Skip to content

IP/Ethernet Storage Networking for AI Infrastructure - BRKDCN-2945#

Cisco Live Europe 2025 (Amsterdam) digest Program Core pick Amsterdam · 9–14 February 2025

Abstract

Speakers: Paresh Gupta (Cisco)

  • Why: Lossless Ethernet for RDMA is squarely in the PM's core, and this is one of the few sessions that grounds PFC/ECN in a concrete workload: storage I/O for AI training and inference, where a single loss can stall an entire job.
  • What: A design-and-operations guide for storage networking under AI workloads: lossless Ethernet foundations (Priority-based Flow Control) for RoCEv2, the interaction of PFC and ECN for RoCE congestion management (RCM), QoS effects on storage traffic, and fabric designs using 100G/400G/800G Nexus switches for RDMA, iSCSI, and NVMe-over-TCP, with NVMe-oF best practices. (from abstract only)
  • How: Explains how to engineer a lossless Ethernet class of service for RoCEv2 (PFC headroom tuning, ECN thresholds, RCM), and how to decide between dedicated and shared storage networks depending on QoS contention, plus operational tooling via Nexus Dashboard.
  • Where applicable: Cisco Nexus switch families and UCS server connectivity; the PFC/ECN/RCM mechanics are general to any RoCEv2-over-Ethernet deployment. Org-specific parts are the Cisco product recommendations; the lossless-fabric engineering is broadly applicable.