Docker: Cluster Setup

Public Docker Image

One can specify a Docker Image through the Runhouse Image class, which is passed into the cluster factory. Call .from_docker(image_id) on the image, passing in the Docker container in the format <registry>/<image>:<tag>.

base_image = rh.Image("base_image").from_docker("nvcr.io/nvidia/pytorch:23.10-py3")

docker_cluster = rh.ondemand_cluster(
    name="pytorch_cluster",
    image=base_image,
    instance_type="CPU:2+",
    provider="aws",
)

Private Docker Image

To use a Docker image hosted on a private registry, such as ECR, you need to additionally provide the user, password, and registry server values, as used in docker login -u <user> -p <password> <registry server>.

These values are propagated to SkyPilot at launch time, which will be used for setting up the base container on the cluster.

There are two approaches to providing this information:

1. Creating a runhouse Secret as follows, and pass it to the Image along with the Docker image above.

values = {
    "username": <user>,
    "password": <password>,
    "server": <server>,
}
docker_secret = rh.provider_secret("docker", values=values)

base_image = rh.Image("base_image").from_docker(
    "pytorch-training:2.2.0-cpu-py310-ubuntu20.04-ec2", docker_secret=docker_secret
)

2. Directly set your local environment variables, as expected by and extracted by Skypilot during launch time. In this case, you do not need to specify the secret during the runhouse Image construction.

   • SKYPILOT_DOCKER_USERNAME: <user>

   • SKYPILOT_DOCKER_PASSWORD: <password>

   • SKYPILOT_DOCKER_SERVER: <registry server>

For instance, to use the PyTorch2.2 ECR Framework provided here, you can set your environment variables as so:

$ export SKYPILOT_DOCKER_USERNAME=AWS
$ export SKYPILOT_DOCKER_PASSWORD=$(aws ecr get-login-password --region us-east-1)
$ export SKYPILOT_DOCKER_SERVER=763104351884.dkr.ecr.us-east-1.amazonaws.com

base_image = rh.Image("base_image").from_docker("pytorch-training:2.2.0-cpu-py310-ubuntu20.04-ec2")

In either case, we can then construct the cluster with the image as follows:

ecr_cluster = rh.ondemand_cluster(
    name="ecr_pytorch_cluster",
    image_id=base_image,
    instance_type="CPU:2+",
    provider="aws",
)

SSH

To SSH directly onto the container, where the Runhouse server is started, you can use runhouse cluster ssh <cluster_name>.

User and Container Name

By default, the remote Docker container, which is set up through Skypilot, will be named sky_container, and the user will be root.