-
Notifications
You must be signed in to change notification settings - Fork 1
2020 05 18 webex
#5/18/20 webex notes
Martin has added some use cases to the slides (https://github.com/mpiwg-sessions/sessions-issues/wiki/SessionsV2-ideas.pptx).
- Individual applications with changing resources
- Coupled application(s) (components)
- (Partly) Co-located independent applications - tools, viz, etc.
- System software to schedule independent applications - e.g. Slurm using Open MPI itself for IPC?
Martin briefly recapped what we covered at the 5/4 meeting since we have some new attendees. Martin defines a bubble - not something explicit in the current Sessions model. Its more like MPI entities associated with a given Session - communicators, file handles, rma window handles, etc. The green things in the slide deck illustrate the idea.a Using the slide entitled "Scenario: Growing (cooperative/coercive)" for this discussion.
Resume on the slide entitled "Scenario: Shrinking (Selfish)". Moved quickly to the coercive case. Discuss example where Process 3 fails. This is broaching somewhat the ULFM model. Dan points out we're not incompatible with ULFM. Bubbles approach is a build-up rather than shrink-down model. There's discussion of the Session error handler - could this be used to handle this situation in terms of cleaning up the existing session and its resources and starting over with just 2 processes. We need to clarify this. Discuss case where the RM gives us a new process slot so we don't have to shrink - "Scenario: Replace (Coercive)/FT or forced Migration". Discuss how to know if a process was "restarted" in the MPI Reinit sense. Can this be done using process set versioning?Need to agree on process set version - almost certainly - but how to do this? Can we handle moving to the new bubble using local operations? Dan says we need to come up with a set of steps that the app needs to do and expected RT responses. Three classes of processes - those that were in old bubble, new process, dead process. See the "Options" slide. We'll work on this slide at the next meeting.
- Martin will update slides