The ABINIT Users Mailing List ( CLOSED )

[gilles zerah <gilles.zerah@wanadoo.fr>]

Dear Maurice,

The best distribution of npbands and npfft depends presumably upon the 
implementation of MPI you are using and of many
other parameters. Now, it should work for this distribution, but the 
case npfft=1 is treated differently.
Is it possible you post your input file?

Gilles
> Dear all,
>
> I noticed strange behavior while using the bands/fft parallelization 
> features of abinip (compiled with -DMPI_FFT preprocessor flag),
> running on an SGI Altix system with 16 CPU´s.
>
> The calculation in question involves a geometry optimization using 
> only the gamma point:
>
> 1) The first calculation was run using the parameters:
>
> npbands=8, npfft=2
>
> This calculation runs fine and the geometry optimization converges 
> without any trouble.
>
> 2) Then I tried a different distribution, using the parameters
>
> npbands=16, npfft=1
>
> According to the documentation, this should be more efficient than the 
> first choice.
>
> But while running this calculation something strange happens. After a 
> while, only a few
> of the 16 CPU´s are actually processing, with the remainder in 
> "uninterruptable sleep".
> This already occurs during the first SCF cycle, effectively 
> interrupting the progress of
> the calculation.
>
> Any ideas?
>
> Cheers,
>
> Maurice