Network Metrics for ML Workloads

Network metrics are essential for understanding distributed ML performance.

Useful signals:

• per‑port throughput and utilization
• packet drops and errors
• retransmissions (for TCP)
• ECN marks, PFC pause frames (for RoCE)
• RTT distributions and tail latency

Symptoms and hints:

• high retransmits → congestion or loss
• high ECN marks → queues and congestion, but not necessarily loss
• frequent PFC pauses → risk of head‑of‑line blocking