Enhanced ECMP for AI Cluster#

Weiqiang Cheng (China Mobile), Changwang Lin (New H3C Technologies).

• Why: In AI training the flow count is small, so per-flow ECMP hash collisions are proportionally more damaging; per-packet ECMP balances perfectly but reorders — unacceptable for RDMA without reordering support.
• What: Two complementary methods: (1) ingress-interface-group ECMP — group ingress interfaces and hash by in-group index so traffic from different ports lands on different ECMP egress ports; (2) egress-group ECMP — the egress leaf advertises a Group ID in a new BGP extended community (Router-ID + Group-ID, 6 bytes), and the ingress leaf maps flows to matching ECMP subgroups.
• How: Purely control-plane/local hashing changes; no datapath encapsulation. Method 1 suits balanced GPU ingress traffic, method 2 suits balanced egress traffic.
• Where applicable: Leaf-spine AI Clos. Individual draft -01 soliciting RTGWG feedback — with an explicit open question whether the BGP extended community should move to IDR. Overlaps in problem space with FARE and the flow-label draft but is a simpler, narrower fix.