[Wien] some comments on parallel execution of wien2k

[Robert Laskowski]

Hi,
I agree iatm_proc_pack is "color" for MPI_Comm_split and according to the specification it should be non negative.
The idea behind the code around line 319

     if (jatom.le.nat.and.myid_vec.eq.0) then
        iatm_proc_pack=jatom_pe
     else if (jatom.le.nat) then
        iatm_proc_pack=nat+npe+jatom_pe
     else
        iatm_proc_pack=-(nat+npe+jatom_pe)
     endif

is to extract the first processors from suncommunicators used to split eigenvectors.
Only communicators that comes from  iatm_proc_pack=jatom_pe are used latter in the code,
but the rest must be separated. I suggest to change this part of the code to:

     if (jatom.le.nat.and.myid_vec.eq.0) then
        iatm_proc_pack=jatom_pe
     else if (jatom.le.nat) then
        iatm_proc_pack=nat+npe+jatom_pe
     else
        iatm_proc_pack=nat+npe+jatom_pe + npe       !!!! change
     endif

I do not have openmpi installed so please test it.

regards

Robert

 


On Tuesday 05 January 2010, Andres Saul wrote:
> Dear wien users:
>    I can also confirm that there is a problem with lapw2_mpi and
> openmpi for some specific combinations of
> the number of atoms and number of processors. The same calculations
> with mpich or hpmpi work fine.
>    Looking at the source files I concluded that the problem was
> related to using a negative second argument
> in MPI_COMM_SPLIT in line 327 of l2main.F (wien2k v9.1).
> 
>    I have found that a simple dirty solution consisting in changing
> line 319 from
>          iatm_proc_pack=-(nat+npe+jatom_pe)
> to
>          iatm_proc_pack=10000-(nat+npe+jatom_pe)
>    was enough to solve the problem. I my opinion the occurrence of the
> error for a given number of atoms or specific number of processors is
> related to the "if then else" structure around this line.
> 
> Regards
> Andres Saul
> _______________________________________________
> Wien mailing list
> Wien at zeus.theochem.tuwien.ac.at
> http://zeus.theochem.tuwien.ac.at/mailman/listinfo/wien
> 
> 



-- 
Dr Robert Laskowski

Vienna University of Technology, 
Institute of Materials Chemistry, 
Getreidemarkt 9/165-TC, A-1060 Vienna

tel. +43 1 58801 15675
Fax  +43 1 58801 15698