[CPMD-list] increased CPU times for wavefunction iterations

[Salomon Billeter]

Dear Martin,

> Dear CPMD developers,
>
> May I have few questions on wavefunction optimization, in particular CPU
> times for individual iterations? The usual pattern is that first part of
> them takes one value of CPU time (about 82 s in the example below) but
the
> other part spends significantly more time (here 121 s). I see that this
> larger value occurs when wavefunctions are almost converged; often it is
> only in last one or two iterations but sometime in more of them.
>
> I can notice this behaviour at least when using Vanderbilt
> pseudopotentials.
>
> What is a more detailed reason for this?
>
> Is it possible to avoid the incresed CPU times somehow (except cutting
> down the number of iterations)?
>
>
> RESETTING WAVEFUNCTION TOLERANCE FACTOR TO:  1.0
>    1  1.274E-05   2.093E-06   -2003.766203    6.517E-06     82.41
>    2  1.418E-05   7.662E-07   -2003.766206   -3.192E-06     81.77
>    3  1.423E-05   6.226E-07   -2003.766206   -3.597E-07     81.00
>    4  1.120E-05   3.037E-07   -2003.766207   -3.911E-07     82.17
>    5  9.869E-06   2.593E-07   -2003.766207   -1.469E-07     81.38
>    6  9.743E-06   2.190E-07   -2003.766207   -2.256E-07     81.62
>    7  8.522E-06   1.914E-07   -2003.766208   -3.139E-07     81.72
>    8  7.540E-06   1.659E-07   -2003.766208   -2.474E-07     83.64
>    9  6.717E-06   1.455E-07   -2003.766208   -1.815E-07     82.54
>   10  5.885E-06   1.346E-07   -2003.766208   -1.956E-07     82.62
>  RELAXING WAVEFUNCTION TOLERANCE, NEW FACTOR:  2.0
>   11  5.428E-06   1.274E-07   -2003.766208   -1.691E-07     82.66
>   12  4.997E-06   1.185E-07   -2003.766208   -1.303E-07    121.48

during geometry optimization, as long as the wavefunction is far from being
converged, the forces acting on the ions are not calculated because their
calculation can cost a lot of CPU time (far from being converged by CPMD
default means more than three times the convergence criterion). If your
wavefunction converges as slowly as in your case, the calculation of the
forces starts unnecessarily early. You can adjust the criterion for the
calculation of the forces from three times the wavefunction criterion to
<your_factor> using
CONVERGENCE CALFOR
  <your_factor>
If <your_factor> is too close to one, in the worst case, an extra cycle is
required because the forces are not there when the wavefunction is ready.
Btw, in your case, CONVERGENCE RELAX 10 is probably a little aggressive
because the convergence is more or less uniformly slow. If you want to do
save CPU time, tuning CONVERGENCE ADAPT will probably produce more
consistent accuracy.

Best regards,
Salomon