The National Research Platform: Stretched, Multi-Tenant, Scientific Kubernetes Cluster

This presentation introduces the National Research Platform (NRP), a collaborative, multi-institutional initiative designed to support scientific computing through a federated, Kubernetes-based infrastructure. The NRP spans over 84 physical sites contributed by more than 66 institutions, aggregating a vast pool of resources including approximately 1,500 GPUs and 29,000 CPUs. It supports over 500 research groups annually, offering multiple entry points such as Kubernetes, JupyterHub, and hosted services.

The talk explores the architectural and operational strategies behind the NRP, including its “stretched” Kubernetes cluster model that leverages VXLAN for virtual networking across administrative domains. Key technical challenges such as site-specific policies, networking (e.g., Science DMZ and 9000 MTU requirements), and storage latency are discussed, along with solutions like regionalized Ceph pools and automated Ansible-based deployment.

The presentation also highlights advanced service offerings like hosted JupyterHubs, VS Code environments, open-model LLM access, and integrated security with Falco for runtime threat detection. Looking forward, the NRP aims to expand access to more educational institutions and implement fair-sharing scheduling using Yunikorn to guarantee resource access for contributing members.