[Pw_forum] questions on molecular dynamics example
Eduardo Ariel Menendez P
emenendez at macul.ciencias.uchile.cl
Thu Dec 28 18:21:29 CET 2006
I think that now I understand the usefulness of extrapolation.
Default options are
I think that wfc_extrapolation acts only for the first self-consistent
iteration at any ionic configuration, i.e., the guess wfc is taken from
the converged wfc at the previous ionic config, and extrapolated using
some differential formula. Its use is to accelerate the first scf
iteration. Once the first iteration is completed, the wfc is the same
either if extrapolation is used or not (if it is converged). Hence, the
guess wfc for the second iteration is always the one that satisfy the KS
equation for the guess potential. Hence, wfc_extrapolation does not affect the
number of scf iterations.
pot_extrapolation influences the guess for the potential, then it has a
direct influence on the number of iterations needed for self-consistency.
Here I assume that the default value is very good and is almost equivalent
to using the second order extrapolation (at least for my insulating system
and for the example 4). I have just made a test calculation (verlet dynamics, no T
rescaling) using pot_extrapolation='none' and the number of scf iterations
for convergence increases.
Please, correct me if this is not true.
It left to find out if the acceleration of the firs scf iteration mades a
big difference in the total CPU time.
> Wow!! You are not taking hollidays yet! Thanks for your cooperation.
> I have rather low memory, then maybe disk_io is good for me.
> Regarding the wfc extrapolation, I just realize that I am doing velocity
> rescaling at every ionic step, hence a 'second order ' extrapolation would
> be fooled. However, I don not understand why in the example 4 I do not see
> differences between the use of extrrapolation or commenting it.
> > From: "Axel Kohlmeyer" <akohlmey at cmm.chem.upenn.edu>
> > To: pw_forum at pwscf.org
> > Subject: Re: [Pw_forum] questions on molecular dynamics example
> > Reply-To: pw_forum at pwscf.org
> > On 12/25/06, Eduardo Ariel Menendez P
> > <emenendez at macul.ciencias.uchile.cl> wrote:
> > > Hi Paolo,
> > > I saw these options in the examples of molecular dynamics woth PW. There
> > > is few information in the Doc files and the forum.
> > >
> > > pot_extrapolation=3D'second-order',
> > > wfc_extrapolation=3D'second-order',
> > >
> > > I guessed that extrapolation options would reduce the number of
> > > self-consistent steps in MD steps. However, in the case I am running
> > > (CdTe) he number of self consistent steps at each ionic steps averages
> > > 16, regardless of the extrapolation keywords.
> > >
> > > Other keyword that I see in the examples is
> > > disk_io=3D'high',etc,
> > > Is this keyword useful to improve the efficiency of molecular dynamics?
> > that depends on the individual example and hardware.
> > with a higher disk_io level, more data is written to disk
> > and re-read instead of being kept in memory. if you have
> > a lot of memory, a lower disk_io level should be better.
> > in fact, for some machines with no local disk storage,
> > it would be nice of have a disk_io level of 'no', i.e. keep
> > everything in memory...
> > >
> > > Other issue. I had this error in step 90
> > > iteration # 4 ecut=3D 30.00 Ry beta=3D0.70
> > > Davidson diagonalization with overlap
> > > %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%=
> > %%%%%
> > > from rdiaghg : error # 1986
> > > info =3D/=3D 0
> > did you try changing the diagonalizer?
> > > Consulting the manual and the forum I see that it is difficult to find
> > > the cause, and that the first choice is restart
> > > and see what happens. I did it and the calculation just continued without
> > > error. ?Does it allows to discard the possible cause of bad
> > > pseudopotential, and bug of the lapack algorithm, in favor of buggy mahin=
> > e
> > > optimization of lapack (MKL) ?
> > i don't think that you have a buggy MKL. it might just be, that it get fed =
> > an
> > input that it cannot converge. the fact that you do not see any improvement=
> > s
> > from wavefunction extrapolations could indicate that the order of your stat=
> > es
> > keeps changing a lot and perhaps the davidson implementation is more
> > sensitive there.
> > (note: i'm not an expert here, but currently trying to learn about the deta=
> > ils
> > of these issues myself).
> > i noticed however, that with different lapack implementations, this kind of
> > error message keeps popping up more or less often (e.g. more often with
> > ATLAS than with ACML than with MKL). also the compiler seems to have
> > some impact, so it might be that very small differences in accuracy may
> > have a large impact here (which would make it difficult to rule out all oth=
> > er
> > possible sources of problems like suboptimal pseudopotentials, as they
> > tend to enhance existing accuracy problems...).
> > ciao,
> > axel.
> > >
> > > Thanks
> > > Eduardo
> > >
> > > Eduardo A. Menendez Proupin
> > > Department of Physics
> > > Faculty of Science
More information about the users