-
Notifications
You must be signed in to change notification settings - Fork 1
2025 06 30 webex
AI-generated content may be inaccurate or misleading. Always check for accuracy.
The team discussed proposals for reintroducing session attributes and updating presentations for an upcoming Forum meeting, with Howard and Sonja planning to review materials before the late September meeting. They explored various technical implementations, including process set operations, spawn mechanisms, and resource management systems, with Dominik and Martin debating different approaches and Martin presenting concepts of persistent requests in resource management. The group concluded by discussing the implementation of persistent requests in MPI, agreeing to rename the feature and consider different versions, while Howard planned to coordinate future discussions with other team members.
- Dominik to update the example with the missing part in the Mpi_Info_get call.
- Howard to ping Martin Schultz and Wes about restarting informal discussions.
- Dominik to prepare a few slides summarizing the high-level concepts of P set operations and process spawning for presentation at the next informal Forum meeting.
- Howard to update the pull request with an example demonstrating the use of session attributes.
- Sonja to review Howard's updated pull request when pinged.
- Howard to present the session attributes proposal at the next MPI Forum meeting.
- Working group to reconsider the naming and functionality of the "persistent" flag in the P set operation.
- Dominik to check if the AI Bot worked for this meeting and add the summary to the notes if available.
Howard discussed a proposal to reintroduce session attributes to address data type persistence issues across sessions. He explained that while sessions were initially designed to be minimal, some members of the Forum had expressed concerns about removing attributes entirely from windows, communicators, and data types. Sonja suggested adding an example using data types and sessions to clarify the motivation for attributes, as she feared the Forum might question the rationale otherwise.
Howard and Sonja discussed updating a presentation for the Forum meeting, with Howard planning to add new content and seek feedback. They agreed to review the material before the late September Forum meeting, with Sonja offering to provide additional comments if needed. Dominik shared updates on his work, including splitting example 2 into separate cases for growing and shrinking, and mentioned he had added some examples with spawn. Sonja confirmed she had already sent information about the partex spawn extension to Howard for preparation of the last Forum meeting.
Dominik and Howard discussed the Crow example without and with spawn operations. Dominik explained the synchronization process across all processes and the need to pass output and set the name of the new main pset. They also discussed the creation of a new communicator. Sonja raised a question about the non-prox parameter in the spawn method, which Dominik noted might be implementation-specific. Sonja clarified that the MPI library would need to map the pset name to the number of processes, suggesting that the resource manager should inform the application about creating a new pset when sufficient resources are available.
The team discussed separating the process set operation from the spawn operation in their resource management system. Martin and Dominik debated whether this additional step was necessary, given that the resource manager could already start processes after allocating resources. Sonja clarified that the spawn function is primarily for starting processes on already allocated resources, rather than allocating resources itself. The group agreed that while optimization information for resource allocation is important, it doesn't necessarily need to be part of the spawn process.
The team discussed the implementation of MPI processes and the potential complexity of introducing new concepts into the standard. Howard suggested presenting the two approaches they've been considering to the Forum for feedback, as they want to avoid going too far down the wrong path. He emphasized the importance of getting input from outside the working group sooner rather than later, given the complexity of the proposed changes.
The group discussed the implementation of a spawn mechanism for resource allocation in Slurm, with Martin expressing concerns about the current approach and suggesting a clearer interface. Sonja clarified that the spawn function is already used in standards and could be useful for spawning with sessions, potentially helping to reduce complexity. Howard proposed opening the discussion to a broader audience in the Forum after the holiday season and considered whether resource allocation and process spawning should be combined into one operation or handled separately.
The team discussed two approaches for process set operations: a monolithic solution and a more complex, flexible solution. Howard expressed concerns about the monolithic approach being difficult to get accepted, suggesting it might be better to present both versions at a non-voting meeting to gather feedback. Martin argued for the more complex solution, stating that it could be hidden behind an API and would be more flexible for different applications. The team agreed to present both versions at the upcoming meeting for feedback.
The team discussed code implementation issues, particularly regarding missing parentheses in the info get function and a missing flag check. Martin explained the concept of persistent requests in resource management, describing how applications can maintain optimization information across resource changes without needing repeated Pset calls, allowing the resource manager to dynamically allocate and reallocate resources while the application continues running. The team raised questions about how applications would be notified of resource changes, with Martin clarifying that the resource manager would communicate these changes to the application through the Pset interface.
The team discussed the implementation of persistent requests in MPI, realizing that the term "persistent" conflicts with existing MPI terminology. They agreed to rename the feature and consider two versions: a one-shot version for single set operations and a persistent version for ongoing operations. Howard suggested rethinking the concept of requests in the context of the whole standard and proposed creating a new thing instead of just renaming. The group decided to postpone further discussion on this topic for two weeks, during which Dominic will prepare slides summarizing the current work on P-set operations and process starting. Howard will also coordinate with Martin Schultz and Wes about restarting informal discussions.