---

name: gke-ai-troubleshooting-tpu-performance-degradation description: >- Diagnose GKE Cloud TPU training throughput drops and step-time regressions (15%+ TPU duty-cycle drop) using ML Diagnostics Workload Monitoring (gcloud alpha mldiagnostics monitored-events / hypercomputecluster.googleapis.com/v1alpha) and 1-minute Cloud Monitoring system metrics (kubernetes.io/node/accelerator/*). Distinguishes hardware and network fabric throttling from workload resource bottlenecks (HBM capacity, host memory, or host CPU saturation). Use when TPU training throughput or duty cycle drops without crashing pods, when PERFORMANCE_DEGRADATION monitored events fire, or when triaging slow multi-slice training steps. Don't use for complete multi-slice XLA execution stalls with HANG_DETECTED logs (use gke-ai-troubleshooting-tpu-mxla-hang) or pod eviction/interruption restarts (use gke-ai-troubleshooting-jobset-interruption). metadata: version: "1.0.1" category: AiAndMachineLearning

Troubleshoot GKE TPU performance degradation with ML Diagnostics Workload Monitoring

Diagnose and mitigate Cloud TPU training throughput drops and step-time regressions (15%+ drop in TPU duty cycle) on Google Kubernetes Engine (GKE) by correlating ML Diagnostics Workload Monitoring MonitoredEvent analyzer reports with 1-minute Cloud Monitoring system metrics and GKE node topology labels.

---

Attribution

MUST READ: Before executing any gcloud, REST API, or MCP command, you MUST read Usage Attribution. For commands executed directly in the terminal (never in generated scripts, Terraform, or user code), prefix gcloud with CLOUDSDK_METRICS_REQUEST_ATTRIBUTION="agent-skills/1.0 (skill:gke-ai-troubleshooting-tpu-performance-degradation)" and append agent-skills/1.0 (skill:gke-ai-troubleshooting-tpu-performance-degradation) to User-Agent headers.

Prerequisites

• Tools: Install the Google Cloud SDK (gcloud with alpha component for gcloud alpha mldiagnostics) and kubectl.
• Cloud Billing & Project Configuration: Verify an active billing account is linked (gcloud billing projects describe {project_id}), authenticate (gcloud auth login), set the target project (gcloud config set project {project_id}), and ensure container.googleapis.com, monitoring.googleapis.com, and hypercomputecluster.googleapis.com are enabled.
• Supported workloads and GKE versions: Google Cloud ML Diagnostics only supports JAX on TPUs (see ML Diagnostics platform). Workload Monitoring is enabled by default, supports the jobset and job GKE job types, and is compatible with GKE versions 1.36.0-gke.4681000 and later, as stated in Configure GKE for ML Diagnostics. If a workload uses another framework (such as PyTorch) or another custom resource type, gcloud alpha mldiagnostics won't list ML runs or monitored events for it. On-demand profiling (Step 4 Path A) additionally requires the cluster setup described in that document.
• Required IAM Roles:
  • Cluster Director Editor (roles/hypercomputecluster.editor), the role listed in the "IAM permissions" section of ML Diagnostics platform, for the ML Diagnostics CLI and API calls in this skill (ML runs, monitored events, and on-demand profiler sessions)
  • Monitoring Viewer (roles/monitoring.viewer) for the PromQL queries in Step 2
  • Kubernetes Engine Viewer (roles/container.viewer) for the kubectl get nodes query in Step 3
  • For remediation ([High Risk] steps): Kubernetes Engine Cluster Admin (roles/container.clusterAdmin)
• Reference Documentation:
  • ML Diagnostics platform (Sections: "IAM permissions")
  • Configure GKE for ML Diagnostics (Sections: "Set up with gcloud CLI, Google Cloud console, or Terraform", "Manual installation", "Connection-operator")
  • Workload monitoring with ML Diagnostics (Sections: "Workload Monitoring and analyzers", "HBM Capacity Analyzer", "Host Memory Utilization Analyzer", "CPU Utilization Analyzer", "Access Workload Monitoring information through the API", "List all monitored events", "System Metrics")
  • Get started with the ML Diagnostics CLI (Sections: "List machine learning runs", "Monitored-events commands", "List profiler targets", "Capture on-demand profiler sessions")
  • Scale container resource requests and limits (Sections: "Identify workloads without resource requests or limits")
  • Auto-repair nodes (Sections: "Repair criteria", "Verify node auto-repair is enabled for a Standard node pool", "Enable auto-repair for an existing Standard node pool", "Node auto repair in TPU slice nodes")
  • Deploy TPU workloads in GKE Standard (Sections: "Configure auto repair for TPU slice nodes")
  • Get support (Sections: "Before you contact support")

Read-only rule: Run read-only diagnostic commands only. Never drain, delete, or re-create nodes, or run any other command that changes the cluster. Give the user any fix to apply themselves.

