Source Routing for AI Fabrics

Author: Kishore Gummadidala & Prashant Baker, Directors of Software Product Management, Marvell Collection: 2023 Global Summit - Networking

Summary

Marvell proposes Flow Path Routing (FPR), a source-routing approach for managing traffic flows in AI training fabrics as an alternative to existing congestion-control methods like packet spraying, virtual output queues (VoQs), and telemetry-assisted adaptive congestion control. The core insight is that AI job mappers already know the full topology and workload placement, so traffic paths can be pre-programmed into the fabric by an SDN controller before workloads begin, eliminating per-hop lookup latency and enabling holistic congestion management from a central vantage point.

The proposal introduces a new FPR shim header inserted between Layer 2 and IP headers, containing five 16-bit hop ID fields that map directly to physical switch port numbers. Access ports perform a single flow lookup (using source IP, destination IP, source port, QP source, QP destination) and encapsulate the FPR header, while fabric ports simply read the next hop ID from the packet and forward without any lookup. The controller requires intimate knowledge of the physical port mappings on each switch. The HLD has been submitted to the SAI community on GitHub for review and collaboration.

The presenters explicitly position FPR as an early-stage alternative, not a replacement for existing methods, and acknowledge open questions around failure handling at hyperscale (100K+ clusters), noting that audience feedback from Meta highlighted that static path assumptions break down when failures occur every two hours at scale. No performance data was available at the time of presentation.