-
Notifications
You must be signed in to change notification settings - Fork 1
2026 02 23 webex
AI can make mistakes. Review for accuracy.
The team discussed preparations for an upcoming MPI forum where Howard will not be attending, including content planning and scheduling considerations for virtual and in-person attendance. The group explored various technical topics related to MPI 6 development, including dynamic process mechanisms and functionality proposals, while also addressing concerns about resource management and interface compatibility. The conversation ended with discussions about meeting frequency and scheduling logistics, including plans for future meetings and forum organization.
- Dominik: Provide summary from the working group to Martin Schulz at the forum
- Dominik: Update the main wiki with links to meeting recordings
- Howard: Clean up the slides/notes and send to the mailing list
- Sonja: Remind Wes about scheduling the plenary presentation at a time suitable for European participants (morning New York time)
- Sonja: Lead discussion on feedback from the spawn presentation at the next working group meeting on March 23rd
- Sonja and Dominik: Decide during the forum whether tools people need working group breakout session time
The meeting focused on preparing for an upcoming forum where Howard will not be attending. Dominik agreed to provide a summary to Martin Schulz regarding the working group's plans for MPI 6. Sonja confirmed she would attend virtually from Germany, and Dominik also planned to attend in person. The group discussed whether to have a working group breakout session, but decided to make this decision on the fly at the forum next week.
The team discussed preparing content for an upcoming MPI 6 forum, focusing on key areas including dynamic process mechanisms, spawn functionality, and a publish lookup dictionary feature. Dominik suggested making the lookup dictionary a separate, more general functionality rather than integrating it with dynamic processes. Sonja raised a concern about ensuring the proposed interfaces don't conflict with ongoing ULFM work. The group also briefly considered including smaller items in an MPI 5.1 bug fix release, though the final decision was left pending further discussion.
Howard expressed concerns about the sustainability of MPI meetings given the current industry direction and vendor focus on alternative technologies like NVIDIA's workflow gluing solutions. Michael explained that HPE cannot attend the forum due to lack of travel budget, though they are interested in GPU-MPI interfaces and offloads. Both participants acknowledged limited resources for contributing to MPI standard updates at this time.
Sonja discussed the tech team's resources for the working group and suggested limiting in-person meetings to once per year due to travel and budget constraints. She expressed concern that reducing meeting frequency could slow down the pace of MPI updates, which she argued cannot afford to be slower given the need to keep up with AI and other emerging technologies. Howard agreed with the idea of reducing in-person meetings and suggested holding them in North America, potentially in Canada, though he noted uncertainty around European immigration to the US due to new travel criteria.
Howard and Sonja discussed the challenges of attracting travelers to the United States, particularly regarding budget constraints and limited time for activities. Howard emphasized the importance of improving MPI's usability in accelerator contexts as a key driver for future development, noting ongoing funding opportunities through Tony Skellum's group and Patrick Bridges. He also highlighted the potential for better interoperability with complex mixed AI and HPC workflows.
The team discussed an open question about how much dynamic resources management content should be included in the main document versus the site document of the standard. Howard acknowledged this topic needed follow-up and mentioned cleaning up related information about process set name conventions and COL (cooperative optimization language) for the scheduler. The discussion concluded with plans to address this topic in more detail in the following week.
The group discussed scheduling their next meeting, with Sonja confirming it would be in two weeks on March 9th. Howard mentioned he would be away in South America and wouldn't return until March 23rd. The team decided to proceed with the March 9th meeting to discuss feedback from a recent plenary session, though Howard would not be able to attend.
The team discussed the organization of the MPI forum, where Sonja will present the spawn proposal in the plenary session. They agreed to potentially schedule the Sessions working group update before the spawn details, with Sonja requesting a late afternoon slot to accommodate time zones. Howard reminded everyone to remind Wes about the scheduling considerations, and the conversation ended with well-wishes for Howard's upcoming vacation.