Files
wild-central-api/internal/setup/cluster-services/nvidia-device-plugin/README.md

98 lines
2.9 KiB
Markdown

# NVIDIA Device Plugin
The NVIDIA Device Plugin for Kubernetes enables GPU scheduling and resource management on nodes with NVIDIA GPUs.
## Overview
This service deploys the official NVIDIA Device Plugin as a DaemonSet that:
- Discovers NVIDIA GPUs on worker nodes
- Labels nodes with GPU product information (e.g., `nvidia.com/gpu.product=GeForce-RTX-4090`)
- Advertises GPU resources (`nvidia.com/gpu`) to the Kubernetes scheduler
- Enables pods to request GPU resources
## Prerequisites
Before installing the NVIDIA Device Plugin, ensure that:
1. **NVIDIA Drivers** are installed (>= 384.81)
2. **nvidia-container-toolkit** is installed (>= 1.7.0)
3. **nvidia-container-runtime** is configured as the default container runtime
4. Worker nodes have NVIDIA GPUs
### Talos Linux Requirements
For Talos Linux nodes, you need:
- NVIDIA drivers extension in the Talos schematic
- nvidia-container-toolkit extension
- Proper container runtime configuration
## Installation
```bash
# Configure and install the service
wild-cluster-services-configure nvidia-device-plugin
wild-cluster-install nvidia-device-plugin
```
## Verification
After installation, verify the plugin is working:
```bash
# Check plugin pods are running
kubectl get pods -n kube-system | grep nvidia
# Verify GPU resources are advertised
kubectl get nodes -o json | jq '.items[].status.capacity | select(has("nvidia.com/gpu"))'
# Check GPU node labels
kubectl get nodes --show-labels | grep nvidia
```
## Usage in Applications
Once installed, applications can request GPU resources:
```yaml
apiVersion: apps/v1
kind: Deployment
metadata:
name: gpu-app
spec:
template:
spec:
containers:
- name: app
image: nvidia/cuda:latest
resources:
requests:
nvidia.com/gpu: 1
limits:
nvidia.com/gpu: 1
```
## Troubleshooting
### Plugin Not Starting
- Verify NVIDIA drivers are installed on worker nodes
- Check that nvidia-container-toolkit is properly configured
- Ensure worker nodes are not tainted in a way that prevents scheduling
### No GPU Resources Advertised
- Check plugin logs: `kubectl logs -n kube-system -l name=nvidia-device-plugin-ds`
- Verify NVIDIA runtime is the default container runtime
- Ensure GPUs are detected by the driver: check node logs for GPU detection messages
## Configuration
The plugin uses the following configuration:
- **Image**: `nvcr.io/nvidia/k8s-device-plugin:v0.17.1`
- **Namespace**: `kube-system`
- **Priority Class**: `system-node-critical`
- **Tolerations**: Schedules on nodes with `nvidia.com/gpu` taint
## References
- [Official NVIDIA Device Plugin Repository](https://github.com/NVIDIA/k8s-device-plugin)
- [Kubernetes GPU Scheduling Documentation](https://kubernetes.io/docs/tasks/manage-gpus/scheduling-gpus/)
- [NVIDIA Container Toolkit Documentation](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/)