I checked MPICH2, and it doesn't invoke the attribute destroy functions in the order required here. It wouldn't be hard to change, but I haven't had a chance to do that. Bill On Apr 2, 2009, at 3:26 PM, MPI Forum wrote:
#71: Add routine for registering callback when finalize occurs ------------------------------------- +-------------------------------------- Reporter: koziol | Owner: koziol Type: Enhancements to standard | Status: assigned Priority: Waiting for reviews | Milestone: 2009/04/06 Chicago Version: MPI 2.2 | Resolution: Keywords: | Implementation: Waiting ------------------------------------- +--------------------------------------
Comment(by jsquyres):
Other than missing a changelog, this proposal looks good (read: we need a changelog entry ASAP before we can discuss it next week in Chicago!).
Rajeev -- I think the issue with enforced ordering was inspired by potentially using multiple different middleware packages that may not be able to coordinate their shutdown callbacks. E.g., if package A uses package B which uses MPI, then you ''need'' to have B's finalize callback called before A's. A and B would be hard-pressed to coordinate this if they didn't know about each other. And so on.
FWIW, Open MPI does ''not'' invoke COMM_SELF callbacks in any particular order -- it'll actually be a little painful to convert to assign an order (because we use a hash table to store this stuff). But it's all just bookkeeping -- it's certainly do-able. So someone needs to do an open- source implementation.
Bill -- did you get a chance to see if MPICH2 destroys COMM_SELF attributes in order?
-- Ticket URL: <https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/71#comment:17
MPI Forum <https://svn.mpi-forum.org/> MPI Forum
William Gropp Deputy Director for Research Institute for Advanced Computing Applications and Technologies Paul and Cynthia Saylor Professor of Computer Science University of Illinois Urbana-Champaign