[Wien] WARN: Not all eigenvectors are orthogonal, and CLUSTERSIZE parameter

[Peter Blaha]

> I have met problems in lapw1.
> I work in fine grained parallel mode, using 32 cpus, simulating a system
> After 4 or 5 iterations, the program diverges (eventough the mixing
> parameter is not large, i.e. 0.1).
> In all the iterations the program gives the warning:
>
> :WARN :      WARNING: Not all eigenvectors are orthogonal
>
> and in the 'output1' it suggests to increase the CLUSTERSIZE parameter
> in seclr4.F
> Default value of CLUSTERSIZE is 200.
>
> I have two questions:
> 1) Might this warning have something to do with the divergence of my system
> 2) Is there a way to judge the new value of CLUSTERSIZE, from, say, some
> input parameters, or
> one should just just increase it till the warning disappears?

I was told that these SCALAPACK algorithm may have problems to
orthogonalize eigenvectors for "nearly degenerate" eigenvalues exactly.
Thus I insisted on some warnings, but I was also told that "usually this
should NOT be a problem". My own experience is very limited.
(We checked once the resulting densities from a parallel and serial
eigenvector and this was ok).

Thus I don't think this is the reason for your crash.
grep :DIS case.scf to see the divergence and also check the individual
charges (QTLxxx), if they are still reasonable or if anything is
"missing" (eg. if you expet some semicore p-charge of an atom, but this
is not present anymore,...). Most likely you need an even smaller mixing
factor.

Let me know if you have found out more about the "orthogonal" problem and
if it really was a problem.
                                      P.Blaha
--------------------------------------------------------------------------
Peter BLAHA, Inst.f. Materials Chemistry, TU Vienna, A-1060 Vienna
Phone: +43-1-58801-15671             FAX: +43-1-58801-15698
Email: blaha at theochem.tuwien.ac.at    WWW: http://info.tuwien.ac.at/theochem/
--------------------------------------------------------------------------