SRv6 For AI Networks

Overview

SRv6 (Segment Routing over IPv6) applied to AI backend and DCI networks provides explicit source-routed paths for low-entropy elephant flows that are poorly served by ECMP-based routing. By encoding the full forwarding path in the IPv6 destination address using compact uSID encoding, SRv6 enables deterministic traffic engineering without distributed signaling protocols. Multiple hyperscalers — Microsoft and Alibaba — are deploying SRv6 on SONiC Ecosystem-based AI backend networks, replacing traditional MPLS/RSVP-TE stacks with a simpler BGP+SRv6-only architecture.

SRv6 is particularly relevant to AI workloads because GPU-to-GPU collective communication generates large, long-lived elephant flows with strict latency requirements that ECMP hash collisions degrade unpredictably. Source routing allows NICs and host stacks to make path selection decisions at flow origination, avoiding the per-packet load balancing limitations of conventional IP routing. Both Microsoft and Alibaba have contributed SRv6 features back to the SONiC Ecosystem open source, extending it with SRv6 VPN, SRv6 policy, and C-marking capabilities. The SAI-Switch-Abstraction-Interface has been extended with SRv6 endpoint flavors (uN, uDT, uA) to support these deployments.