MPI groups and communicators

[sawyl]

Investigating a problem with one of my MPI codes — it abended with a communicator error when I increased the number of CPUs — I uncovered an interesting difference between groups and communicators: it seems that groups are defined on all processors, regardless of whether they're members of the group, whereas communicators are only valid on processors within the group. Thus my code, which checked the validity of the group rather than the communicator, kept crashing in MPI_Comm_free() whilst attempting to free MPI_COMM_NULL because it foolishly tried to free up memory on all processors, rather than just group members.

When you think about it, it's actually kinda obvious why it works the way it does: the invalid communicator provides an easy test for group membership.

Comments

[doctor-squale.livejournal.com]

No no no it does not make sense to me what you are saying.
Communicators operations are defined within the calling universe.
MPI_Comm_free is necessarily called on a superset of the processors
of the group.

Or maybe I did not understand what you're saying.

email me your test case.