When you recommend a fix, link the doc section that describes it.

---

Analyzer routing overview

By default, Workload Monitoring treats a 15% drop in TPU duty cycle as a performance degradation, raises a MonitoredEvent (type: PERFORMANCE_DEGRADATION), and runs its analyzers. Consult the Workload Monitoring and analyzers section in the official Cloud TPU documentation for the complete analyzer definitions, detection criteria, and subsections, and route remediation by category:

• Category A: workload resource bottlenecks (never cordon or replace nodes):
  • When the firing analyzer (detectionState: "DETECTED") indicates a workload memory or host CPU capacity bottleneck (see the "HBM Capacity Analyzer", "Host Memory Utilization Analyzer", and "CPU Utilization Analyzer" sections of Workload monitoring with ML Diagnostics), route to Path A: Workload resource bottlenecks (workload optimization and profiling).
• Category B: infrastructure and network fabric throttling (culprit nodes):
  • When the firing analyzer (detectionState: "DETECTED") indicates an interconnect, thermal/power throttling, memory bandwidth, or Top-of-Rack (ToR) network fault on specific instances (see Workload Monitoring and analyzers), route to Path B: Infrastructure or network fabric throttling (map the culprit Compute Engine instance IDs to GKE nodes, handle node repair through GKE, and escalate to Google Cloud Support).
• Category C: PERFORMANCE_DEGRADATION event fired, but no analyzer reports detectionState: "DETECTED":
  • When a PERFORMANCE_DEGRADATION event exists (as in the sample output in List all monitored events, where analyzerReports entries have detectionState: "NOT_DETECTED"), route to Path C: Event fired with no DETECTED analyzer.

---

Diagnostic workflow

Step 0: Collect context and set the investigation window [Low Risk]

Collect the target parameters. By default, query a 60-minute window `[T - 30m, T

• 30m]around{issue_time}`:

• {project_id}: Google Cloud project ID
• {location}: Google Cloud region where the ML run and GKE cluster reside (for example, us-central1)
• {cluster_name}: GKE cluster name
• {workload_name} / {ml_run_id}: JobSet / workload name or ML Diagnostics run ID
• {issue_time}: Timestamp when throughput degradation was observed (T, ISO-8601 UTC)
• {start_time}: T - 30m
• {end_time}: T + 30m

---

Step 1: Query ML runs and monitoredEvents [Low Risk]

1. List active or recent ML runs: Give the user gcloud alpha mldiagnostics machine-learning-run list with a link to List machine learning runs in the ML Diagnostics CLI reference, and locate {ml_run_id} matching {workload_name}.
2. List performance degradation events: Give the user gcloud alpha mldiagnostics monitored-events list with a link to Monitored-events commands, or the hypercomputecluster.googleapis.com/v1alpha API with a link to Access Workload Monitoring information through the API, to check for PERFORMANCE_DEGRADATION events during [{start_time}, {end_time}].
3. Describe the MonitoredEvent: Give the user gcloud alpha mldiagnostics monitored-events describe with a link to the same Monitored-events commands section, and inspect the analyzerReports array (analyzer, detectionState, details, and recommendedActions).
   • If a PERFORMANCE_DEGRADATION event fired, run Step 2 to corroborate with the 1-minute system metrics; if none of its analyzerReports entries has detectionState: "DETECTED", follow Path C: Event fired with no DETECTED analyzer.
   • If no PERFORMANCE_DEGRADATION event exists and duty cycle is steady in Step 2, rule out TPU performance degradation by following Path D: Healthy telemetry.

---

Step 2: Correlate with 1-minute Cloud Monitoring system metrics [Low Risk]

Consult the System Metrics section in the Workload Monitoring documentation for the 1-minute Cloud Monitoring metrics exported for TPU and host devices, and run read-only PromQL queries over [{start_time}, {end_time}] to corroborate the analyzer report:

# 1. Node TPU duty cycle (look for the drop on the affected nodes)
kubernetes_io:node_accelerator_duty_cycle{
  monitored_resource="k8s_node",
  project_id="{project_id}",
  cluster_name="{cluster_name}"
}

# 2. HBM utilization ratio by node (around 0.90 is approaching the limit)
sum by (node_name) (
  kubernetes_io:node_accelerator_memory_used{
    monitored_resource="k8s_node",
    project_id="{project_id}",
    cluster_name="{cluster_name}"
  }
)
/
sum by (node_name) (
  kubernetes_io:node_accelerator_memory_total{
    monitored_resource="k8s_node",
    project_id="{project_id}",
    cluster_name="{cluster_name}"
  }
)

# 3. Host memory and CPU allocatable utilization
kubernetes_io:node_memory_allocatable_utilization{
  monitored_resource="k8s_node",
  project_id=