-
Notifications
You must be signed in to change notification settings - Fork 1
2021 08 23 webex
#08/23/21 webex notes
Attending: Howard Pritchard, Thomas Hines, Joseph Schuchart, Dan Holmes, Isaias Alberto Com, Martin Schreiber, Martin Schulz
- Restart discussions where we left off on dynamic process sets and RT/app interaction use cases.
Resume discussion of notes accompanying the miro docs. Start back with an option for a query function to get group of processes that are in the old and new pset version - but for MPI 4.0 we would not allow that. Isaias asks if we can use something more out of band with an RM or add something else that doesn't require the application to use group operations to determine deltas in psets between sessions. Or could we add new group methods? Lot of burden on the users.
Another option would be to add a handler which would be invoked when there's a change in psets associated with a session.
This would be as an alternative to calling session init/get psets/session finalize (aka polling approach)..
Possible info key for this callback function - or when registering the callback function to provide constraints to the runtime as to
when to invoke the callback.
Some delta for providing info back from RM about psets:
- last pset created
- any pset(s) created
- all changes?
Could this be encapsulated in an info object - psets as keys, values are reasons for change. Should a session handle encapsulate the delta? Martin Schulz and Dan came to the conclusion no this isn't a good idea. Discuss limitations of info object, max val in particular may limit its use.
In a nutshell for finding delta in psets
- use the group concept, but will need to relax group comparison constraints - allowing groups from different psets
- some other api completely.
Examples of handler registration straw man proposal:
- change to any pset invoke the handler (wildcard version perhaps)
- monitor a particular pset
At the end of Martin Schulz and Dan's conversation, they captured ideas from the google docs as text in the miro diagrams either around the arrows or a sideblock.
Isaias mentions probabilistic scheduling - Ralph Castain seems to think this has promise.