[gpaw-users] Calculations in LCAO mode

[Ask Hjorth Larsen]

Hi

On Tue, 21 Feb 2012, Ulrik Grønbjerg wrote:

> Dear GPAW users
> 
> I am doing calculations on some rather large metal-slabs using LCAO, and I
> would like some advice in this regard.
> 
> I have been told that performance improves if ScaLAPACK is used for systems
> with more than 50 atoms, which is the case here. I have therefore inserted
> "parallel=dict(sl_default=(4,2,64))," in my GPAW-object. I then see the
> following in my log-file:
> 
> Total number of cores used: 32
> Parallelization over k-points: 4
> Domain Decomposition: 2 x 1 x 4
> Diagonalizer layout: BLACS 4 x 2 grid with 64 x 64 blocksize
> 
> Does this mean that I parallelize over both space and the diagonalization
> operation? I think it might refer to different operations, but is this true?
> If yes, could someone maybe explain it a bit or refer me to the relevant
> documentation?

Within each k-point, there are [domain decomposition] x [band 
parallelization] cpus available for the scalapack operations.  Some 
redistribution of arrays takes place in order for this to be possible, but 
it will work no matter the band/domain parallelization. 
(Performance-wise it is slightly entangled with band/domain 
parallelization, but with just 8 CPUs/k-point you should probably not 
worry about band parallelization yet.)

> I would also like to know if it is possible to set "fprojectors=True" inside
> the script or GPAW-object, or is must always be set at job
> submission/execution?

Try doing this at the top of the script:

   from gpaw import extra_parameters
   extra_parameters['fprojectors'] = True

(This does not *in general* work for "extra parameters", but it will work 
for fprojectors)

> I currently run GPAW version 0.8, but should I switch to version 0.9 to get
> the best performance for LCAO calculations?

Switch to 0.9 when it comes out.

Regards
Ask