[CP2K:652] Re: cp2k speedup on multicore machines

[Teodoro Laino]

Ciao Luigi,

as we said already on a similar topic in this mailing list, there are  
few things that can speed-up
the SCF during an MD.
I will repeat them here hoping that these suggestion will help also  
other people.
The most fundamental one is a good extrapolator and a good  
preconditioner for the SCF.

instead of those used in that tests you run you may want to try:

&QS
   EXTRAPOLATION ASPC
   EXTRAPOLATION_ORDER 3
&END

and in &OT

PRECONDITIONER FULL_SINGLE_INVERSE

Of course the choice of the preconditioner is very dependent from the  
system you're running. But as a preliminar
trial this should be ok.


Moreover every time  OT  is used to optimize the wavefunction it is  
HIGLY suggested to use a nested SCF procedure..
This does not help at all normal diagonalization schemes but improves  
terribly the convergence with OT.
This is the way to go:

&SCF
   MAX_SCF    30
&END
&OUTER_SCF
      MAX_SCF    5
&END

Don't use too many steps for the inner SCF levels (in the range 30-40  
should be ok) and quite few for the outer (5-10)..
Try it and you will see a great improvement for the first SCF  
optimization..

Keep in mind that absolute running time depends on many things..  
CUTOFF, basis set, preconditioner, extrapolation,
threshold for the convergence of the SCF, precision required in the  
integration and collocation of the density and so on..
So a full comparison can be done only with the exactly the same input  
file.

Cheers,
Teo