Skip to content

Designing RoCEv2 Networks for Distributed AI Training with Cisco Nexus 9000 - BRKDCN-2688#

Cisco Live US 2026 (Las Vegas) digest Program Core pick Las Vegas · 31 May – 4 June 2026

Abstract

Praveen Poojary (Cisco)

  • Why: Lossless/low-loss RDMA fabric tuning is the hardest operational problem in AI clusters today; this is the session that actually walks the PFC/ECN/congestion-control knobs.
  • What: A design-and-operate guide for RoCE fabrics purpose-built for training jobs of thousands to tens of thousands of GPUs running days-to-weeks, targeting near-lossless behavior and predictable AllReduce/AllGather throughput (from abstract only).
  • How: Covers the frontend/backend network separation common at hyperscalers and its implications for leaf-spine topology, routing, and failure domains; then congestion control, PFC/ECN, and QoS tuning for bursty, low-entropy AI flows.
  • Where applicable: Nexus 9000-specific in tooling, but the lossless-fabric design principles (entropy, incast, ECN thresholds) apply to any RoCEv2 backend fabric.