[Mpi3-tools] Question about communicator for control variables

[Martin Schulz]

Hi Andreas,

On Oct 29, 2010, at 12:59 PM, Andreas Kuntze wrote:

> Hello all,
> 
> in the current MPIT draft, the MPIT_CTRLVAR_WRITE function takes a 
> communicator as an argument. If I understood the description correctly, you 
> have to pass the communicator returned by MPIT_CTRLVAR_GETINFO. Can someone  
> explain to me, for what reason you have to pass it, if you can execute 
> collective write operations on one communicator only for each control 
> variable?

The idea was that the control variable set operation can have
a global meaning, in which case it has to be set on all processors
of a certain communicator. Passing this communicator into the
routine made that explicit (I agree, we could have just made this
implicit by using the communicator that is returned by the info
call).

We discussed this at length at the last face-to-face meeting and
actually decided to drop it. The problem that many saw in this
was that communicators don't exist before MPI_Init, but we want
to be able to call this routine at that time. The compromise was
to still return a flag during the info call that says that a variable
has a global/collective meaning, but put the burden on the
user to ensure that the variable is set consistently across all
tasks.

This change (along with the other comments I got in the last few
meetings), will be reflected in the next MPIT draft.

Martin


> 
> Thanks,
> 
> Andreas
> _______________________________________________
> Mpi3-tools mailing list
> Mpi3-tools at lists.mpi-forum.org
> http://BLOCKEDlists.mpi-forum.org/mailman/listinfo.cgi/mpi3-tools
> 

________________________________________________________________________
Martin Schulz, schulzm at llnl.gov, http://people.llnl.gov/schulzm
CASC @ Lawrence Livermore National Laboratory, Livermore, USA