-
Notifications
You must be signed in to change notification settings - Fork 1
2025 09 08 webex
AI can make mistakes. Review for accuracy.
The team reviewed project progress and discussed updates to the homepage, including links to previous sessions and recording additions. Dominik presented on resource allocation changes and proposed extensions to the MPI Sessions model, focusing on dynamic process management and interactions between applications and resource managers. The team prepared for upcoming presentations, including slides on MPI operations and set operations, while also discussing implementation details and use cases, with plans to present Sessions Attributes at a future forum.
Next steps:
- Dominik to update the main session's homepage with links to the most recent recordings.
- Dominik to add an explanation about why MPI groups are not sufficient for dynamic process management in the presentation slides.
- Dominik to remove or rename the "persistent" parameter in the slides and indicate it's still under consideration.
- Dominik to reorder slides to show the overview of arguments before going into details.
- Dominik to add his email address to the final slide for follow-up questions from the community.
- Dominik to present the slides at the MPI meeting on Wednesday.
- Martin to prepare for potential questions about the complexity of the proposed approach.
- Howard to open an issue and move the pull request for Sessions Attributes to the MPI standard repository.
Dominik presented his slides on the project's goals and progress, noting that the content hadn't changed much since the last meeting. He mentioned that he had added a recording to the session's homepage but no transcript was available since the AI tool only works when he's in the meeting. The team discussed updating the homepage with links to previous sessions, including one from two weeks ago that featured a discussion with Sonya, Tony Skellum, and others.
Dominik presented a discussion on changing resource allocation and processes for applications, focusing on both system and application perspectives. He outlined use case scenarios, including swapping resources, dynamic power capping, and adapting to varying application requirements. Dominik also highlighted limitations in the current MPI specifications and proposed extensions to the MPI Sessions model to enable dynamic process management. The presentation concluded with an overview of interactions needed between applications and resource managers, as well as a graphical representation of a set operation in the MPI Sessions IP setup function.
The team discussed their upcoming MPI meeting presentation, focusing on the sequence of a static operation throughout the stack and the need to address specific questions from the community. They agreed to add a slide with key questions, including optimization object design and PSAP interface complexity. Dominik confirmed he would go into more detail on all slides during the presentation. The team decided to postpone presenting their complex topic until the December forum to allow more time for integration of feedback.
The team discussed slides for a presentation, with Sonja and Martin providing feedback to Dominik. They agreed to add an explanation about why MPI group operations are not suitable for the current modification of process constellations, as this point might be frequently asked. Dominik acknowledged this was an important addition and committed to including it in the slides.
The team discussed the need for p-set operations in MPI, which are unique in considering process-level operations not visible to resource managers and runtimes. Sonja noted that in MPitch, a p-set is an MPI group with a name, and Martin highlighted that this approach allows for use beyond MPI, though they agreed not to emphasize this aspect too much in the MPI forum. The group also touched on issues with MPitch and Slurm's PMIX support, noting that some systems experience long timeouts, but concluded that this might be a Slurm-specific problem rather than an MPitch issue.
Dominik and Martin discussed function parameters and mathematical set operations, deciding to focus on set operations for explaining different possibilities. They agreed to remove the persistent argument for now to avoid complexity. iPhone-DX3GXN6E0DXP raised questions about output parameters, and Dominik explained the need for an in-out parameter to handle set operation cancellations and race conditions when the application terminates. Sonja suggested providing an overview slide before going into detailed explanations.
The team discussed how to handle a persistent parameter in a proposal, deciding to leave it unstruck with a note that they are still considering it but won't discuss it on Wednesday. They agreed to rename the parameter to avoid misunderstandings. The group also reviewed implementation-specific slides about set operations and resource management, noting these could be presented as part of a complete flow from application to resource manager. Martin suggested starting with an initial version in the MPI standard before moving it to supplementary documents as it grows.
The team discussed the complexity and coverage of their use cases, emphasizing the need to ensure no use cases are missed. Martin suggested collecting feedback via email to address any potential gaps. Dominik mentioned a potential student project on fault tolerance with MPI Sessions and ULFM, though the timeline remains uncertain. The group also touched on the Phoenix framework, noting its limited user base but potential utility for fault tolerance. The conversation ended with plans to present Sessions Attributes at an upcoming forum and to review slides for further feedback.