The ABINIT Users Mailing List ( CLOSED )

[chabotj@courrier.umontreal.ca]

Dear Andriy

> Dear Developers of ABINIT,
>
> I have the following question to ask. What is the present status of the
> parallelization over bands in ABINIT?
>
>        The 'Features' file gives the following info:
> "B.1. For ground-state calculations, the code has been parallelized
> on the k-points, on the spins, on the bands,
> and on the FFT grid and plane wave coefficients."
>  From what is written above one would assume that parallelizations of all
> kinds should work equally well (with MPI).
>             However:
>  I was running the parallel version of ABINIT_3.4.3 on IBM-SP3 machine,
> but whatever I tried, the band parallelization wouldn't work.
> For example, in the case of 1 k-point
> all the load is carried by only ONE processor, while the others are idle!
>
> I used the following parameters in the input file, as documented in
> the list of input variables:
>    wfoptalg 1     # blocked CG algorithm minimization
>    nbdblock 8     # here, say, 8 - number of available processors

 nbdblock defines the number of bands in a block, not the available
 processors. To determine the number of processors you should respect the
 following formula:

 nproc = nkpt * nbdblock * nsppol
 nproc must be a multiple of the number of processor you want to use

 so if you want to use 8 processors for 1 kpt (assuming that nsppol = 1),
 you have to set nbdblock to a multiple of 8. but the nbdblock must also be
 a multiple of nband.



> I played around with different values of <nbdblock>, being integer
> divisors
> of the total number of processors, but it had no effect on the run.
> In fact, the sequential version of ABINIT did better job in all cases!

 this is it, but for the nband and not the number of processors

 I wish it would help you

 Jean-Francois Chabot

> Am I missing something essential, which is why the parallel version
> doesn't work, or is there any other reason for this?
> I would be very grateful for your help.
>
> Yours Sincerely,
> Andriy Nevidomskyy
>
> P.S> The parallelization over k-points seems to be working fine.
> Unfortunately only when nkpt=integer*number_of_processors.