[Pw_forum] Restarting PWSCF - usability issues

Carlo Sbraccia sbraccia at sissa.it
Sat Mar 13 18:38:52 CET 2004

On Fri, 12 Mar 2004, Konstantin Kudin wrote:

> > restarting with atomic wfc usually implies that
> > several iterations are needed 
> > to diagonalize the hamiltonian because the threshold
> > is more strict than the 
> > one used for he first scf iteration of a
> > "from_scratch" calculation.
> > Moreover, since there is no way for the code to know
> > at restart time whether 
> > the starting potential is close to the scf one or
> > not, the code doesn't know 
> > the correct value for diagonalization threshold.
>  Dear Carlo
>  It appears that the code is already quite smart and
> can detect such problems and lower the threshold on
> the fly at the same cycle. So restarting from *.rho
> file alone costs only that first diagonalization with
> large threshold when it finds out that the threshold
> should be smaller - hardly a big price to pay for the
> convenience of not dealing with *.wfc* files.
>  Here:
>      The initial density is read from file     
> yz_1s.rho
>      Starting wfc are atomic
>      total cpu time spent up to now is     34.38 secs
>      iteration #  1     ecut=    28.00 ryd     beta=
> .70
>      Davidson diagonalization with overlap
>      ethr =  1.00E-05,  avg # of iterations =  5.0
>      Threshold (ethr) on eigenvalues was too large:
>      Diagonalizing with lowered threshold
>      Davidson diagonalization with overlap
>      ethr =  2.29E-07,  avg # of iterations =  4.0

Dear Kostya,
I konw that part of the code quite well (I wrote it some weeks ago): it 
was written to avoid these "restart" problems, but, since then, it has not 
been widely tested. In the case reported in your mail it seems to work 
(nice news), but I'm not sure that it is always so and I'd not say that old wfcs are 
always useless.
Indeed you see that starting from atomic 9 davidson iterations are needed 
(in average) to diagonalize the hamiltonian: using the old wfcs as 
starting wfc I guess that a smaller number of iterations was probably needed.
Of course if you have serious problems in dealing with *.wfc* files this 
can be considered a worthy price.

More information about the users mailing list