[Mpi3-hybridpm] [MPI Forum] #357: thread-safety of MPI_INITIALIZED, MPI_QUERY_THREAD and MPI_IS_THREAD_MAIN in all uses for library correctness

[Jeff Hammond]

MPI_INIT and MPI_FINALIZE are collective, are they not?  How does
reference counting help to make my example code correct in the case
where only a subset of the processes in world are going to use the
threaded library?

Jeff

On Tue, Apr 9, 2013 at 1:20 PM, MPI Forum <mpi-forum at lists.mpi-forum.org> wrote:
> #357: thread-safety of MPI_INITIALIZED, MPI_QUERY_THREAD and MPI_IS_THREAD_MAIN
> in all uses for library correctness
> -------------------------------------+-------------------------------------
>                   Reporter:          |                          Owner:
>   jhammond                           |                         Status:  new
>                       Type:  Text    |                      Milestone:
>   (only) changes                     |  2013/03/11 Chicago, USA
>                   Priority:  Forum   |                     Resolution:
>   feedback requested                 |          Implementation status:
>                    Version:  MPI     |  Unnecessary
>   <next>                             |            Author: Rich Graham:  0
>                   Keywords:          |        Author: Torsten Hoefler:  0
>   threads                            |  Author: Jesper Larsson Traeff:  0
>         Author: Bill Gropp:  0       |             Author: David Solt:  0
>         Author: Adam Moody:  0       |          Author: Rajeev Thakur:  0
>      Author: Dick Treumann:  0       |      Author: Alexander Supalov:  0
>     Author: George Bosilca:  0       |
> Author: Bronis de Supinski:  0       |
>       Author: Jeff Squyres:  1       |
>  Author: Rolf Rabenseifner:  0       |
> -------------------------------------+-------------------------------------
>
> Comment (by balaji):
>
>  Jeff,
>
>  We had some discussion on this on the last Hybrid working group (we were
>  actually discussing the ref-counted Init/Finalize ticket and the
>  discussion moved to include this ticket as well).
>
>  We thought that the init/finalize issue (which includes MPI_INITIALIZED)
>  is different from the MPI_QUERY_THREAD and MPI_IS_THREAD_MAIN issue.  The
>  former is perhaps better handled by the ref-count ticket, though that has
>  its own set of concerns and might not fix everything.  The latter is with
>  respect to how a thread can identify whether it can make MPI calls or not
>  (which, unfortunately, requires it to make the MPI calls MPI_QUERY_THREAD
>  and MPI_IS_THREAD_MAIN.
>
>  Perhaps we should tackle these two problems independently to start with,
>  while making sure the solutions make sense as a whole?
>
> --
> Ticket URL: <https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/357#comment:8>
> MPI Forum <https://svn.mpi-forum.org/>
> MPI Forum



-- 
Jeff Hammond
Argonne Leadership Computing Facility
University of Chicago Computation Institute
jhammond at alcf.anl.gov / (630) 252-5381
http://www.linkedin.com/in/jeffhammond
https://wiki.alcf.anl.gov/parts/index.php/User:Jhammond