The ABINIT Users Mailing List ( CLOSED )

[Fabien Bruneval <fabien.bruneval@polytechnique.fr>]

Dear Philippe and Don,

The parallelization over the bands takes advantage of the summation over 
occupied and empty states in the expression of the screening and of the 
self-energy. I mean that the calculation of the elements in the sums is 
distributed among the processors.
So the parallelization is not over the matrix elements, but over the 
internal summations. As you said, the wavefunctions are then distributed 
accordingly among the different CPUs. This allows us to save memory.

Concerning the problem mentioned by Don, something needs to be fixed for 
sure. Don, have you also noticed a problem in the final result for the 
band structure?

I don't know what is the current state of development of the band 
parallelism. Maybe Riad could say something more accurate than me...

Have a nice week-end.

Fabien



Philippe Blaise wrote:
> Please, someone could explain more in details what is written in the 
> release notes for the 5.3 ?
>
> B.6 *Parallelism over bands is available for GW*. See the input variable 
> gwpara.
>    This parallelism allows for memory savings, and is quite efficient.
>    At present, either the band parallelism or the k point parallelism is 
> available
>    in GW, they cannot be used simultaneously.
>    This has been implemented by R. Shaltaf.
>    (Note, however, that parallelism over bands in GW is not compatible 
> with nsppol=2 as of v5.3.0)
>    Automatic test for band parallelism : paral#N
>
> and the corresponding action of the parameter gwpara=2 : parallelisation 
> on bands.
>
> I'm not sure to understand : "bands" stand for matrix elements ? 
> wavefunctions ?
>
> Thanks a lot,
>
> Philippe