[Pw_forum] Benchmarking QE 4.1

[Marcos Veríssimo Alves]

Hi all,

I have expressed myself pretty badly, as I see from looking at the
email I wrote. So I'll try to use the correct terminology and express
myself better - my sincere apologies.

What I meant to say is that, right after the initial output of the
data (k-points, largest allocated arrays, etc), it displays "total cpu
time spent up to now is", and then another measure of the time spent
before displaying the total energy, Harris-Foulke estimate and scf
accuracy (I'll call the difference between these times dt_0). Then,
the time dt1, that is, the difference between the times displayed
between the subsequent successive DAVIDSON DIAGONALIZATION STEPS
(there probably lies the source of the confusion) is much smaller than
dt_0. Maybe a "picture" would be worth a thousand words, so here is an
excerpt from the output of my run for 48 k-points:

     starting charge  157.98451, renormalised to  164.00000
     Starting wfc are  118 randomized atomic wfcs

     total cpu time spent up to now is     81.63 secs

     per-process dynamical memory:    30.0 Mb

     Self-consistent Calculation

     iteration #  1     ecut=    28.00 Ry     beta=0.40
     Davidson diagonalization with overlap
     ethr =  1.00E-02,  avg # of iterations =  5.0

     total cpu time spent up to now is   1436.72 secs

     total energy              =   -1204.37619760 Ry
     Harris-Foulkes estimate   =   -1206.19047782 Ry
     estimated scf accuracy    <       3.31218627 Ry

     iteration #  2     ecut=    28.00 Ry     beta=0.40
     Davidson diagonalization with overlap
     ethr =  2.02E-03,  avg # of iterations =  1.8

     total cpu time spent up to now is   1730.39 secs

     total energy              =   -1203.96488414 Ry
     Harris-Foulkes estimate   =   -1206.39476160 Ry
     estimated scf accuracy    <      10.70119539 Ry

As for the number of points being so high, you are right, Stefano. In
principle I wouldn't need so many points, but I'm using no symmetries
in my calculation (actually pwscf could find none). This high number
of k-points comes from a 12x12x8 Monkhorst-Pack mesh thus generated
for a tetragonal cell and anti-ferrodistortive geometry (rotated
oxygen octahedra). What I find funny is that pwscf cannot find any
symmetries, even if I give crystal coordinates up to the 9th figure.
Maybe the rotation of the oxygen tetrahedra is not as precise as it
would be necessary for pwscf to find the correct symmetries - anyway,
I'd like to see a result for a plane-wave calculation without any
symmetries, just for the sake of perfectionism. I just find it very
weird that the time between the initialization of the calculation and
the end of the first Davidson step takes so much longer to complete
than the other Davidson steps.

Cheers,

Marcos

On Thu, Jan 21, 2010 at 2:35 PM, Stefano Baroni <baroni at sissa.it> wrote:
> just as a side note, the fact that subsequent SCF steps require less
> davidson iterations should be seen as an advantage (due to the use of smart
> iterative diagonalization techniques), rather than a disadvantage. I do not
> quite understand why you need so many k-points, given the symmetry of
> SrTiO3, and its insulating character ... SB
>
> On Jan 21, 2010, at 2:26 PM, Lorenzo Paulatto wrote:
>
> On Wed, 20 Jan 2010 20:04:40 +0100, Marcos Veríssimo Alves
> <marcos.verissimo.alves at gmail.com> wrote:
>  The time bewteen the end of
>
> the initialization of the calculation (starting mode="from scratch")
>
> and end of the first calculation step are much longer than the typical
>
> times for the subsequent davidson diagonalization steps (between 4-5
>
> times longer):
>
> Dear Marcos,
> when performing the first diagonalization the initial trial wavefunctions
> used by the algorithm are just a superposition of atomic wavefunctions,
> which are only a very rough estimate of the eigenvaleus. On subsequent
> steps the initial trial wavefunctions are the ones from the previous scf
> step, which are usually quite a good estimate. If after finishing the
> calculation you restart it from scratch, but setting startingwfc='file',
> you should find that now the first diagonalization step is as fast as the
> subsequent ones.
>
> best regards
>
>
> --
> Lorenzo Paulatto
>
> *** Note: my affiliation has changed! please send future
> correspondence to: <Lorenzo.Paulatto at impmc.upmc.fr> ***
>
> post-doc @ IMPMC/UPMC - Université Paris 6
> phone: +33 (0)1 44 27 74 89
> www:   http://www-int.impmc.upmc.fr/~paulatto/
>
> previously:
> phd student @ SISSA  &  DEMOCRITOS (Trieste)
> phone: +39 040 3787 511
> www:   http://people.sissa.it/~paulatto/
> _______________________________________________
> Pw_forum mailing list
> Pw_forum at pwscf.org
> http://www.democritos.it/mailman/listinfo/pw_forum
>
> ---
> Stefano Baroni - SISSA  &  DEMOCRITOS National Simulation Center - Trieste
> http://stefano.baroni.me [+39] 040 3787 406 (tel) -528 (fax) / stefanobaroni
> (skype)
> La morale est une logique de l'action comme la logique est une morale de la
> pensée - Jean Piaget
> Please, if possible, don't  send me MS Word or PowerPoint attachments
> Why? See:  http://www.gnu.org/philosophy/no-word-attachments.html
>
>
>
>
>
>
>
> _______________________________________________
> Pw_forum mailing list
> Pw_forum at pwscf.org
> http://www.democritos.it/mailman/listinfo/pw_forum
>
>