kind-gpu-sim

Simulate NVIDIA or AMD (ROCm) GPUs in a Kubernetes in Docker (kind) cluster, without requiring actual GPU hardware.

This is perfect for:

• Testing GPU scheduling
• Validating device plugin behavior
• Learning how GPU workloads interact with Kubernetes
• Building GPU-related Kubernetes infrastructure (where no real workloads are required).

⚠️ Important: No Real GPU Support

This project simulates the presence of GPU resources in a Kind cluster. It does not provide access to actual GPU hardware, and real GPU workloads (like CUDA or ROCm kernels) will not run.

Prerequisites

Make sure the following tools are installed on your system before running the GPU simulator script:

Tool	Purpose
docker OR podman	Required by kind, runs the local registry and all cluster nodes
kind	Creates the local Kubernetes cluster inside Docker
kubectl	CLI to interact with the Kubernetes cluster
git	Clones the GPU device plugin repositories (NVIDIA / ROCm)
sed	Used to patch Dockerfiles for public registry compatibility

Features

• Kind cluster with 1 control-plane + 2 workers
• Simulated amd.com/gpu or nvidia.com/gpu resources
• Automatically taints and labels GPU nodes
• Uses a local container registry
• Builds and deploys the AMD ROCm device plugin (locally)
• Builds and deploys NVIDIA plugin (locally)
• Includes GPU test pod manifests

Quick Start

1. Clone the repo and make the script executable

chmod +x kind-gpu-sim.sh

2. Start the simulated GPU cluster

Choose your simulation type:

# Simulate AMD GPUs
./kind-gpu-sim.sh create rocm

# Simulate NVIDIA GPUs
./kind-gpu-sim.sh create nvidia

3. (Optional) Test a simulated GPU pod

Create a pod that requests GPU resources:

For NVIDIA

kubectl create -f pods/nvidia-gpu-test-pod.yaml

Check pod logs

kubectl logs nvidia-gpu-test
Hello from fake NVIDIA GPU node

For AMD

kubectl create -f pods/rocm-gpu-test-pod.yaml

Check pod logs

kubectl logs gpu-rocm-test
Hello from fake ROCm GPU node

4. Tear down the cluster

./kind-gpu-sim.sh delete

File Structure

.
├── kind-gpu-config.yaml          # Kind cluster config: 1 control-plane, 2 workers
├── kind-gpu-sim.sh               # Main script to create/delete simulated GPU clusters (ROCm or NVIDIA)
├── pods
│   ├── nvidia-gpu-test-pod.yaml  # Pod spec to test NVIDIA GPU simulation (uses nvidia.com/gpu)
│   ├── rocm-gpu-test-pod.yaml    # Pod spec to test AMD ROCm GPU simulation (uses amd.com/gpu)
│   └── triton-pod.yaml           # Pod that installs and runs Triton-lang, useful for simulating kernel compilation
└── Readme.md                     # Project overview and usage instructions

How It Works

Component	Description
kubectl patch	Fakes amd.com/gpu or nvidia.com/gpu on nodes
taint + toleration	Ensures only GPU workloads land on simulated nodes
DaemonSet	Deploys either AMD or NVIDIA device plugin DaemonSets
localhost:5000	Local registry, connected to Kind

Tested With

• kind v0.23.0

Why Simulate?

This project helps:

• Devs test GPU workloads without expensive hardware
• CI environments validate GPU scheduling logic
• Anyone learn Kubernetes GPU primitives

Loading container images to the kind cluster

# ./kind-gpu-sim.sh load --image-name=<Image-Name> --cluster-name=<KIND_CLUSTER_NAME)>
# for e.g.
./kind-gpu-sim.sh load --image-name=public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:v0.9.1