[Q-e-developers] Band parallelization

Ivan Girotto igirotto at ictp.it
Mon Jun 17 17:19:28 CEST 2013


Based on my experience I can stand that -ntg and -nbgrp do not work 
together.
As far as I know there is no track of this anywhere, including the paper 
on PRACE website and CPC.

My two cents,

Ivan

On 17/06/2013 17:10, Paolo Giannozzi wrote:
> On Mon, 2013-06-17 at 16:30 +0200, nicola varini wrote:
>
>> Paolo, what kind of material do you think is useful to have?
>> A tutorial or a user guide or some guidelines?
> as a first approximation, any kind of material. There is
> nothing (rien nichts niente nada ничего موشي ) right now
> in the documentation. There is just a single mention of
> option "-nband" or "-nb" in Doc/user_guide.tex, and this
> mention in PW/examples/EXX_example/README:
>    An experimental parallelization on the band structure
>    is also available (pw.x -nbgrp N)
> My specidic original motivation for my question is to
> figure out what the "smallmem" option exactly does and
> in which cases the distribution of arrays dimensioned
> (nkb,nbnd) hidden in "calbec" is used. For calculations
> in large unit cells, O(1000) atoms or so, distributing
> such arrays (after reducing them to the strict minimum)
> is often needed, but it seems to me that using band
> distribution slows down things. An alternative solution
> is to split the sum over beta functions.
>
> P.




More information about the developers mailing list