Google Cluster Toolkit (formerly HPC Toolkit)

Cluster Toolkit is an open-source software provided by Google Cloud that makes it easy to deploy AI/ML and high performance computing (HPC) environments following Google Cloud best practices.

Cluster Toolkit is highly customizable and extensible, addressing the deployment needs of a broad range of workloads, such as compute, networking, and storage, in a repeatable manner.

Cluster Toolkit comes with a suite of tutorials, examples, and full documentation for modules designed for AI/ML and HPC use cases.

The main components of Cluster Toolkit include:

• Cluster blueprints: YAML files that define the cluster's infrastructure and configuration.
• Modules: Reusable building blocks (Terraform or Packer) used to compose a blueprint.
• gcluster engine: The command-line tool that processes blueprints to create a deployment folder.
• Deployment folder: A self-contained folder containing the Terraform or Packer code needed to provision the environment.

For more information, see Google Cluster Toolkit overview.

Cluster Toolkit is an integral part of Google Cloud AI Hypercomputer. For more information about GKE and Slurm deployments in AI Hypercomputer, see Create an AI-optimized GKE cluster with default configuration and Create a self-managed Slurm cluster for AI workloads.

To get started with Cluster Toolkit deployments, you can follow one of the following quickstart guides:

• Deploy an HPC cluster with Slurm.
• Create a self-managed Slurm cluster with an A4 VM.
• Create a Cloud RDMA-enabled HPC Slurm cluster with H4D instances.

To create a cluster using Cluster Toolkit, you can use either Cloud Shell, or a workstation that is running Linux or macOS.

Cloud Shell is an interactive development and operations environment that is accessible from your web browser. If you use Cloud Shell, the following dependencies are already pre-installed and you don't need to manually install dependencies.

If you want to work from a Linux or macOS client or workstation, you must follow the steps in Install dependencies before you continue.

You can use two different methods to install Cluster Toolkit:

• Using the pre-built bundle (recommended)
• Building from source

For the easiest setup, download the appropriate bundle for your operating system and architecture (e.g., gcluster_bundle_linux_amd64.zip, gcluster_bundle_linux_arm64.zip, gcluster_bundle_mac_amd64.zip, or gcluster_bundle_mac_arm64.zip) from the Releases page. These bundles include the pre-compiled gcluster binary, the examples folder, and the community/examples folder.

The pre-built bundles are compiled for Linux and macOS execution environments and support the deployment of the following cluster operating systems.

1. Download and extract the bundle:

   For versions v1.89.0 and newer (Multi-architecture Tarball):

   # Find all available releases at: https://github.com/GoogleCloudPlatform/cluster-toolkit/releases
   # Set the desired version TAG (e.g., v1.89.0)
   TAG=vX.Y.Z
   # Set your OS (linux or mac) and architecture (amd64 or arm64)
   OS="linux"
   ARCH="amd64"
   # Download and extract the platform-specific bundle in a single step
   mkdir -p cluster-toolkit && curl -L https://github.com/GoogleCloudPlatform/cluster-toolkit/releases/download/${TAG}/gcluster_bundle_${OS}_${ARCH}.tgz | tar -xz -C cluster-toolkit && cd cluster-toolkit

   For versions v1.85.0 through v1.88.0 (Multi-architecture Zip):

   # Find all available releases at: https://github.com/GoogleCloudPlatform/cluster-toolkit/releases
   # Set the desired version TAG (e.g., v1.85.0)
   TAG=vX.Y.Z
   # Set your OS (linux or mac) and architecture (amd64 or arm64)
   OS="linux"
   ARCH="amd64"
   # Download and extract the platform-specific bundle
   curl -LO https://github.com/GoogleCloudPlatform/cluster-toolkit/releases/download/${TAG}/gcluster_bundle_${OS}_${ARCH}.zip
   unzip gcluster_bundle_${OS}_${ARCH}.zip -d cluster-toolkit/
   cd cluster-toolkit

2. Verify the installation:

   ./gcluster --version
   ./gcluster --help

If you prefer to build the gcluster binary from source, you can use the following commands:

git clone https://github.com/GoogleCloudPlatform/cluster-toolkit
cd cluster-toolkit
make
./gcluster --version
./gcluster --help

Before deploying your first cluster, ensure the following are configured in your Google Cloud project.

Several APIs must be enabled to deploy your cluster. While Terraform identifies missing APIs during terraform apply, enabling them upfront saves time. Required APIs typically include:

• Compute Engine API
• Filestore API
• Cloud Storage API
• Service Usage API

For more information, see Set up Cluster Toolkit.

HPC and AI workloads often require significant resources. You might need to request additional quota to deploy your cluster. For more information, see Request additional quotas.

Terraform can provide credentials for authenticating to Google Cloud in several ways. We recommend using gcloud on your workstation or using service accounts attached to cloud environments.

On your local terminal, Cloud Workstations, or Cloud Shell, generate Application Default Credentials (ADC) associated with your Google Cloud account:

gcloud auth application-default login

Follow the prompts in your browser to authenticate. You are provided a token to copy and paste back into your terminal to complete the process. Once finished, Terraform automatically uses these "Application Default Credentials."

