[ofiwg] OFIWG 5/19/2026 Meeting Minutes

[Xiong, Jianxin]

* Participants

  Andriy Kot (Cornelis Networks)
  Bob Cernohous (Cornelis Networks)
  Jerome Soumagne (HPE)
  Jianxin Xiong (Working Group Chair)(Intel)
  John Byrne [HPE]
  Ken Raffenetti
  Lindsay Reiser (Cornelis Networks)
  Naveen (HPE)
  Rajalaxmi (Intel)
  Sai Sunku [AWS]
  Shi Jin (AWS)
  Stephen Oost (Intel)

* Summary

* Libfabric release v2.6.0 schedule: 
	
  RC1 on 6/1, RC 2on 6/8, GA on 6/15. Please mark PRs targeting v2.6.0 with the "for-2.6.x" label. It is expected that only bug fixes get in after RC1.
	
* Accelerator API discussion
	
  A new parameter "scope" is added to the accelerator side API calls. The scope can be "work item (thread)", "subgroup(warp)" or "work group (thread block)". 
All the threads in the same scope are required to make same call. This allows the implementation to optimize the processing of work requests (concurrent write
to WQ and batch submission). 
	
  Naveen (HPE) raised the question that even without the new parameter the provider can still do batch processing. The argument is that the parameter acts 
as a hint with which the provider can do the optimization more proactively.
	
  Naveen (HPE) also asked about how the provider sets up the queues to be device accessible. This is done in the host side setup stage.  More information is 
available in previous part of the presentation. There was also a question on whether it is possible to expose the internal data structure and let user to handle 
it (copy, etc). More discussions are welcome.
	
  Shi Jin (AWS) mentioned user's interest on separate request preparation and submission phase, similar to RDMA core's WR interface. This is not limited to 
accelerator API.  


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