Towards Fully-Controllable Packet Steering for AI Backend Networks with SRv6

Distributed AI training and inference demand precise traffic control to achieve optimal network performance, yet current traffic management methods remain passive, coarse-grained, and fragmented. We argue that future AI backend optimization requires holistic, proactive, packet-level controllability.
In this position paper, we propose a new vision of leveraging Segment Routing over IPv6 (SRv6), a mature WAN technology, to achieve comprehensive traffic controllability in AI backend networks.

We discuss the key advantages of SRv6, including enhanced load balancing, improved failure recovery, and efficient network monitoring. Through our preliminary design and experiments, we demonstrate SRv6’s feasibility and potential performance benefits. At the same time, we explicitly outline critical open problems and challenges in adopting SRv6, such as designing effective path assignment algorithms in diverse workloads, scalable control-plane design, high-performance data-plane integration, and effective hardware-software co-design. Furthermore, we identify and present additional open research directions and questions necessary for realizing fully controllable AI backend networks.