[Wien] Wien post from pascal.boulet at univ-amu.fr (Errors with DGEMM and hybrid calculations)

[Laurence Marks]

Continuing the thread for clarification; please correct my errors:

If I remember right, lmaxv & lmaxe change how the LM expansion is done
inside the spheres, so need to be larger with higher L. The CPU scales
something like a square, and they are CPU expensive.

GMAX is a plane wave expansion, so (like lapw0) would need to be larger for
small RMT such as H or odd cases. Again something like a square dependence,
although I am not sure how CPU expensive except for a surface.

The CPU time will scale something like the product of the outer k-mesh and
the inner (for which a reduced mesh can be used). Standard accuracy issue.

NBand in case.inhf is I think linear in CPU; I am not exactly sure why it
needs to be larger than the number of occupied bands (forgotten).

RKMAX is standard, with a 2-3 power dependence of CPU time.

N.B., I think there is also a dependence upon how different the straight
lapw1 & hf are, for instance the band gap in lapw1. This matters for how
large the number of states used in lapw1 has to be.



_____
Professor Laurence Marks
"Research is to see what everybody else has seen, and to think what nobody
else has thought", Albert Szent-Gyorgi
www.numis.northwestern.edu

On Sat, Aug 15, 2020, 08:38 Tran, Fabien <fabien.tran at tuwien.ac.at> wrote:

> That's difficult since it may depend on the system, RMT, property, and
> required accuracy, but roughly (concerning  the parameters in case.inhf):
>
> -For geometry: none of the parameters really need to be increased compared
> to default. In particular, nband can be set to minimum required (number of
> occupied bands plus one). Maybe gmax (from case.inhf) should be tested if
> RMT are chosen clearly smaller than default. The option "-nonself" could
> also be used for non-magnetic systems.
>
> -For cohesive energy: preferably test nband, gmax, lmaxe and lmaxv.
> Especially the latter two for systems with an open  f-shell.
>
> -For electronic structure: test lmaxe and lmaxv for systems with an open
> f-shell. For the band gap of non-magnetic systems, the option "-diaghf" (
> https://urldefense.com/v3/__https://doi.org/10.1016/j.physleta.2012.01.022__;!!Dq0X2DkFhyF93HkjWTBQKhk!AN3JSEcq-L36m1Ickxc1xjixv8KB9lE8fbCvyAvgSj7ySnH9QhVXJB8tDGGm-rjxGqJUjQ$
> ) can be used to reduce the computational time by two orders of magnitude.
>
> -For electron density (e.g., for EFG): Test carefully all parameters.
>
> -k-mesh: In general, the convergence of properties with HF/hybrids is
> slower than with LDA/GGA/MGGA (fortunately only slightly slower with
> screened hybrids like HSE06).
>
> -RKmax: the convergence is the same as for LDA/GGA/MGGA.
>
> -Computational time: strongly influenced by k-mesh, RKmax, nband, lmaxe
> and lmaxv. Less by gmax (from case.inhf).
>
>
> From: Wien <wien-bounces at zeus.theochem.tuwien.ac.at> on behalf of
> Laurence Marks <laurence.marks at gmail.com>
> Sent: Saturday, August 15, 2020 2:40 PM
> To: A Mailing list for WIEN2k users
> Subject: Re: [Wien] Wien post from pascal.boulet at univ-amu.fr (Errors with
> DGEMM and hybrid calculations)
>
> It would be good to have a brief summary of what matters with hf:
> a) For speed
> b) For accuracy
>
> It is probably somewhere in the docu, but another cite for the list would
> be useful (to me as well as others).
> _____
> Professor Laurence Marks
> "Research is to see what everybody else has seen, and to think what nobody
> else has thought", Albert Szent-Gyorgi
> http://www.numis.northwestern.edu
> _______________________________________________
> Wien mailing list
> Wien at zeus.theochem.tuwien.ac.at
>
> https://urldefense.com/v3/__http://zeus.theochem.tuwien.ac.at/mailman/listinfo/wien__;!!Dq0X2DkFhyF93HkjWTBQKhk!AN3JSEcq-L36m1Ickxc1xjixv8KB9lE8fbCvyAvgSj7ySnH9QhVXJB8tDGGm-rgXzLvd8g$
> SEARCH the MAILING-LIST at:
> https://urldefense.com/v3/__http://www.mail-archive.com/wien@zeus.theochem.tuwien.ac.at/index.html__;!!Dq0X2DkFhyF93HkjWTBQKhk!AN3JSEcq-L36m1Ickxc1xjixv8KB9lE8fbCvyAvgSj7ySnH9QhVXJB8tDGGm-rjETTfF-Q$
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://zeus.theochem.tuwien.ac.at/pipermail/wien/attachments/20200815/9041db86/attachment.htm>