GPU GitOps for OpenShift / Kubernetes

This repository uses GitOps to deploy NVIDIA GPU support into your cluster. It automates:

• 👉 Node Feature Discovery (NFD) to detect GPU nodes
• 👉 NVIDIA GPU Operator to install drivers, device plugin, runtime, monitoring, etc.
• 👉 NVIDIA DCGM Dashboard to visualize GPU telemetry exposed in Prometheus by using the NVIDIA DCGM Exporter.

✅ Prerequisites

1. A Kubernetes or OpenShift cluster with at least one GPU-capable node.
2. oc (or kubectl) logged into the cluster as a cluster-admin.
3. kustomize CLI (optional if you use oc apply -k).

---

🛠️ Deploying via GitOps

You can bootstrap these components declaratively without cloning this repo with:

# 1. Log in
oc login <your-cluster-url> --token=<token> --server=<api-url>

# 2. Bootstrap GPU GitOps
oc apply -f https://github.com/na-launch/gpu-gitops/main/gitops/manifests/cluster/bootstrap/base/bigbang-app.yaml

# 3. Monitor progress
oc get pods -n nvidia-gpu-operator -w
oc get nodes -l feature.node.kubernetes.io/pci-10de.present=true -o name

# 4. Verify GPU functionality
oc -n nvidia-gpu-operator get pods -l openshift.driver-toolkit=true -o name | while read pod; do
  echo "$pod"
  oc exec -n nvidia-gpu-operator -it "$pod" -- nvidia-smi
  echo
done

---

📊 Viewing GPU Metrics

In the OpenShift Container Platform web console:

• Go to Observe > Dashboards → select NVIDIA DCGM Exporter Dashboard from the list.

The NVIDIA DCGM Exporter Dashboard displays GPU-related graphs such as utilization, temperature, memory usage, and more.

🛡️ Next Steps

Once the GPU infrastructure is installed:

• Create GPU-capable workloads using resources.limits.nvidia.com/gpu: 1.
• Integrate with ML platform (e.g. OpenDataHub or OpenShift AI).
• Deploy monitoring tools like Prometheus to collect DCGM metrics from GPUs.
• Scale your GPU node pools and GPU workloads using standard GitOps.

---

🔗 References

• NVIDIA GPU Operator + NFD setup with GitOps workflows
• GPU detection using NFD labels (pci-10de)
• Typical GPU Operator installation patterns and validation via nvidia-smi