Uploaded February 2026 | Updated September 2026, 1 week ago
Abstract
RoCEv2 (RDMA over Converged Ethernet) has created a significant visibility gap. Because RDMA utilizes kernel bypass technology to achieve ultra-low latency, traditional host-based monitoring tools like tcpdump are blind to the traffic. A live dashboard showing San Diego Supercomputer Center’s Expanse cluster production traffic will demonstrate how sFlow instrumentation in modern data center switch hardware provides the real-time visibility into RDMA operations needed to fill the visibility gap.
Key Takeaways:
• Why kernel bypass makes traditional host-based agents ineffective for AI clusters
• How to decode packet headers from sFlow telemetry to monitor RDMA I/O operations
• Using sFlow dropped packet notifications to instantly diagnose the location and cause of packet loss
Peter Phaal:
Original inventor and a co-author of industry standard sFlow network monitoring technology. Head of the sFlow.org industry consortium responsible for developing, maintaining, and promoting sFlow. Founder and President of InMon Corp., a leading provider of performance analysis software based on sFlow.
nanog.org/events/nanog-96/content/5718
Abstract
RoCEv2 (RDMA over Converged Ethernet) has created a significant visibility gap. Because RDMA utilizes kernel bypass technology to achieve ultra-low latency, traditional host-based monitoring tools like tcpdump are blind to the traffic. A live dashboard showing San Diego Supercomputer Center’s Expanse cluster production traffic will demonstrate how sFlow instrumentation in modern data center switch hardware provides the real-time visibility into RDMA operations needed to fill the visibility gap.
Key Takeaways:
• Why kernel bypass makes traditional host-based agents ineffective for AI clusters
• How to decode packet headers from sFlow telemetry to monitor RDMA I/O operations
• Using sFlow dropped packet notifications to instantly diagnose the location and cause of packet loss
Peter Phaal:
Original inventor and a co-author of industry standard sFlow network monitoring technology. Head of the sFlow.org industry consortium responsible for developing, maintaining, and promoting sFlow. Founder and President of InMon Corp., a leading provider of performance analysis software based on sFlow.
nanog.org/events/nanog-96/content/5718