If you receive a quota project error, then set the quota project to your current project ID:

gcloud auth application-default set-quota-project ${PROJECT_ID}

To help improve Cluster Toolkit, feature usage statistics are collected and sent to Google. You can opt-out at any time by executing the following command:

./gcluster telemetry off

Cluster Toolkit telemetry overall is handled in accordance with the Google Privacy Policy. When you use Cluster Toolkit to interact with or utilize GCP Services, your information is handled in accordance with the Google Cloud Privacy Notice.

After installing gcluster, you can deploy, manage, and destroy infrastructure using cluster blueprints:

1. Create deployment folder: Use a blueprint YAML file to generate deployment files:

   ./gcluster create examples/hpc-slurm.yaml

2. Deploy cluster: Provision infrastructure with Terraform:

   ./gcluster deploy hpc-slurm

3. Destroy cluster: Clean up provisioned resources:

   ./gcluster destroy hpc-slurm

For detailed guides, blueprint syntax, and configuration options, see the Google Cloud documentation site.

The gcluster job submit command provides a unified interface to submit batch and distributed containerized workloads (JobSets) to your GKE clusters.

Ensure your environment is set up before submitting jobs:

• A deployed GKE cluster managed by Cluster Toolkit (with Kueue and JobSet enabled).
• kubectl and gke-gcloud-auth-plugin installed and authenticated (gcloud auth login).
• For on-the-fly builds (--build-context), set export GCLUSTER_IMAGE_REPO=<repository-name> to specify your Artifact Registry repository name (e.g., export GCLUSTER_IMAGE_REPO=gcluster-repo).

There are two ways to submit a JobSet with the gcluster job submit command:

• Submit with a pre-built image: Submit a job using an existing container image from an image registry.
• Submit with on-the-fly image building: Package local application code into a container image and push it automatically.

To submit with a pre-built image, run ./gcluster job submit with the following properties in your shell:

./gcluster job submit \
  --project <PROJECT_ID> \
  --cluster <CLUSTER_NAME> \
  --location <REGION_OR_ZONE> \
  --name my-job \
  --image us-docker.pkg.dev/samples/gke/hello-app:1.0 \
  --command "echo Hello from Cluster Toolkit" \
  --compute-type n2-standard-32

To submit with on-the-fly image building, run ./gcluster job submit with the following properties in your shell:

./gcluster job submit \
  --project <PROJECT_ID> \
  --cluster <CLUSTER_NAME> \
  --location <REGION_OR_ZONE> \
  --name my-python-job \
  --base-image python:3.11-slim \
  --build-context ./app_dir \
  --command "python main.py" \
  --compute-type n2-standard-32

You can manage submitted jobs directly using the gcluster job commands:

• List jobs: Check workload status across the cluster.
• View job logs: View container stdout and stderr logs.
• Cancel a job: Clean up a running workload.

Check workload status across the cluster:

./gcluster job list

View container stdout and stderr logs:

./gcluster job logs my-job

Clean up a running workload:

./gcluster job cancel my-job

For complete step-by-step tutorials, advanced multi-slice topologies, storage mounts, and Kueue queue management, refer to the full Gcluster Job Submission Guide.

The gcluster job submit command integrates natively with advanced Google Kubernetes Engine (GKE) hardware and scheduling capabilities:

• Dynamic TPU slicing: Reconfigure and pack TPU v7x (and future TPU generation) nodes into logical slices dynamically (supporting both superslicing and subslicing) by using the --topology flag and Kueue Topology-Aware Scheduling.
• Pathways distributed AI orchestration: Compile and deploy multi-slice Pathways training jobs (Resource Manager, Proxy, Workers, and optional headless mode) by using the --pathways and --pathways-gcs-location flags.
• GKE node auto-provisioning (NAP): Automatically scale and provision compute nodes dynamically by using the --gke-nap-provisioning (spot, reservation, on-demand) and --gke-nap-reservation flags.

For cluster blueprint Terraform deployment instructions and comprehensive workload orchestration guidelines, see the guide to advanced GKE infrastructure features and the gcluster job submission guide.

Cluster Toolkit provides specialized modules for Slurm images, and support for standard OS images.

Cluster Toolkit provides specialized modules for Slurm.

The toolkit also supports standard OS images for general-purpose modules:

• HPC Rocky Linux 8
• Debian 11
• Ubuntu 22.04 LTS

For more details, see VM image support.

A cluster blueprint is the core configuration file (YAML) for your deployment. Cluster Toolkit includes validator functions that perform basic tests on the blueprint to ensure variables are valid and resources can be provisioned. See Blueprint validation for more details.

To track the costs of your deployment, use the Cloud Billing Reports page. You need a role with the billing.accounts.getSpendingInformation permission.

1. In the Google Cloud Console, go to Billing.
2. Select Reports.
3. In the Filters pane on the right, filter by label using the key ghpc_deployment or ghpc_blueprint and specify your deployment name.

Ensure you have properly set up Google Cloud credentials.

See Slurm Troubleshooting.

Common deployment failures:

• Project Access: Ensure your account has the necessary roles in the IAM section of the console.
• Filestore resource limit: If you see "System limit for internal resources has been reached," see the Filestore troubleshooting guide for the solution.

Install the following tools to ensure your changes pass validation:

• pre-commit
• TFLint (requires version compatible with .tflint.hcl)
• ShellCheck

Additional development dependencies can be installed with a single command:

make install-dev-deps
pre-commit install

Please refer to CONTRIBUTING.md.