The ABINIT Users Mailing List ( CLOSED )

[Emmanuel Arras <emmanuel.arras@cea.fr>]

I got the same results as you : It works just fine without band parallelization.
The parameters I use are :
paral_kgb 1
fftalg 401
wfoptalg 4
noalg 4

I'm afraid I'm not qualified to look into this any further... I'll just have to avoid the use of ecutsm in the futur, and hope the bug has no other consequences...

Thanks a lot for the help.

Emmanuel ARRAS


PS : the input file I gave you was an old version. I agree with everything you say. (However, it is interesting to notice that it is actually not necessary to define paral_kgb 1 to turn on the band parallel stuff. Some how, it works without it, but twice slower. Thus it can be confusing).




Josef W. Zwanziger a écrit :

937205.52685.qm@web1112.biz.mail.sk1.yahoo.com" type="cite">

OK, I finally get something like you with ecutsm, and it seems to be strictly related to band parallelization. If I run exactly your non-primitive input cell (strained in one direction) either in serial mode or the default kpt parallelization, the scf cycle converges just the same with ecutsm = 0 or 0.5. However, if I then turn on the band parallelization also, it converges just as fast with ecutsm = 0 as before but doesn't fully converge with ecutsm = 0.5. 

[by the way, a couple of comments about your input file--first, you had nshiftk 1 but then three lines of shiftk data, it will only read the first one since it reads nshiftk of them only--is that what you wanted? since your strained cell actually has body-centered tetragonal symmetry, I think nshiftk 1 and shiftk 0.5 0.5 0.5 would be most efficient; Also, your input file didn't define paral_kgb 1, I think that is necessary to turn on all the band parallel stuff]

Anyway, I suspect there is a bug in the band parallelization set up. My first suspicion would be to look in kpgsph.F90, as I was just tracking down a different bug there concerning the possibly related case of running when getcell /= 0. In that code there is extensive branching to account for various parallel possibilities, so maybe there's a problem there.

Good luck--

Joe

 

Josef W. Zwanziger
Professor of Chemistry
Canada Research Chair in NMR Studies of Materials
Director, Atlantic Region Magnetic Resonance Centre
Department of Chemistry
Dalhousie University
6274 Coburg Road
Halifax, Nova Scotia B3H 4J3 Canada
tel: +1.902.494.1960
fax: +1.902.494.1310
web: http://jwz.chem.dal.ca
jzwanzig@jzwanzig.org,jzwanzig@dal.ca

---

From: Emmanuel Arras <emmanuel.arras@cea.fr>
To: forum@abinit.org
Sent: Tuesday, December 16, 2008 11:41:09 AM
Subject: Re: Re: [abinit-forum] ecutsm and electronic convergence

ok, I tried to improve k-mesh (16x16x16 => 288 kpts (kptrlen = 87) as compared to 8x8x8 => 40 kpts (kptrlen = 32) previously ) and increase nband (up to 32 bands (now 13 are empty for spin up, and 19 for spin down) as compared to 24 bands previously ( 5 and 11 empty) ). It gives the same result : convergence is very very slow with ecutsm, but not without !

Concerning the pseudopotential, I used the one given one the abinit download page (http://www.wfu.edu/~natalie/papers/pwpaw/periodictable/atoms/Fe/index.html). I guess Josef Zwanziger used the same.

I think the problem may indeed be connected to the parallelization. I still have to perform the calculations with abinis to see.

Another thing I mentioned earlier : if I remove the symmetries (nsym 1), the convergence is really difficult, but the calculation without ecutsm is still faster. (Fe-BCC_ecutsm_aniso_nsym_1.png attached).

And finally, a plot of convergence as a function of ecutsm (Fe-BCC_ecutsm_aniso.png attached). The dependence is obvious.

I'll keep you updated.
Thanks

Emmanuel ARRAS


matthieu verstraete a écrit :

On Tue, Dec 16, 2008 at 12:08 PM, Emmanuel Arras <emmanuel.arras@cea.fr> wrote:

I'm not quite sure I understand what you are saying, so correct me if I'm wrong :
- You tried my input file, and found a very very slow convergence (the same as me).
 

- To see the difference with ecutsm, you tried to remove ecutsm from my input, and launched it. You also got a very very slow convergence (=> that's why you say there is no big difference with ecutsm). Then I don't understand why we get different results.

maybe the pseudopotential. I don't think you misran or miscompiled abinit, it's just a conjunction of factors (underconvergence of some, magnetism, etc) which allow this effect of ecutsm.

I tried to inprouve my calculations as you told, and as soon as these calculations with a denser kpt coverage and more bands are finished, I will let you know. Still, I just need to remove ecutsm to get a good convergence...

It may be something to do with the interplay between ecutsm and the parallelization. The block-parallel algorithm is trickier than the standard one where all the bands are kept together. This may be the source of differences in the SCF.

Joe - were you using the parallelization on bands?
 

Matthieu

--
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Dr. Matthieu Verstraete

European Theoretical Spectroscopy Facility (ETSF) 
Dpto. Fisica de Materiales, 
U. del Pais Vasco,            
Centro Joxe Mari Korta, Av. de Tolosa, 72,   Phone: +34-943018393
E-20018 Donostia-San Sebastian, Spain        Fax  : +34-943018390

Mail : matthieu.jean.verstraete@gmail.com
http://www-users.york.ac.uk/~mjv500