BGP Next-next Hop Nodes#

Kevin Wang, Jeff Haas, Changwang Lin, Jeff Tantsura (Juniper)

• Why: AI/ML Clos fabrics carry high-throughput, low-entropy flows that random ECMP hashing cannot keep drop-free; local remedies (dynamic load balancing) can't see remote-link congestion. This draft is the BGP standardization of the topology discovery that Global Load Balancing (GLB) needs.
• What: The NNHN characteristic for the NHC attribute, letting a leaf learn <next-hop, next-next-hop> tuples per route. Presented with GLB test results: on a 3-stage Clos with 100G links, a 90G flow that drops without GLB runs drop-free with it.
• How: The key insight is that Clos symmetry means only next-hop and next-next-hop link congestion matters — congestion beyond that is absorbed by symmetric GLB at higher tiers — so a non-link-state protocol like BGP suffices for GLB topology discovery. BGP carries next-next-hop BGP IDs in a new NHC TLV; each spine broadcasts link quality to direct neighbors (sub-millisecond, via the companion UDP router-info protocol draft-zzhang-rtgwg-router-info's Neighbor Path Information TLV), and each leaf combines the two-hop link qualities to steer load. Works for 3-stage, 5-stage, and general Clos.
• Where applicable: IP/IPv6-routed BGP Clos fabrics — i.e., modern AI back-end networks. WG draft (draft-ietf-idr-next-next-hop-nodes-00) at first presentation: early on the WG track, but with working multi-party test evidence. The single most directly relevant AI-fabric item in the program.