[Octopus-users] orthogonalization in td

[Miguel Marques]

  Hi Matthieu,

  There are no orthogonalizations when you do a TDDFT propagation. The
slowdown has a very simple reason: the recalculation of the external
potential every time you move the ions (every time-step). This part can
be extremely time-consuming (relatively to the time to propagate the
orbitals). For the moment, Xavier has optimized already the code, and he
has made it parallel in atoms. Of course, if you have small systems, of
a couple of atoms, it can not scale much... I am not sure of what is the
solution to the problem. One can try to optimize the code further, of
course. Another possible way out is only to update the external
potential every X time-steps or, even better, only after the ions have
moved more than a certain distance. This is very easy to implement and
may speed-up your calculations considerably, as the ions move very
slowly when compared to the electrons. Of course, there might be issues
with energy conservation...

  On a side note, it is not stupid at all to calculate vibrational
spectrum from TDDFT. Actually, you should really take a look at the fast
Ehrenfest stuff of Xavier et al, that basically allows you to perform a
"TDDFT" calculation with a larger time-step.

  hope this helped,
  Miguel

matthieu verstraete wrote:
> Hello everyone,
> 
> I am trying a very long td run for a small molecule, to get
> vibrational spectra from tddft (I know this is not the efficient way
> to go, it's more for educational purposes). In practice, I need a
> timestep of 0.01 or 0.001, which makes for horribly long simulations
> if I want the slow phonon modes to integrate well. I have tried
> parallelization (8, 16, or 60 procs), but this makes little difference
> on average (most steps are faster, and a few are much slower, the
> average is even slightly worse in parallel wrt sequential).
> 
> In the purely electronic dynamics case (no moving ions and par_domains
> only) the speedup is brilliant, almost linear. So what can I do? Are
> things slowed because of an added orthonormalization step or something
> like that? Can I use some modified Ehrenfest (for the moment it's
> normal) to improve my time step? My energy is constant for slightly
> larger dt, but I also need the long time stability in this case, for
> the low w phonons, so it gets delicate.
> 
> Cheers
> 
> Matthieu
> 
> 
> 


-- 
Dr. Miguel A. L. Marques
marques at tddft.org

Laboratoire de Physique de la Matière Condensée et Nanostructures
(LPMCN) - Université Lyon I
Bâtiment Brillouin, Domaine scientifique de la DOUA
69622 Villeurbanne Cedex
Tel +33 (0)4 72448187
Fax +33 (0)4 72432648