Re: 6.5.2 Intercommunicator collective operations
William Gropp (gropp@mcs.anl.gov)
Tue, 04 Feb 1997 11:12:45 -0600
| The current proposal for rooted intercommunicator collective operations is
| unclear.
|
| Consider the broadcast operation defined on page 124. The processes in group B
| are all receiving data from the root (in group A). This is locally known since
| the communicator is an inter-communicator, and the special (MPI_ROOT) value is
| NOT passed as the root. The root process in group A is sending data to all the
| processes in group B. This is locally known since the communicator is an
| inter-communicator, and the special (MPI_ROOT) value IS passed as the root.
| How do the other processes in group A know that they should not participate?
| What should their root value be?
|
| I believe the original intent here is that ONLY the root process in group A
| calls the MPI_Bcast. Thus we redefine the semantics of collective (must be
| called by all processes in group B, and ONLY the root in group A.)
|
| This ugliness must either be repaired or detailed in the document.
I have added the text that makes this clear. As an alternative, I've offered
up for sacrifice MPI_NOT_ROOT, which all processes in the group that the root
belongs to would provide as the root argument, leaving the call collective.
Bill