RTGWG

Summary

The Routing Working Group (RTGWG) met at IETF 125 in Shenzhen to discuss the status of active drafts and several new proposals focusing on AI/ML network fabrics, fast network notifications, and multicast requirements for large-scale data centers. Significant time was dedicated to the "Fast Network Notification" (FAN) initiative, which is progressing toward the potential formation of a new working group. Other major themes included hardware-accelerated IP Fast Reroute (FRR) and congestion control collaboration across Data Center Networks (DCN) and Wide Area Networks (WAN).

Key Discussion Points

WG Status and Administration

01-rtgwg-status

Draft Updates:
- draft-ietf-rtgwg-multisegment-sdwan has been submitted for publication.
- draft-ietf-rtgwg-srv6-egress-protection finished its WGLC and will proceed to a joint last call with the LSR WG.
- draft-ietf-rtgwg-bgp-pic has been updated with Yingzhen Qu as a co-author to improve readability and structure.
- draft-ietf-rtgwg-net-notif-ps was recently adopted.
FAN (Fast Network Notification): Discussion continues regarding the charter for a potential new working group (formerly referred to as Fun-Tell).

YANG Models for Quality of Service (QoS)

02-QoS_YANG

Aseem Choudhary presented updates to draft-ietf-rtgwg-qos-model (v15).
Technical Updates: Addressed DSCP range violations by clarifying that gateway routers must respect RFC 8436. Clarified Per-Hop Behavior (PHP) selection at edge and core nodes. Data types for filters/actions were refined (e.g., using leaf-list for IP addresses and leafref for policy references).
Next Steps: The chairs noted that directorate reviews are complete and the document is nearing its Working Group Last Call (WGLC).

Fast Network Notifications (FAN)

03-Fast Network Notifications Problem Statement

Jie Dong provided an update on draft-ietf-rtgwg-net-notif-ps.
Scope: The document focuses on the need for timely network operational status (congestion/failure) to support AI/ML training and cloud services.
Discussion: Jeff Tantsura noted that security mechanisms should not introduce latency, potentially leveraging the "limited domain" concept. A data model for these notifications is in development.

IP Fast Reroute for AI/ML Fabrics

04-IP Fast Reroute for AI/ML Fabrics

Roy Yang presented a framework for sub-millisecond convergence in AI fabrics.
Proposals: Hardware-accelerated protection activation (bypassing the CPU), complete topology visibility in BGP-based networks (exchanging link-state via BGP), and quality-aware remote protection.
Feedback: Francois Clad questioned how 100-microsecond detection is guaranteed without hardware-level probing. Jeff Tantsura advised separating Data Center and WAN scenarios clearly, noting that existing technologies like TI-LFA are available for WAN but DCN requirements differ.

Efficient Remote Protection

05-Efficient Remote Protection

Francois Clad discussed limitations of local-only FRR, such as traffic "hair-pinning" and lack of load awareness.
Solution: A mechanism where a node detecting a failure or quality degradation (e.g., a partial link-bundle failure) notifies remote nodes to reroute traffic earlier.
Discussion: Jeff Tantsura cautioned against network-based "bypass" solutions in AI networks, as hair-pinning can be detrimental to collective performance. He emphasized the importance of ensuring network-level FRR does not conflict with host-level congestion control mechanisms.

BGP-based Adaptive Routing for Scale-Up Networks

06-xu-fare-in-sun

Roy Yang presented draft-xu-rtgwg-fare-in-sun, proposing BGP on the host for scale-up (GPU) networks.
Proposal: Uses BGP to perform adaptive routing (WCMP) based on link capacity.
Questions: Jeff Tantsura asked about non-IP encapsulation in scale-up frameworks and the necessity of BGP in single-tier networks where no routing typically occurs.

Flow Control Collaboration Across DCNs and WAN

07-Use cases and Requirement for Flow Control Collaboration Across DCNs and WAN

Hanjunxin proposed Fine-Grained Flow Control (FGFC) to bridge PFC (DCN) and WAN congestion management.
Mechanism: Edge nodes coordinate protocol conversion and semantic mapping between PFC frames and FGFC packets (carried via ICMP/UDP).
Feedback: Jeff Tantsura requested the addition of a table detailing buffer memory requirements relative to WAN distances (e.g., 1MB per 20km).

Multicast in AI Data Centers

08-Multicast Use Cases for Large Language Model Synchronization 09-Requirements and Gap Analysis of Multicast in AI Data Centers

Yisong Liu and Jian Zhang discussed multicast for LLM model distribution and collective communication (All-Reduce/MOE).
Requirements: Interactivity (MP2P feedback), reliability (lossless), and dynamics (microsecond membership changes).
Gaps: Existing technologies (PIM, SR-P2MP, BIER) were analyzed. Tony Przygienda noted that "U-BEER" (uncompressed BIER) might address sparse group concerns.

Satellite Networking and SRv6

10-SDAF for LEO Satellite Networks

Kemin Liang presented Symmetry-Driven Asynchronous Forwarding (STEP) for LEO satellites, leveraging Torus/Ring topology symmetry to prevent micro-loops. 11-Congestion Control Based on SRv6 Path
Yisong Liu proposed a hop-by-hop congestion notification for SRv6 paths to alleviate the processing load on the head-end node.

Decisions and Action Items

draft-ietf-rtgwg-qos-model: Authors to incorporate final review comments and prepare for WGLC.
draft-ietf-rtgwg-srv6-egress-protection: Chair (Jeff Tantsura) to perform final review and progress to joint last call with LSR.
draft-ietf-rtgwg-bgp-pic: Chairs to shepherd the document following recent updates.

Next Steps

Interim Meeting: The chairs indicated a potential interim meeting before IETF 126 to handle the high volume of AI Data Center-related submissions.
FAN Working Group: Continued refinement of the charter on the mailing list and GitHub.
Data Model: Jeff Tantsura and Yingzhen Qu plan to publish a data model draft related to fast network notifications.

Automatic IETF Minutes