Hi all,
We had a “reading” of PR #50 (
https://github.com/mpi-forum/mpi-standard/pull/50) about error behavior during MPI_INIT and MPI_FINALIZE during the December meeting. There were a few things that came out of that which we need to discuss to decide next steps here.
Some of this is probably simple and just needs to be added to the text:
- Add MPI_ERROR_CLASS and MPI_ERROR_STRING to the list of functions callable outside of INIT/FINALIZE
The other is a more fundamental shift of the proposal to move specifying the error behavior from mpiexec to MPI_INFO_ENV (which obviously requires allowing info objects to be manipulatable outside of INIT/FINALIZE). It might still be a good idea to have command line flags to set these too.
- default_errhandler - An info key that defines the default error handler for all situations unless replaced by another.
- This would include situations currently handled by MPI_COMM_SELF such as errors that have no other place to be raised.
- This can only be MPI_ERRORS_ARE_FATAL or MPI_ERRORS_RETURN to simplify the requirement of a “universal error handler”
- predefined_errhandler - Attached to all predefined objects when they are created.
- This cannot be changed after MPI_INIT because the communicators would have already been created. If their error handlers need to be changed after that, change them directly.
- If this isn’t defined, it would fall back to default_errhandler.
Do people agree with these changes? Should we move forward with updating the proposal to reflect them?