Networking Minerva User Guide

Traffic and Congestion

The Traffic and Congestion Dashboard is a focused view within the Professor-X Dashboard. It covers high-level bandwidth trends and the immediate health of the RoCE fabric. Use this dashboard during active training jobs to determine if the network is causing performance issues.

Fabric Throughput

  • Max BW Per Level: Displays 5-minute averaged bandwidth trends across four roles: leaf to NIC, leaf to spine, spine to leaf, and super spine to spine.

  • SU Activity: Shows the highest bandwidth from NIC to leaf in the last 6 hours, grouped by pod and Scalable Unit (SU). Use it to identify underperforming SUs.

  • Link Status Heatmaps: Displays any operationally down interface via the reported operState for the following connections:

    • Leaf to NIC

    • Leaf to spine

    • Spine to super spine

Congestion Control (RoCE)

  • Adaptive Retransmissions: Visualizes hwRoceAdpRetrans counts. Use it to identify packet loss in the fabric.

  • CNP Handled: Tracks Congestion Notification Packets at the NIC level.

  • Pause Frames: Monitors TX and RX pauses per role (leaf, spine, and super spine).

Discards

  • Discards - NIC: Tracks ifInDiscards at the host level.

Last updated: