New OT Algorithms

vweber valer... at hotmail.com
Wed May 9 22:00:12 CEST 2007


Hi Nichols
The speed up is always there, i.e. new OT has less matrix multiplies
than the
standard OT (especially when the WF doesnt change much).
Now the problem is how much time is spent in the other parts of the
code.
For small systems the time spend in OT is usually small compare to
the
other parts, then you wont see much difference.
For large and very large systems OT starts to dominate the
calculations.

I would say that your 4000+ atoms is a good candidate for the new OT.

Valery


On May 8, 11:24 pm, "Nichols A. Romero" <naro... at gmail.com> wrote:
> Valery,
>
> Thanks for your comments. The benefits are for large systems and MD.
>
> Is the speed up for MD _only_ for large system though? I'm am doing
> MD on system sizes as small as 500 atoms as well as some large ones
> (e.g. 4000+ atoms).
>
> Thanks in advance for clarifying.
>
> On 5/8/07, vweber <valer... at hotmail.com> wrote:
>
>
>
>
>
> > Hi Nichols
>
> > The new OT is basically a mapping of the orthogonality constrained
> > energy functional to an unconstrained one.
> > The orthogonality is enforced via a refinement polynomial.
> > The minimization gets faster and faster while the wavefunction
> > converges.
> > It is, in the best case (e.g. during MD), about 2.5 faster than the
> > standard OT (for what it replaces!).
> > It converges the SCF like the standard OT or slightly faster
> > especially if you dont update the preconditionner.
> > As said Juerg, you really see the difference for large systems. For
> > example a box of 512 waters will run (total time per SCF) 20% faster
> > and
> > 35% for 4096 waters. The speedup is much higher if no preconditionner
> > is used.
> > I would appreciate any feedback.
>
> > Valery
>
> > On May 7, 11:19 pm, "Nichols A. Romero" <naro... at gmail.com> wrote:
> > > Hi,
>
> > > It would be nice if someone could comment on the advantages of the new
> > OT
> > > algorithm.
>
> > > Thanks,
>
> > > --
> > > Nichols A. Romero, Ph.D.
> > > 1613 Denise Dr. Apt. D
> > > Forest Hill, MD 21050
> > > 443-567-8328 (C)
> > > 410-306-0709 (O)
>
> --
> Nichols A. Romero, Ph.D.
> 1613 Denise Dr. Apt. D
> Forest Hill, MD 21050
> 443-567-8328 (C)
> 410-306-0709 (O)




More information about the CP2K-user mailing list