openshift-eng/scripts

Node Tuning Helper Scripts

Detailed instructions for invoking the helper utilities that back /node-tuning commands:

• generate_tuned_profile.py renders Tuned manifests (tuned.openshift.io/v1).
• analyze_node_tuning.py inspects live nodes or sosreports for tuning gaps.

When to Use These Scripts

• Translate structured command inputs into Tuned manifests for the Node Tuning Operator.
• Iterate on generated YAML outside the assistant or integrate the generator into automation.
• Analyze CPU isolation, IRQ affinity, huge pages, sysctl values, and networking counters from live clusters or archived sosreports.

Prerequisites

• Python 3.8 or newer (python3 --version).
• Repository checkout so the scripts under plugins/node-tuning/skills/scripts/ are accessible.
• Optional: oc CLI when validating or applying manifests.
• Optional: Extracted sosreport directory when running the analysis script offline.
• Optional (remote analysis): oc CLI access plus a valid KUBECONFIG when capturing /proc//sys or sosreport via oc debug node/<name>. The sosreport workflow pulls the registry.redhat.io/rhel9/support-tools image (override with --toolbox-image or TOOLBOX_IMAGE) and requires registry access. HTTP(S) proxy env vars from the host are forwarded automatically when present, but using a proxy is optional.

---

Script: generate_tuned_profile.py

Implementation Steps

1. Collect Inputs

   • --profile-name: Tuned resource name.
   • --summary: [main] section summary.
   • Repeatable options: --include, --main-option, --variable, --sysctl, --section (SECTION:KEY=VALUE).
   • Target selectors: --machine-config-label key=value, --match-label key[=value].
   • Optional: --priority (default 20), --namespace, --output, --dry-run.
   • Use --list-nodes/--node-selector to inspect nodes and --label-node NODE:KEY[=VALUE] (plus --overwrite-labels) to tag machines.

2. Inspect or Label Nodes (optional)

   # List all worker nodes
   python3 plugins/node-tuning/skills/scripts/generate_tuned_profile.py --list-nodes --node-selector "node-role.kubernetes.io/worker" --skip-manifest
   
   # Label a specific node for the worker-hp pool
   python3 plugins/node-tuning/skills/scripts/generate_tuned_profile.py \
     --label-node ip-10-0-1-23.ec2.internal:node-role.kubernetes.io/worker-hp= \
     --overwrite-labels \
     --skip-manifest
   

3. Render the Manifest

   python3 plugins/node-tuning/skills/scripts/generate_tuned_profile.py \
     --profile-name "$PROFILE" \
     --summary "$SUMMARY" \
     --sysctl net.core.netdev_max_backlog=16384 \
     --match-label tuned.openshift.io/custom-net \
     --output .work/node-tuning/$PROFILE/tuned.yaml
   

   • Omit --output to write <profile-name>.yaml in the current directory.
   • Add --dry-run to print the manifest to stdout.

4. Review Output

   • Inspect the generated YAML for accuracy.
   • Optionally format with yq or open in an editor for readability.

5. Validate and Apply

   • Dry-run: oc apply --server-dry-run=client -f <manifest>.
   • Apply: oc apply -f <manifest>.

Error Handling

• Missing required options raise ValueError with descriptive messages.
• The script exits non-zero when no target selectors (--machine-config-label or --match-label) are supplied.
• Invalid key/value or section inputs identify the failing argument explicitly.

Examples

python3 plugins/node-tuning/skills/scripts/generate_tuned_profile.py \
  --profile-name realtime-worker \
  --summary "Realtime tuned profile" \
  --include openshift-node --include realtime \
  --variable isolated_cores=1 \
  --section bootloader:cmdline_ocp_realtime=+systemd.cpu_affinity=${not_isolated_cores_expanded} \
  --machine-config-label machineconfiguration.openshift.io/role=worker-rt \
  --priority 25 \
  --output .work/node-tuning/realtime-worker/tuned.yaml

python3 plugins/node-tuning/skills/scripts/generate_tuned_profile.py \
  --profile-name openshift-node-hugepages \
  --summary "Boot time configuration for hugepages" \
  --include openshift-node \
  --section bootloader:cmdline_openshift_node_hugepages="hugepagesz=2M hugepages=50" \
  --machine-config-label machineconfiguration.openshift.io/role=worker-hp \
  --priority 30 \
  --output .work/node-tuning/openshift-node-hugepages/hugepages-tuned-boottime.yaml

---

Script: analyze_node_tuning.py

Purpose

Inspect either a live node (/proc, /sys) or an extracted sosreport snapshot for tuning signals (CPU isolation, IRQ affinity, huge pages, sysctl state, networking counters) and emit actionable recommendations.

