[mpiwg-sessions] [EXTERNAL] Notes from today / slides for next time

Jeff Squyres (jsquyres) jsquyres at cisco.com
Tue May 17 09:57:04 CDT 2016


Put simply: your use case is grow then shrink, right?

The case I show in slides 83-91 is just shrink.  But I think it generalizes to any case of shrink (i.e., regardless of whether you previously grew).  I think that the main issues are:

- previously-obtained sets may become stale
- the release manager may want to (implicitly or explicitly) release nodes that are no longer being used (this may or may not be in MPI's prevue)



> On May 16, 2016, at 6:31 PM, Teranishi, Keita <knteran at sandia.gov> wrote:
> 
> Jeff,
> 
> Thanks for the update. Our potential use case in Sandia is less aggressive than the model in pages 83-91. Typically, the main application occupies the 70-90% of the total nodes for the entire execution time.  The application data analytics routines are spawned and joined (shrunk) on the rest of the nodes.  Data movement happens between the ranks running base application and the spawned ranks, but I believe that doing within MPI is faster than using external fast-IO services (such as DataSpaces).
> 
> Maybe, the best use case I can map to your example is reduced order modeling, which can be seen as interactions between “big” and “small" simulation instances.  Of course, it can be done with master-worker model, but spawn+shrink (join+leave) model would allow more flexible use of big machines.
> 
> Thanks,
> Keita
> 
> 
> 
> 
> 
> On 5/16/16, 3:00 PM, "mpiwg-sessions on behalf of Jeff Squyres (jsquyres)" <mpiwg-sessions-bounces at lists.mpi-forum.org on behalf of jsquyres at cisco.com> wrote:
> 
>> Thanks for everyone's time today.  Here's notes and slides from today:
>> 
>>   https://github.com/mpiwg-sessions/sessions-issues/wiki/2016-05-16-webex
>> 
>> I consolidated the open issues, and added a bunch of information about the "grow" and "shrink" use cases that we want to support, and made slides for 2 weeks from now -- please have a look:
>> 
>>   https://github.com/mpiwg-sessions/sessions-issues/wiki/2016-05-30-webex
>> 
>> Hilights:
>> 
>> • slides 14-15: required/provided for thread level
>> • slide 20: resolved -- only static sets (no dynamics)
>> • slide 29: only get set names for which you are a member
>> • slide 30: Updated the "open questions"
>> • slides 78-70: minor updates to MPI_EXEC
>> • slides 83-91: A bunch of new slides about "shrink the app" use case
>> 
>> -- 
>> Jeff Squyres
>> jsquyres at cisco.com
>> For corporate legal information go to: http://www.cisco.com/web/about/doing_business/legal/cri/
>> 
>> _______________________________________________
>> mpiwg-sessions mailing list
>> mpiwg-sessions at lists.mpi-forum.org
>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-sessions
> _______________________________________________
> mpiwg-sessions mailing list
> mpiwg-sessions at lists.mpi-forum.org
> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-sessions


-- 
Jeff Squyres
jsquyres at cisco.com
For corporate legal information go to: http://www.cisco.com/web/about/doing_business/legal/cri/



More information about the mpiwg-sessions mailing list