22Mar2017 - openpmix/openpmix GitHub Wiki

Mar 22, 2017

Ralph Castain, Intel.
Aurelien Bouteiller, UTK
Geoff Paulsen, IBM
Josh Hursey, IBM
Alexandre Eichenberger, IBM Howard Pritchard, LANL David Bernholt, ORNL Geoffroy Valley, ORNL (several others - my apologies for missing the names)

We started with a brief review of the last meeting's minutes and where we last left off. We then called up the PMIx RFC for cross-model coordination and worked our way through it.

There seemed to be general agreement with the overall approach in the draft. Alexandre pointed out that OpenMP actually does have an "init" function - it just isn't explicitly called, but rather gets executed on first use. So no barriers to the proposed approach were identified.

Geoff raised a point about passing down any info available to the RM - e.g., if the user told the RM that the application would utilize certain programming models. PMIx already brings down all the RM known info about the application, and so this could be included. At the meeting, we felt that we would always have to be able to dynamically declare anyway as we cannot know that the user told the RM, and so the potential benefit wasn't clear. We left it as a future consideration.

It was noted that PMIx would now need two features:

• PMIx already supports multiple calls to PMIx_Init from inside a process, but it ignores the "info" arguments from all calls after the first. This would need to change as the info will contain the new directives. Questions were raised about how to handle conflicting and/or overlapping directives. Some head scratching is required.

• event notifications must now be delivered across threads within the same process. Need to add a "proc_local" range for this purpose, and extend the event notification tracking system to handle multiple handlers registered to the same status code.

These will be added to the RFC.

1. Ralph offered to create a strawman set of API definitions based on conversations so-far and circulate it for comment purely as a "stake-in-the-ground" to make the discussion a little more concrete [DONE]
2. Ralph agreed to send to the mailing list some links to background on PMIx [DONE]
3. Alexandre agreed to do the same for OpenMP, and provide an overview presentation at future meeting [presentation provided]

1. Geoffroy to arrange meeting with SNL "RM" project to discuss possible collaboration on APIs for coordinating their "worker pool" of threads

1. Everyone will review the RFC and provide comments!
2. Ralph will continue working on the document to reflect the discussion, provided comments
3. Ralph will create a PMIx branch and begin updating it per the RFC
4. Alexandre offered to create a branch and update the OpenMP runtime to support the RFC
5. Aurelien volunteered to send a link to a suggested prototype application
6. Howard agreed to also look for something suitable - anyone else who would like to offer an application is welcome to do so!