ProDeploy — Learn the fabric that keeps AI clusters alive

← Labs ← Chapter 6: Efficient Load BalancingDiagnose uneven spine utilisation

Desktop recommendedChecking session

Diagnose uneven spine utilisation

intermediate

Elapsed

00:00

Scenario

AllReduce throughput has dropped 25% across the 16-node training cluster. JCT has increased by 35%. No packet drops are visible. PFC pause frames are present but modest. No link failures. No alerts have fired. The cluster uses 4-way ECMP from the leaf to Spine 1-4. The current workload has large RoCE flows with the same UDP dst-port, so static hashing may be pinning traffic to a subset of spines. Your task: 1. Check leaf and spine interface counters for traffic skew 2. Confirm adaptive routing is currently disabled 3. Enable adaptive routing on all four leaf uplinks 4. Verify spine utilisation has equalised

Conditions

Spine utilisation checked
Load balancing mode identified
Per-packet load balancing enabled
Spine utilisation verified (balanced)

Need Help?

Open the step-by-step command guide for this lab.

Topology

Filled green circle = Active

Unfilled red circle = Down

Solid line = 400G link

Dashed line = degraded

Click a device node to open its terminal session

Preparing lab environment...