CNCF Atlas

English

日本語

English

日本語

Appearance

Getting Started ​

Verified against the README at commit 2487a24 (near tag v2.9.0). Commands assume a Kubernetes cluster with NVIDIA GPU nodes and Helm.

Prerequisites ​

For the NVIDIA device-plugin path, the README lists (README.md:94-102):

  • NVIDIA driver >= 440
  • nvidia-docker version > 2.0
  • NVIDIA configured as the default runtime for containerd, Docker, or CRI-O
  • Kubernetes >= 1.23
  • glibc >= 2.17 and < 2.30
  • Linux kernel >= 3.10
  • Helm > 3.0

Install ​

Label the GPU nodes so HAMi manages only those, then install the chart into kube-system (README.md:104-134):

bash
kubectl label nodes <node-name> gpu=on

helm repo add hami-charts https://project-hami.github.io/HAMi/
helm repo update

helm install hami hami-charts/hami -n kube-system

A first working setup ​

The shortest path to a shared GPU is one labelled node, the chart, and the bundled example pod.

  1. Confirm the scheduler and device plugin are running.

    bash
    kubectl get pods -n kube-system

Wait until both hami-scheduler and hami-device-plugin show Running.

  1. Submit the example workload, which asks for one GPU with a memory and core budget.

    bash
    kubectl apply -f examples/nvidia/default_use.yaml

The example pod requests nvidia.com/gpu: 1, nvidia.com/gpumem: 3000, and nvidia.com/gpucores: 30, so it takes one physical GPU limited to 3000 MB and 30 percent of the cores (examples/nvidia/default_use.yaml).

Verify it works ​

Check that the pod scheduled and is running:

bash
kubectl get pod gpu-pod

Inside the container, a tool like nvidia-smi reports the capped memory rather than the card's full amount, because HAMi-core enforces the injected CUDA_DEVICE_MEMORY_LIMIT. Cluster-wide GPU usage is exposed by the scheduler monitor, whose default port is 31993 (README.md:152-158):

text
http://<scheduler-ip>:31993/metrics

Where to go next ​

For scheduling policies (binpack, spread, topology-aware, dynamic MIG), Volcano and Koordinator integration, WebUI, and production configuration such as high availability and per-vendor setup, follow the official documentation at https://project-hami.io/docs/. The complete Helm install guide is at https://project-hami.io/docs/get-started/deploy-with-helm/.