[Octopus-devel] [Octopus-notify] svn commit: r4908 - trunk/src/sternheimer by xavier

[David Strubbe]

Xavier,

I think this orthogonalization you added is very similar in effect to the
orthogonalization I added a few months ago, controlled by the
"Preorthogonalization" variable, so we shouldn't have both.

The differences are:
1. Mine occurs at the beginning of the SCF iteration, whereas yours is at
the end.  The effect will be the same, except that the initial guess will
not be orthogonalized.  For zero initialization, as default, this is not an
issue; but for the non-dispersive kdotp initialization, the
orthogonalization should be done before entering the iterations.  Of course
that could just be done at the initialization stage.
2. Mine is done all at once for all states, rather than state by state; I
assumed since there was a function available to do this, it was better in
some way.
3. I didn't take into account what should be done in the occ_response case,
whereas you did.
4. Mine can be turned off by a variable, since as you wrote in your JCP
paper, the orthogonalizations are increasingly expensive for large systems.
I wonder if this is not a significant issue in the orthogonalizations you
added to the BiCGstab solver too.

David

On Sat, Jan 31, 2009 at 11:51 AM, <octopus-notify at tddft.org> wrote:

>      Author: xavier
>        Date: Sat Jan 31 19:51:14 2009
> New Revision: 4908
>   Changeset: http://www.tddft.org/trac/octopus/changeset/4908
>
> Modified:
>   trunk/src/sternheimer/sternheimer_inc.F90
>
> Log:
> The resulting vector after solving the Sternheimer equation has to be
> re-orthogonalized, this greatly increases convergency (we had this
> before and it was removed at some point).
>
>
> _______________________________________________
> Octopus-notify mailing list
> Octopus-notify at tddft.org
> http://www.tddft.org/mailman/listinfo/octopus-notify
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.tddft.org/pipermail/octopus-devel/attachments/20090204/e0d1bb0e/attachment.html