[ofiwg] OFIWG 8-19-2025 Minutes

[Xiong, Jianxin]

08/19/2025

* Participants

Alex McKinley (Intel)
Alexia Ingerson (Intel)
Ben Lynam [Cornelis]
Howard Pritchard
Jerome Soumagne (HPE)
Jianxin Xiong (Intel)
John Byrne [HPE]
Ken Raffenetti (ANL)
Shi Jin (AWS)
Stephen Oost (Intel)
Zach Dworkin (Intel)
Rajalaxmi (Intel)

* Summary

Libfabric 2.3.0 is scheduled for September release, with RC1 on 9/1, RC2 on 9/8, and GA on 9/15.

Update on the system memory monitor issue: PR#11282 tries to change the ordering of default
memory monitors by making uffd_monitor with the highest priority. A few past issues with the 
uffd monitor were examined: #5554 was fixed by #5562; #5580 was closed as user error; #5662
was closed for inactivity, with #5666 as workaround. Shi added another past PR #6268 that tried
to work around similar issue in efa. Before the issues are resolved, it is recommended not to 
make  uffd monitor the default. Going to make the PR on hold and update the related info in the 
comments.

DL providers have their own copies of the global states. Some of the initialization happen inside
fabric.c which is not part of the DL provider. DL providers need to initialize them separately. Today
some initialization is not present in some DL providers and that leads to a few reported issues. 
For example #11312 for global parameters and a report in the slack channel for default monitor
setting. PR#11320 addresses on of the issues and there will be more PRs to address the rest.     

-----------------
Jianxin Xiong
Fabric Software
Intel Corporation
Jianxin.xiong at intel.com