[Mpi3-ft] RTS - 4th Draft

Josh Hursey jjhursey at open-mpi.org
Wed Jun 22 14:00:48 CDT 2011

I just uploaded the 4th draft version of the RTS proposal to the wiki
that reflects the notes from todays teleconf. Change log at bottom.

On the teleconf we decided to meet again on Monday, June 27 at 11:30
am EDT - in addition to meeting on Wednesday, June 29 at the regular
time. The document should be ready to ship by the Wednesday meeting.
We stopped at section 17.9 on todays teleconf.

A couple of issues came up on the teleconf that other might want to comment on:
 - In Section 17.6.2, we need to cleanup the wording here a bit to
make it clear about the possible impact on implementations. In
particular, does this require that all MPI operations have a check for
initialized in them? and how could that check be used to provide the
necessary semantics.
 - Folks were wondering about what we should say regarding how long
MPI_Comm_kill should block considering it is a non-collective
operation. I added a sentence to the document, not sure if it is good
or not but it is a place to start. Anyone have other suggestions?


Change Log
 * 17.5.2: Add forward reference to collective chapter in
MPI_Comm_validate_all explanation of collective active.
 * 17.6:   Minor wording cleanup
 * 17.6.1: Tense correction on 'error classes'
 * 17.6.2: Loosen up the wording a bit 'should try' and add an advice
to implementors regarding critical errors.
 * 17.6.3: Clarify that this is a local operation (non-collective).
Also clarify 'process group'
 * 17.6.3: If the MPI_Comm_kill operation cannot provide the necessary
semantics then it should return MPI_ERR_UNSUPPORTED_OPERATION.
 * 17.6.3: Add a first try at some language about how long
MPI_Comm_kill will block - must block until locally finished, may
block until globally finished.
 * 17.7.1: 'effect' -> 'affect'
 * 17.7.1: Slight clarification improvement to the use of
MPI_Comm_validate_set_state_null on an inactive persistent request.
 * 17.7.1: 'effect' -> 'affect'
 * 17.8:   Clarify advice to implementors
 * 17.8:   Highlight 'undefined' for buffers.
 * 17.8.1: 'therefor' -> 'therefore'
 * 17.8.2: Re-order the sentence so it reads a bit better
 * 17.8.3: Add back reference to Validation operations regarding
overlapping collectives with validate_all.
 * 17.8.4: Bcast Example - Add citation for 'recovery blocks'
 * 17.9.2: Add missing word ("group") to 'Rationale'
 * 17.9.2: Clarify that MPI_Comm_is_collectively_active is a local operation.
 * 17.12.1: Add back reference to window validation operations to
advice to users.
 * 17.13:   Put the question in an 'e.g.,' parenthetical.
 * 17.13.1: Add a bit more clarification to Rationale regarding
semantics for MPI_File_close.

Joshua Hursey
Postdoctoral Research Associate
Oak Ridge National Laboratory

More information about the mpiwg-ft mailing list