The ABINIT Users Mailing List ( CLOSED )

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

Hi Francois
 First, I'd like to thank you for the time you spend on this problem.
Now concerning the problem itself, I'm surprised by the results of Carl 
Greeff, since I only find this problem in parallel, and especially with 
wfopttalg=4. Now this may indeed be a deeper problem only revealed by my 
particular settings...
 As for increasing nline, it works but at a cost of a computation time 
4 times greater per iteration (in the Fe_BCC case).
 Also you are talking about a track you have. Is that some kind of a 
bug you found?

Thanks

Emmanuel ARRAS



BOTTIN Francois a écrit :
> Dear Carl, Joe and Emmanuel,
>
> The problem comes from LOBPCG. And especially from the preconditioning 
> within this algorithm.
> This one depends on the kinetic energy, so that's the reason why 
> ecutsm reveals (and not causes) the problem.
> This trouble of convergence is also present in a few other systems 
> which have been studied in our lab, with or without ecutsm.
> Up to now, we have circumvent this problem by tuning some input 
> parameters.
> Try to increase: nline and/or nband and/or npband. Your calculations 
> should work fine. Keep me in touch.
>
> I'm aware that these tricks are only patchs and that the problem is 
> deeper.
> I'm interested these days in the preconditioning of LOBPCG and have a 
> track.
> The first results on various systems, which show no or poor 
> convergence, are very good (these ones include the iron bcc system of 
> emmanuel).
> This correction improves strongly (several order of magnitude) the 
> residual of the wave functions.
> Before giving it to the community, I have to check that this 
> correction is safe.
> I will try to commit  it for the abinit 5.7 realease.
>
> Francois
>
> greeff@lanl.gov a écrit :
> > Sorry if this a repetition. I tried to post before, but I don't see 
> > that it
> > showed up. I was having a problem with convergence of the parallel 
> > code, and
> > saw this thread, so I did some tests of the ecutsm option. The result 
> > is that
> > there is an interaction of ecutsm /= 0 and wfotpalg = 4 that causes the
> > problem. The problem occurs with parallel over band/fft, parallel 
> > over k-pt,
> > and serial runs.
> >
> > My test case is for bct Sn, using a homegrown PAW from the Holzwarth 
> > code. I am
> > using abinit version 5.6.4, except the serial tests use 5.4.4. The 
> > compiler is
> > intel 10.0.023 with openmpi. I can provide full input files if desired.
> >
> > This table summarizes the results. The number of iterations given is 
> > what is
> > required for convergence with toldfe  1.0d-6. The entries with >n never
> > converged to this tolerance.
> >
> > parallel ecutsm wfoptalg iter
> >
> >  band       1.0        4      >30
> >  band       0.0        4       6
> >  k-pt       1.0        4      >30
> >  k-pt       0.0        4       6
> >  k-pt       1.0       10       6
> >  k-pt       0.0       10       6
> >  serial     1.0     10       7
> >  serial      1.0      4    >12
> >
> > The last serial run needed me to set fftalg 401, which wasn't used in 
> > the other
> > serial run. (It was used in all the parallel runs).
> >
> > Carl Greeff
> >