-
Notifications
You must be signed in to change notification settings - Fork 1
2025 09 22 webex
AI can make mistakes. Review for accuracy.
The team discussed upcoming presentations and scheduling for various MPI-related events, including the MPI Forum and Euro MPI meeting, with specific focus on working group activities and session attributes. They explored potential features and improvements for MPI, including parallel-in-time methods, spawn functionality, and object attribute listings, while considering the need for a strategic roadmap for MPI 6 development. The group also addressed concerns about complexity in the API and discussed scheduling a sessions working group breakout to further coordinate efforts. Next steps
- Howard to ping Wesley Bland about scheduling the Sessions Working Group meeting at the MPI Forum and share the information with the group.
- Tony to set up a Zoom link for the Sessions Working Group meeting once Howard confirms the time.
- Howard to prepare slides for the MPI BOF presentation at the Forum, including content about dynamic process and sessions attributes.
- Sonja to prepare materials about the refactored SPOM mechanism for a future meeting before Supercomputing.
- Howard to prepare slides for the plenary session about sessions attributes at the MPI Forum.
- Sessions Working Group to consider developing a roadmap for MPI 6 as suggested by Tony.
- Sonja to prepare notes about ParTech spawn functionality for future discussion.
- Dominik to check his vacation schedule to see if he can join the MPI Forum meeting online.
Howard discussed the upcoming MPI BOF at the Forum, where they will have 10 minutes to talk about the session's working group activities, including the dynamic process set ideas presented by Dominic on the 10th. Howard mentioned he would like to present a refactored SPOM mechanism that NPI already uses as a potential quicker feature to get into the standard, and invited Sonja to prepare something or use a session working group meeting before the supercomputing event to discuss this further. Howard also confirmed that Tony, Dominik, Rajat, and Raja Bhatari would attend SC, with Rajat attending for the first time. MPI Forum and Euro MPI Planning
The group discussed the status and applications of a feature related to parallel-in-time methods, noting that while some work continues, it is not limited to this specific application. Howard mentioned plans to present on adding session attributes to the next MPI standard at the upcoming MPI Forum meeting, though the presentation may be combined with Dominic's. The team also discussed scheduling for the upcoming Euro MPI meeting, with Howard planning to coordinate with Wes regarding the working group's time slot, which they requested to be scheduled in the morning for participants in Eastern Time zones. MPI 6 Strategy and Scheduling
The group discussed scheduling for an upcoming meeting, with Tony confirming his attendance for Tuesday and Wednesday, and Howard agreeing to send out options for the end of the week. They also discussed the need for a strategy for MPI 6, as suggested by Tony after his meeting with Martin Schultz, which would involve creating a roadmap for important problems to solve and potentially incorporating fault tolerance concepts. The group agreed to consider this strategy for MPI 6 development. MPI Spawn Functionality Usage Discussion
The discussion focused on the underutilization of MPI's spawn functionality, which Tony attributed to external policies rather than inherent limitations, sharing an example of a system that worked on TAC but failed on a DOE system due to spawn restrictions. Howard highlighted NERSC's use of spawn within Jupyter notebooks for parallel Python scripts, suggesting potential applications in data science. The conversation also touched on the hierarchical features of Flux, which could enable more complex job structures than traditional MPI or Slurm, though these features are not currently widely used with MPI. Presentation Review and Schedule Planning
Howard and Dominik discussed the presentation slides, with Dominik confirming he had sent the same PDF to Wes and hadn't made any changes since. Howard expressed uncertainty about the schedule but mentioned they were targeting December for a revisit with front of the form, and they might work out their roadmap before then. They briefly touched on an issue opened in the upstream PI Forum and the associated pull request, which was related to previous discussions. MPI Object Attribute Simplification Discussion
Howard discussed modifying object attribute listings and proposed a generalized approach inspired by MPI's handling of void pointers, suggesting it could simplify the API. Tony warned against adding unnecessary complexity to MPI, emphasizing the need for clear use cases and practical benefits, and suggested extracting the data types chapter into a separate standard. The group agreed to schedule a sessions working group breakout, with Howard planning to coordinate with Wes and Tony for a suitable time.