Fully Adaptive Routing Ethernet in Multi-Plane Scale-Out Networks#

Yinben Xia, Weifeng Zhang (Tencent), Peilong Wang (Baidu), Haibo Wang (Huawei), and a large RNIC/switch author group (Broadcom, Intel, Ruijie, Yandex, China Mobile, GPU vendors).

• Why: ~100K-GPU clusters are moving to multi-plane scale-out topologies (RNIC multi-homed across disjoint planes, no inter-plane links) — which pushes path selection into the NIC. Route scale is the blocker: 4 planes × 100K host routes ≈ 400K routes, impractical on an RNIC.
• What: FARE-in-MPSON extends FARE-BGP from switches to RoCE NICs for multi-plane scale-out (complementing FARE-in-SUN for scale-up). It contributes two route-table suppression techniques: route aggregation with unreachability notification (zero-bandwidth community, or unreachable-prefix advertisement) and selective route filtering via Prefix ORF (RFC 5292), plus a transitive FIB-Suppress extended community so switches can omit RNIC host routes from FIB.
• How: BGP sessions terminate on RNICs; the RNIC evaluates per-plane reachability and path bandwidth for weighted ECMP. Because collectives mean each GPU talks to a limited peer set, RNICs can request only needed host routes; on unreachability they install a host route from the aggregate's remaining next-hops.
• Where applicable: Hyperscale multi-plane AI scale-out with RNIC-resident BGP — relies on predictable collective-communication peer sets. Version -00, very early, but the author list signals real industry intent.