Fully Adaptive Routing Ethernet using BGP#

Qingliang Zhang (H3C), Hang Wu (Ruijie), Yadong Liu, Yinben Xia (Tencent), Peilong Wang (Baidu), Xiaohu Xu (China Mobile), Shraddha Hegde (HPE), Keyur Patel (Arrcus), Zongying He (Broadcom), et al.

• Why: Static ECMP in Clos fabrics breaks down for elephant-flow AI training traffic — collisions on equal-cost paths cost throughput exactly when a multi-thousand-GPU job is running. This is the most implementation-proven adaptive-routing item in the routing area.
• What: FARE extends BGP so that each node advertises real-time path bandwidth in a path-bandwidth extended community; receivers compute weighted ECMP (WECMP) next-hops and rebalance per the live capacity of each spine path. New in -04: multihoming handling — when a prefix is multihomed to N leaf nodes, the advertised path bandwidth is divided by N (or the local link bandwidth multiplied by N) before taking the minimum, so weights stay honest.
• How: Leaf takes min(local link bandwidth, advertised path bandwidth) as the route weight for weighted load-balancing; congestion or link-state changes propagate via ordinary BGP updates carrying the bandwidth community. The slides state the draft is ready for WG adoption call.
• Where applicable: Broadly applicable to scale-out IP fabrics — implemented by at least 2 cloud providers (China Mobile Cloud, Baidu Cloud) and 5+ vendors, adapted to four mainstream merchant switch chips including Broadcom Tomahawk 5, Nvidia Spectrum-4, HiSilicon SD58XX, and Centec Arctic. Ethernet Clos fabrics only; not a WAN mechanism.