Usage Patterns

• Live node analysis

  python3 plugins/node-tuning/skills/scripts/analyze_node_tuning.py --format markdown
  

• Remote analysis via oc debug

  python3 plugins/node-tuning/skills/scripts/analyze_node_tuning.py \
    --node worker-rt-0 \
    --kubeconfig ~/.kube/prod \
    --format markdown
  

• Collect sosreport via oc debug and analyze locally

  python3 plugins/node-tuning/skills/scripts/analyze_node_tuning.py \
    --node worker-rt-0 \
    --toolbox-image registry.example.com/support-tools:latest \
    --sosreport-arg "--case-id=01234567" \
    --sosreport-output .work/node-tuning/sosreports \
    --format json
  

• Offline sosreport analysis

  python3 plugins/node-tuning/skills/scripts/analyze_node_tuning.py \
    --sosreport /path/to/sosreport-2025-10-20
  

• Automation-friendly JSON

  python3 plugins/node-tuning/skills/scripts/analyze_node_tuning.py \
    --sosreport /path/to/sosreport \
    --format json --output .work/node-tuning/node-analysis.json
  

Implementation Steps

1. Select data source
   • Provide --node <name> (with optional --kubeconfig / --oc-binary). By default the helper runs sosreport remotely from inside the RHCOS toolbox container (registry.redhat.io/rhel9/support-tools). Override the image with --toolbox-image, extend the sosreport command with --sosreport-arg, or disable the curated OpenShift flags via --skip-default-sosreport-flags. Pass --no-collect-sosreport to fall back to the direct /proc snapshot mode.
   • Provide --sosreport <dir> for archived diagnostics; detection finds embedded proc/ and sys/.
   • Omit both switches to query the live filesystem (defaults to /proc and /sys).
   • Override paths with --proc-root or --sys-root when the layout differs.
2. Run analysis
   • The script parses cpuinfo, kernel cmdline parameters (isolcpus, nohz_full, tuned.non_isolcpus), default IRQ affinities, huge page counters, sysctl values (net, vm, kernel), transparent hugepage settings, netstat/sockstat counters, and ps snapshots (when available in sosreport).
3. Review the report
   • Markdown output groups findings by section (System Overview, CPU & Isolation, Huge Pages, Sysctl Highlights, Network Signals, IRQ Affinity, Process Snapshot) and lists recommendations.
   • JSON output contains the same information in structured form for pipelines or dashboards.
4. Act on recommendations
   • Apply Tuned profiles, MachineConfig updates, or manual sysctl/irqbalance adjustments.
   • Feed actionable items back into /node-tuning:generate-tuned-profile to codify desired state.

Error Handling

• Missing proc/ or sys/ directories trigger descriptive errors.
• Unreadable files are skipped without raising an error and noted in observations where relevant.
• Non-numeric sysctl values are flagged for manual investigation.

Example Output (Markdown excerpt)

# Node Tuning Analysis

## System Overview
- Hostname: worker-rt-1
- Kernel: 4.18.0-477.el8
- NUMA nodes: 2
- Kernel cmdline: `BOOT_IMAGE=... isolcpus=2-15 tuned.non_isolcpus=0-1`

## CPU & Isolation
- Logical CPUs: 32
- Physical cores: 16 across 2 socket(s)
- SMT detected: yes
- Isolated CPUs: 2-15
...

## Recommended Actions
- Configure net.core.netdev_max_backlog (>=32768) to accommodate bursty NIC traffic.
- Transparent Hugepages are not disabled (`[never]` not selected). Consider setting to `never` for latency-sensitive workloads.
- 4 IRQs overlap isolated CPUs. Relocate interrupt affinities using tuned profiles or irqbalance.

Follow-up Automation Ideas

• Persist JSON results in .work/node-tuning/<host>/analysis.json for historical tracing.
• Gate upgrades by comparing recommendations across nodes.
• Integrate with CI jobs that validate cluster tuning post-change.