[Mpi3-hybridpm] Better Forum INIT/FINALIZE slides

Steve Oyanagi sko at cray.com
Tue Jul 16 22:25:00 CDT 2013


On 07/15/2013 12:45 PM, Jeff Squyres (jsquyres) wrote:
> Here's a better cut of slides that are more Forum-friendly.
>
> Comments welcome.
>
> (Pavan: I'm correct in thinking that you're getting plenary time from Martin for this, right?)
>

Hi,

Comments.  Not sure how helpful they are.  Feel free to ignore.


   1. Slide 2 - Please describe Jeff Hammod's problems rather than point 2.

   2. Slide 5 - Might be more understandable if you leave
      "- New defs based on inside/output epoch" off this slide.

   3. Slide 9 - maybe obvious, but point 3 should perhaps say:
      "Essentially a no-op if MPI is already initialized (except for
      increment of internal ref count)"

   4. Slide 10 - Can the thread level change at the beginning of each epoch?
      Is this optional or required? (Later slide 20 shows that the thread
      level does not need to change per epoch).

   5. Slide 17 - Why are Initialized/Finalized not thread safe?  Rational
      behind this decision.

   6. Slide 20/21 - Not sure how, but I think you want to mention earlier
      the information on this page.  The obvious assumption up to now
      is that the suggested change requires implementations do a real
      finalize at the end of each epoch and a full initialize at the start
      of each epoch.  Also the thread level change issue for slide 10.

   7. Not sure what slide 22 means.  Maybe a "for example..."
      This might be me and my lack of exposure to different MPI
      implementations.

   8. Slide 27 contradicts slide 17 that says Initialized and Finalized are
      *not* thread safe.  Slide 27 says "Must obey MPI thread level" implying
      thread safe for thread multiple, but slide 17 says "Do NOT make these
      functions thread safe".

Hope this was helpful.
         - Steve



More information about the mpiwg-hybridpm mailing list