[Pw_forum] PWscf and highly parallel machines

[Nichols A. Romero]

Nicola et al.,

I am attempting to setup such a test under controlled conditions.

It would be helpful to clarify a point, because we want to avoid comparing
apples with oranges.

There are two ways to solve for the KS ground state.
1. Direct minimization of KS functional
2. Self-consistently solving the KS eqn.

The conjugate-gradient (cg) in PWSCF and CP are not doing different things
same thing. In PWSCF we are doing #2, while in CP we are doing #1.

What would best for comparison purposes then? Have each do 10 iterations
with CG using random wavefunctions as a starting guess. I would think that
even with those settings that PWSCF still has the advantage due to have
startingpot = atomic.

Hopefully, I am not saying jibberish. I having not thought how the PW codes
work in a while so I am rusty.


On 2/15/07, Nicola Marzari <marzari at mit.edu> wrote:
>
>
>
> Paolo Umari has also implemented a conjugate gradients minimization
> in CP.
>
> so, for a CP calculation, in which the only thing you need is the
> electronic ground state of a given structure, this would be the best
> strategy - our few tests were that it was 3-4 times more effective than
> damped dynamics. If you want to relax a structure, you should do, after
> this first electronic calculation, a combined damped dynamics on
> electrons and ions.
>
> One thing we haven't tested is the relative performance of PWSCF and
> CP - my guess is that the algorithms of PWSCF are much faster, although
> the balance might change as the system becomes larger and larger.
> In principle, the CP algorithm is variational, so should always
> converge, but again this is rarely a problem.
>
> Bottom line - advantages for PWSCF is that it should be faster for most
> systems of interest. CP uses much less memory, and should be very
> efficient in calculating H|psi>, so the balance when going towards very
> large scale or very parallel calculations would be extremely interesting
> to explore.
>
> Any comparative test, under controlled conditions, would be very useful.
>
>                                 nicola
>
>
> Paolo Giannozzi wrote:
> >
> > On Feb 14, 2007, at 18:48 , Nichols A. Romero wrote:
> >
> >> If one were to use CP across such a large number of processors,
> >> what algorithm is your best bet for minimizing your electronic d.o.f.,
> >> sd or cg
> >
> > steepest descent is very inefficient and should be used only
> > for the initial steps of a simulation or in desperate cases. The
> > "standard" way of minimizing electronic degrees of freedom
> > in CP is damped dynamics
> >
> > Paolo
> > ---
> > Paolo Giannozzi, Democritos and University of Udine, Italy
> >
> >
> > _______________________________________________
> > Pw_forum mailing list
> > Pw_forum at pwscf.org
> > http://www.democritos.it/mailman/listinfo/pw_forum
>
>
> --
> ---------------------------------------------------------------------
> Prof Nicola Marzari   Department of Materials Science and Engineering
> 13-5066   MIT   77 Massachusetts Avenue   Cambridge MA 02139-4307 USA
> tel 617.4522758 fax 2586534 marzari at mit.edu http://quasiamore.mit.edu
> _______________________________________________
> Pw_forum mailing list
> Pw_forum at pwscf.org
> http://www.democritos.it/mailman/listinfo/pw_forum
>



-- 
Nichols A. Romero, Ph.D.
1613 Denise Dr. Apt. D
Forest Hill, MD 21050
443-567-8328 (C)
410-306-0709 (O)
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.quantum-espresso.org/pipermail/users/attachments/20070215/1716f31f/attachment.html>