The ABINIT Users Mailing List ( CLOSED )

[BOTTIN Francois <francois.bottin@cea.fr>]

Hi Yann and Max,

Yann Pouillon a écrit :
> Hi,
>
> Le Mon, 23 Nov 2009 09:21:16 +0100, BOTTIN Francois <francois.bottin@cea.fr>
> a écrit :
>
>   
> > 2) I suspect the diagonalisation to be responsible for the crash. 
> > Indeed, the current
> > implementation of ScaLAPACK is not satisfactory and shows a very poor 
> > scaling as a function of the number of band/FFT-processors. 
> >     
>
> I'm not using ScaLAPACK for the moment.
>   
Without using ScaLAPACK it is expected to be worse.
The diagonalisation within the subspace is then perfomed in sequential.
See for example p18 of:
http://wwwold.abinit.org/workshop_07/Presentations/Liege_Bottin.pdf

What I mean is: by using the present implementation of ScaLAPACK,
which gives a great improvement for subdiago.F90 with respect to LAPACK,
we don't scale very well as a function of processors.
This part may become a bottleneck.
I experience that on systems with 1000<nbands<2000 and 200<nprocs<500.

At this time we are trying to optimize the implementation in order to 
overcome this trouble.
> > PS: Yann, do you launch Abinit on 1000 band/FFT-cores? Or more?
> >     
>
> Not yet. I wanted first to start with a smaller system (~ 400 bands) and tune
> the convergence parameters, before going to the big stuff.
>
>   
OK. Good continuation...
> Best regards,
>
> Yann.
>   
Max and Yann, thanks for all your reports. Please, sent your experience 
about this point.
Perhaps it is not subdiago which is responsible for your trouble!
Thanks a lot,
Francois

--
##############################################################
Francois Bottin                    tel: 01 69 26 41 73
CEA/DIF                            fax: 01 69 26 70 77
BP 12 Bruyeres-le-Chatel         email: Francois.Bottin@cea.fr
##############################################################