Skip to content

Clustermapper: Deterministic path diagnosis for AI backend networks using SRv6#

OCP Global Summit 2026 digest Program Core pick 12–15 October 2026

Abstract Paper Session recording Materials

Ryder Lewis (OpenAI), Guohan Lu (OpenAI)

  • Why: Turns path diagnosis from statistical inference into ground truth — a genuinely clever use of SRv6 source routing, and the operational companion to the MRC spec from the same OpenAI team.
  • What: A path-diagnosis system for MRC/SRv6 AI backend networks: a lightweight agent on every node sends SRv6 source-routed self-probes through T0/T1 switch paths, producing a deterministic link-health map used to localize faults (NIC-T0 links, T0-T1 links, switch forwarding) and turn urgent link-failure incidents into scheduled maintenance (from abstract only; the linked arXiv paper — titled "Resilient AI Supercomputer Networking using MRC and SRv6", 18 pages, 22 figures — was not retrievable beyond its landing page).
  • How: SRv6 deterministic source routing pins each probe to an exact path, eliminating ECMP ambiguity and routing-convergence noise so failures localize to specific links even between training jobs; output feeds path-avoidance policy for MRC.
  • Where applicable: Requires SRv6 (and works best with MRC) in the backend network — OpenAI production, but the technique is general to any SRv6 fabric and notably valuable for small network teams operating huge clusters.