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)" <[email protected] on behalf of [email protected]> 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 [email protected] For corporate legal information go to: http://www.cisco.com/web/about/doing_business/legal/cri/
_______________________________________________ mpiwg-sessions mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-sessions