GoogleCloudPlatform/cluster-toolkit

What it solves

Deploying AI/ML and high-performance computing (HPC) environments on Google Cloud is often complex, requiring precise configuration of compute, networking, and storage. Cluster Toolkit simplifies this by providing a repeatable, customizable way to provision these environments following Google Cloud best practices.

How it works

The toolkit uses a combination of configuration files and automation tools to build infrastructure:

• Cluster Blueprints: YAML files that define the desired infrastructure and configuration.
• Modules: Reusable building blocks based on Terraform or Packer.
• gcluster engine: A command-line tool that processes blueprints to generate a deployment folder containing the necessary Terraform or Packer code.
• Job Submission: The gcluster job submit command allows users to deploy containerized workloads (JobSets) to GKE clusters, with support for on-the-fly image building and advanced hardware orchestration.

Who it’s for

It is designed for engineers and researchers deploying AI/ML workloads or HPC clusters (such as those using Slurm) on Google Cloud Platform.

Highlights

• AI Hypercomputer Integration: An integral part of the Google Cloud AI Hypercomputer ecosystem.
• Advanced GKE Orchestration: Supports dynamic TPU slicing, Pathways distributed AI orchestration, and GKE node auto-provisioning.
• Flexible Deployment: Supports both GKE and self-managed Slurm clusters.
• Unified Job Interface: Provides a single CLI for submitting, listing, and managing batch and distributed containerized workloads.