Fully Adaptive Routing Ethernet using BGP#

Qingliang Zhang (H3C), with co-authors from Tencent, Baidu, China Mobile, Yandex, HPE, Arrcus, Broadcom, Ruijie, Centec, Huawei and IEIT SYSTEMS.

• Why: AI training on 3–5 stage Clos fabrics needs non-blocking behavior; static per-flow ECMP collapses when a few large flows hash-collide onto one path. This is the most implementation-backed routing draft at the meeting.
• What: FARE-BGP — a BGP extension (path-bandwidth extended community) that lets switches do weighted ECMP based on real residual path capacity rather than uniform hashing. Version -05 adds handling for multiple BGP sessions per switch pair (bandwidth computed at node granularity, not session granularity) and packet forwarding over tunnels (WECMP applies to tunneled packets with loopback endpoints). The authors state the draft is ready for a WG adoption call.
• How: The prefix-originating node (e.g., leaf) attaches a maximum path-bandwidth value; each node computes path bandwidth = min(aggregate parallel-link bandwidth, advertised value) and either spreads the weight across ECMP paths or prunes the ECMP set so remaining links match the computed bandwidth.
• Where applicable: BGP-only AI/DC fabrics. Implemented by at least two cloud providers (China Mobile Cloud, Baidu Cloud) and six vendors, and adapted to 6+ switch ASIC generations — Broadcom Tomahawk 5 (51.2T) and Tomahawk 6 (102.4T), Nvidia Spectrum-4, Huawei SD58XX, Centec Arctic, Clounix CLX86. Explicitly multi-vendor, not org-specific.