[mpich-discuss] [EXTERNAL] Re: Calling MPI_Finalize or not

Raffenetti, Kenneth J. raffenet at mcs.anl.gov
Sun Feb 21 16:35:04 CST 2021


Yes, that is correct. Spawned processes should be sure to call MPI_COMM_DISCONNECT on their parent communicator before MPI_FINALIZE.


On 2/21/21, 4:13 PM, "Mccall, Kurt E. (MSFC-EV41) via discuss" <discuss at mpich.org> wrote:

    So are you saying that the worker can call MPI_Finalize and it won’t wait for the manager that spawned it to do the same?   That would be ideal.
    From: Zhou, Hui <zhouh at anl.gov> 
    Sent: Sunday, February 21, 2021 4:07 PM
    To: discuss at mpich.org
    Cc: Mccall, Kurt E. (MSFC-EV41) <kurt.e.mccall at nasa.gov>
    Subject: [EXTERNAL] Re: Calling MPI_Finalize or not
    The worker should call `MPI_Finalize` before exit. If somehow it hangs, assuming all the resources have been freed/disconnected, it must be a bug, please file an issue and we’ll track it down. Each spawned process has a different `MPI_COMM_WORLD` from the process who spawned it.
    Hui Zhou
    From: Mccall, Kurt E. (MSFC-EV41) via discuss <discuss at mpich.org>
    Date: Sunday, February 21, 2021 at 3:06 PM
    To: discuss at mpich.org <discuss at mpich.org>
    Cc: Mccall, Kurt E. (MSFC-EV41) <kurt.e.mccall at nasa.gov>
    Subject: [mpich-discuss] Calling MPI_Finalize or not
    Two questions:
    I have a group of manager processes, each of which uses MPI_Comm_spawn() to create individual worker processes one at a time.   
    Occasionally, a worker has to exit and is replaced by another worker via MPI_Comm_spawn().    If the exiting worker calls MPI_Finalize(), 
    will it sit there waiting for all other managers and workers in MPI_COMM_WORLD to do the same, or will it wait only for the manager 
    that created its inter-communicator to call MPI_Finalize?
    In either case, maybe it is better for the worker to NOT call MPI_Finalize at all so that it doesn’t sit there taking up a core’s resources 
    while it waits for the others.   What are the implications of not calling MPI_Finalize in the workers?

More information about the discuss mailing list