MPI 4.1 - opaque data handles and state preservation across sessions

[hppritcha]

Some wording went in to the MPI 4.1 standard that is potentially impactive on the way Open MPI currently handles MPI data types and Session finalize.

See #12193
and entry 7 in section B1.1 of the MPI 4.1 standard

[hppritcha]

I am not sure we really need to do anything concerning this change log entry. The text referenced in the change log is section 5.1.9 page 156 of the MPI 4.1 standard

The optimizations chosen during MPI_TYPE_COMMIT may no longer be optimal if a
session (or the World Model) is initialized or finalized. (End of advice to implementors.)

Some in the forum believe this implies a datatype that has been committed may persist after all sessions are finalized and if the world process model is used, MPI_Finalize is called. If we consider "optimal" though to mean that the data type won't be valid if a session is finalized/WPM is finalized then we don't need to do anything. I'm of the opinion that we document the behavior of Open MPI somewhere and consider this done.

[devreal]

In my mind, this requires user-created datatypes to take a reference on the internal session so that the datatype survives the end of all sessions and the WPM and can still be used in a new session. Just like info objects (IIRC). The "optimization" point is a distraction.

[hppritcha]

What is the internal session?

[hppritcha]

Also there is no type destructor function call allowed after last session finalize

[devreal]

Sorry, I wasn't clear: I meant that each user-created datatype needs to retain a reference to the OMPI instance (via ompi_mpi_instance_retain()). It would seem that not being able to destroy an object that should outlive all sessions after it has outlived all sessions is an oversight...