/20200817/0c05e045/attachment.html>