[Mpi3-tools] Meaning of meaning of unique_id in mqs_communicator?

Geoffrey R. Paulsen geoffrey.paulsen at hp.com
Mon May 11 09:54:20 CDT 2009


  In reading HP-MPI source code, which has implemented the MSQ interface
quite a while back, I believe that we return a value which conforms to
the definition (b) below.

  I'm not sure I understand the difference between (b) and (c).

  I hope this helps,
  Geoff Paulsen

Regarding Ashley Pittman's message:
> All,
> I've raised this point before on the anl mpi-debugger mailing list
> however I'd still like resolution to the issue so I hope it's not
> in-appropriate to raise it again in this forum.
> The problem I have is that when inspecting a parallel job in a debugger
> I find I'm unable to match communicators across MPI processes so it's
> impossible to tell which communicators in process A match which
> communicators in process B.
> Each MPI process reports the existence of a number of communicators to
> the debugger and through the unique_id field of the mqs_communicator
> type it reports a ID for this communicator.  It's the meaning of this ID
> which I'd like clarified.
> There are three possible interpretations for this field.
> (a) The ID could be unique to this process, typically a pointer to some
> internal struct.
> (b) The ID could be unique to this process and also common to all
> members of the communicator, regardless of the MPI process on which they
> reside.
> (c) The ID could be both unique to this process, common to all members
> of the communicator and also unique to this communicator.
> Current MPI implementation's seem to favour (a) however this doesn't
> give the debugger enough information to match communicators in one
> process with communicators in a different process.
> I believe the intention of the code when written was (b) and this does
> allow the debugger to match communicators across processes.
> This issue presents me with real problems, I have a collective state
> inspection tool which is capable of detecting deadlock and spotting
> errant processes however without the ability to match communicators
> across processes it's functionality is severely crippled.
> If possible could we discuss this on the call with a view to clarifying
> the current meaning and if necessary devising other means of exporting
> the required information.
> Ashley Pittman,
> _______________________________________________
> Mpi3-tools mailing list
> Mpi3-tools at lists.mpi-forum.org
> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-tools

More information about the mpiwg-tools mailing list