RFC: should we document local_team-is-always-contiguous-in-world behavior?

Currently our implementation will only ever create a local_team with members that have contiguous ranks in world. If the processes in any GASNet-EX shared-memory "neighborhood" are not numbered contiguously in world then one gets a fatal error by default, or singleton local_teams if configured using --enable-discontig-ranks.

This Request For Comment (RFC) is to discuss whether we should document the contiguous ranks behavior, essentially entering into a contact saying that we do not intend to ever change it.

Our implementation takes advantage of the contiguous rank numbering of local_team members, and I suspect the assumption is pervasive enough that we'd not likely ever implement anything else. It would not surprise me if some/many/most users of local_team are also making this assumption. If they are not, then documenting this behavior would allow then to start doing so

I am entering this as a documentation issue in the implementation repo (not the spec) under the assumption that we'd document this in docs/implementation-defined.md. However, there is also to possibility that we'd elect to elevate this to the specification.

Thoughts?

Comments (3)