GCP Supercomputer Solutions Support (8)

Google Cloud Platform Supercomputer Solutions

Google is seeking a supplier to provide engineering, maintenance, and enhancement services for its Google Cloud Platform ("GCP") Supercomputer Solutions. The supplier will be responsible for supporting and enhancing two key product areas: Cluster Toolkit and HyperCompute Cluster Service (HCS). This work involves a combination of ongoing operational tasks, testing, documentation, and specific development deliverables.

Scope of Work & Deliverables

The supplier will be responsible for the services and deliverables detailed below.

Ongoing Maintenance

• The contractor must provide ongoing maintenance and enhancements for all 6 projects covered under the original Statement of Work.

Cluster Toolkit

Cluster Toolkit is an open-source software solution that simplifies the deployment of high-performance computing (HPC), artificial intelligence (AI), and machine learning (ML) workloads on Google Cloud.

Ongoing Responsibilities:

• Stability Testing: Test the stability of new products, beginning with A3U. This includes:
• Building NVIDIA Collective Communications Library (NCCL) tests on a Slurm cluster.
• Setting up and running pairwise tests to identify and report bad nodes.
• Integration Test Triage: Perform rotational duties to manage and triage integration test failures. This includes:
• Monitoring daily failure chats and flake tools.
• Reporting on failures and performing advanced handling, such as creating new bug reports and categorizations.
• Documentation: Improve, organize, and maintain the Cluster Toolkit documentation. This process involves:
• Gathering existing documents and identifying information gaps.
• Creating new documentation and updating existing materials.
• Organizing the information in g3docs, consolidating it in a team Google Drive, and establishing a review process.
• Project Cleanup: Once a week, clean up the 'hpc-toolkit-dev' project by identifying and deleting unused resources.
• Security: Triage and address security alerts by checking for them, creating PageRanks (PRs) to resolve them, and applying the necessary updates.

Key Deliverables:

• HPC VM Image Releases: Deliver 4-6 High-Performance Computing Virtual Machine (HPC VM) image releases during 2025.
• Software Widget Releases: Release new software widgets every two weeks during 2025, including managing any necessary hotfixes.

HyperCompute Cluster Service (HCS)

HCS is a service that enables the deployment and management of resilient, high-performance AI and HPC systems at scale.

Key Deliverables:

• API Integration Testing: Add comprehensive integration tests for all HCS Application Programming Interface (API) surfaces. Coverage must include:
• HypercomputeClusters: Create, Delete, Update, Get, and List requests and responses.
• Network: NetworkInitialize params.
• Storage: StorageInitialize, FileStoreInitialize, Filestore tier, ParallelstoreInitialize, and GcsInitialize params.
• Compute: Resource request, Guest accelerator, Disk, Provisioning model, Reservation affinity and type, Orchestrator, Slurm, Node test, Storage configuration, and Slurm partition.
• Critical User Journey (CUJ) Validation: Add integration tests to validate the following critical user journeys:
• Creating a cluster that consumes a reservation.
• Creating a cluster with a new network and new storage.
• Creating a cluster using a pre-existing network and storage created both outside of HCS and by a previous HCS deployment.
• Destroying all components of an HCS-created cluster.
• Destroying a cluster while leaving the network and storage intact.

Updating a Slurm cluster to add a new reservation to both new and existing partitions