SRv6 for AI Backend

Author: Abhishek, Microsoft; Changong, Microsoft Collection: 2025-OCP-Global-Summit-Networking

Summary

Microsoft is deploying Segment Routing over IPv6 (SRv6) on its AI backend networks using SONiC as the network operating system, addressing two core challenges: the low-entropy nature of AI training traffic (small numbers of large elephant flows that don't distribute well across equal-cost paths) and the need to connect geographically distributed GPU clusters across multiple physical data centers. SRv6 provides source routing with explicit path specification, enabling fine-grained traffic engineering and fast failover without complex control-plane protocols.

The specific SRv6 implementation uses compressed micro-SIDs (uSIDs) in the IPv6 destination address field rather than an explicit Segment Routing Header (SRH), which simplifies hardware implementation. A 32-bit locator block plus 16-bit per-node SIDs allows a full three-tier path (T0 → T1 → T0 → T1) to be encoded in a single IPv6 DA field for a traditional three-tier network. The control plane is intentionally simple: static SID allocation (one SID per switch), static routes for each neighbor, and a private centralized SDN controller that distributes the list of valid paths to NICs. Path failover is handled at the NIC level using packet trimming signals for congestion/failure detection, and the NIC simply replaces the affected SID in the IPv6 DA with a new one. ASIC stress testing was done using a "snake topology" innovation that fully saturates switch ASIC bandwidth using only two traffic generators, avoiding expensive external traffic generator costs.