Session Date/Time: 25 Jul 2025 07:30

rift

Summary

The IETF rift working group meeting covered several topics including the publication of RFC 9692, updates on the KV store and segment routing, discussion around multicast in AI/ML environments, and a proposal for dragonfly routing. The discussion highlighted the increasing relevance of rift in data center networking.

Key Discussion Points

RFC 9692 Publication: RFC 9692, the rift Protocol Spot, has been published.
KV Store: The KV store document has completed Shepard Review and will be put to last call after this IETF meeting. IPR review is a to-do item for the chairs.
Segment Routing: Continued work on segment routing with good progress on services.
rift Multicast: Revival of multicast work is anticipated due to increasing interest in multicast for AI data centers. BEER integration with rift was also suggested.
Multicast for MOE: Discussion focused on using multicast for Mixture of Experts (MOE) architectures in large language models (LLMs). The motivation is to reduce pressure on source GPUs and reduce congestion between nodes.
MOE Multicast Challenges: The discussion acknowledged the challenges including inconsistent sending methods (memory reading, DMA), coordination between large model coding and network layer, packet encapsulation and reliability issues.
Dragonfly Routing: A proposal for dragonfly routing was presented, aiming for dynamic routing without silicon changes or complex VPN policies. The proposal involves a split horizon approach based on topological ordering.
Reliability of Multicast: Reliability was identified as a major obstacle to deploying multicast in AI networks, necessitating consideration of mechanisms for handling packet loss and retransmission.

Decisions and Action Items

KV Store: Chairs to initiate IPR review for the KV store document. Authors to address document edits after IETF 123.
MOE Multicast: Sandy to share the recording link of the MOE side meeting to the mailing list.
Dragonfly Routing: Tony to update the dragonfly routing draft.

Next Steps

Continue working on segment routing and auto-assess work.
Further investigate the use of BEER in rift for multicast scenarios, specifically considering reliability aspects.
Explore the potential of dragonfly routing within the rift framework, taking into account power constraints and reconfigurable optical switches.