You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Howard Pritchard edited this page May 20, 2024
·
2 revisions
Agenda - 5/06/24
Sonja Happ and Simon Pickartz (ParTec) present some work they have been doing with Sessions
Change log PR
other
Discussion
ParTec presentation
Sonja Happ goes through a presentation about their work - malleability with MPI Sessions
based on Parastation MPI (currently based off MPICH 4.1.2)
notion of varying # of MPI ranks over time for a complex workflow/multi-state application.
implemented via proposed MPIX_Session_reinit function. MPIX_Spawn_async to start new processes (possibly after a shrink).
Focus on case of active maleability - driven by application not by an external agent. Think passive could be built
on top of this - requires some sort of RM interface to MPI runtime system.
Testing concepts at Juelich.
Tony points out that this model does force kind of a "world" like behavior in that reinit is effectively a synchronization point.
Discuss MPIX_Comm_async. Howard and Tony think this can be split out as a separate proposal. Tony says by adding some
pset notion to this method some of the reinit complexity could be avoided.
Dominik asks whether this spawn functionality could be expressed via the pset expansion/shrinkage approach the WG has been
considering.
Briefly discuss cancellation of allocation requests. This has not yet been pushed to the standard.
TODOS:
[x]: Sonja will get her slide deck vetted and send to Howard. He will add to the Sessions WG wiki