Use cases and Requirement for Flow Control Collaboration Across DCNs and WAN#

Zhengxin Han, Ran Pang, Yi Yue, Jie Dong, Zheng Ruan, Quan Xiong (China Unicom)

• Why: Distributed AI training and inference now span data centers over WAN, but PFC — the DCN lossless mechanism — fails badly in WANs: head-of-line blocking, congestion spreading, and deadlock under multipathing, with long RTTs amplifying everything. Extending lossless RDMA end-to-end across DCN+WAN requires the two flow-control regimes to interwork.
• What: Use cases and requirements for co-deploying PFC inside DCNs with fine-grained flow control (fgfc) in the WAN: fgfc is a data-plane L3 backpressure mechanism (carried in ICMPv6 or UDP) operating at tenant/flow granularity, bounded to specific paths and slices, with ms-level response.
• How: Edge gateways (R1/R5 in the examples) do bidirectional translation: DCN PFC pause → fgfc packet (priority, slice ID, pause time) upstream; WAN fgfc → standard PFC frame toward the DC. Requirements on edge nodes: PFC↔fgfc protocol conversion, port-level↔tenant/flow-level semantic mapping, policy coordination, slicing and HQoS support, buffer management. Supports both hop-by-hop and tunneled cross-hop WAN backpressure for legacy intermediate nodes.
• Where applicable: DC-interconnect WANs carrying RDMA/AI traffic. Field trials reported over 200+km (Hangzhou–Jinhua storage/compute separation training) achieving >95% computational efficiency with RDMA lossless over the WAN — operator-specific (China Unicom) evidence but the mechanism is general. Individual requirements draft, seeking review.