[ofiwg] OFIWG Meeting 2/24/2026 Minutes

[Xiong, Jianxin]

2/24/2026

** Participants **

  Alex McKinley (Intel) 
  Alexia Ingerson (Intel)
  Bob Cernohous (Cornelis Networks)
  Howard Pritchard
  Jerome Soumagne (HPE)
  Jianxin Xiong (Intel)
  John Byrne [HPE]
  Ken Raffenetti (ANL)
  Sai Sunku [AWS]
  Shi Jin (AWS)
  Stephen Oost (Intel)
  Zach Dworkin (Intel)
  rajalaxmi


** Notes **

* Release v2.5.0 update
    - RC1 is expected in less than a week
    - Please mark PRs that need to get into RC1 with the "for-2.5.x" label
    - New shm provider (PR#11877) is now cleared for v2.5.0.  The IMB-EXT accumulate test runs very slow with Open MPI. 
      This test is to be disabled in AWS CI and a new Open MPI issue is to be created to track this. 
    - opx provider update is being worked on, may not be ready this week

* Discussion on IPC support for allocation made by cudaMallocAsync()
    - cudaMallocAsync() and some other allocation routines use new CUDA driver VMM API. The resulting buffer does not
       work with cudaMemGetIpcHandle() (or the driver API cuIpcGetMemHandle()). 
    - New API calls are needed to export / import sharable handle for IPC
    - There is no way to identify the method of allocation for a given device pointer
    - Could add the new calls as a fallback path: extra overhead
    - Can the new API replace the old API completely? Need to verify
    - The new API may export shared handle as file descriptor. Need extra steps to pass to other processes:
       pidfd, or control message in sockets

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