[mpiwg-hybridpm] Feedback on init/finalize slides
james.dinan at gmail.com
Wed Aug 21 15:26:45 CDT 2013
- Slide 20: Jim wants "low-quality implementations" to be called "simple
implementations" to save face :-).
JMS 20 Aug 2013: Is this a real comment or a joke?
Yes, I actually asked for that change, for serious. :) For some networks,
never actually finalizing MPI/network when the init/finalize count reaches
zero is the only way to implement this. So, I wouldn't really consider
this to be a quality of implementation issue, because implementors have no
other other choice.
- Slide 26: There is some conflict between the notion of INIT/FINALIZE
being collective vs. they being synchronizing. We wanted to call them
collective, so it is illegal for some processes in an application to do 3
INITs/FINALIZEs and some processes to do 2. But we also didn't want all
calls to be synchronizing (specifically for the case of connect/accept). We
need to figure out how to phrase this. Maybe we completely get rid of the
notion of these calls being collective, and just say that the initializing
and finalizing calls be collective. Everything else is really local-only.
So this would be a valid program:
- P0 does INIT/INIT/FINALIZE/FINALIZE
- P1 does INIT/FINALIZE
JMS 21 Aug 2013: Good points. I therefore made the definition on slides
8 and 13 stronger: MPI_INIT that does not start an epoch / MPI_FINALIZE
that does not close an epoch is a local-only operation.
On slide 13, "INIT and FINALIZE are still collective ... but only when they
open / close an epoch" is not correct. Init and finalize are *always*
collective, but they may not be synchronizing. Are we over-specifying
here? Can we just call these collective and leave synchronization to an
On Wed, Aug 21, 2013 at 12:02 PM, Jeff Squyres (jsquyres) <
jsquyres at cisco.com> wrote:
> I see the feedback from the call that I missed on the wiki. I updated the
> slides accordingly, and added my responses:
> Note that I had some questions about the comments on slides 20 and 26
> (which are now 21 and 27); see my comments on the wiki.
> Attached are the latest slides with the feedback incorporated from the
> wiki page.
> Jeff Squyres
> jsquyres at cisco.com
> For corporate legal information go to:
> mpiwg-hybridpm mailing list
> mpiwg-hybridpm at lists.mpi-forum.org
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the mpiwg-hybridpm