Workshop: Sovereign AI Supercomputing Cloud Workshop
Authors: Travis Cotton (Los Alamos National Laboratory (LANL))
Abstract: AI workloads at HPC sites increasingly expect Kubernetes, while traditional workloads still run under batch schedulers such as Slurm. Both layers are now shaped by vendors, so sites have good reason to keep control of the layer beneath them: how nodes are discovered, booted, configured and reassigned. If that layer is open and site-controlled, a site can change what runs on top without rebuilding its infrastructure. This talk describes our use of OpenCHAMI, a community-governed, open-source system management platform, as that layer. Building on earlier work presented at SC25, where nodes are moved between Slurm and Kubernetes using kexec-based fast booting, we combined OpenCHAMI's discovery, boot and metadata services with a small set of custom OpenShift artifacts. This lets us add bare-metal nodes to an OpenShift cluster as workers on demand. The immediate goal is to shorten the time from idle hardware to a serving model, often called "time to first token." We also want the provisioning path to be independent of any single workload manager or Kubernetes distribution. We will share the architecture, where the integration was simple and where it depended on distribution-specific pieces, and what that tells us about portability. We will close with ongoing work: extending the approach to other Kubernetes management stacks, and the experimental OpenCHAMI Operator, which runs the OpenCHAMI control plane itself as a Kubernetes-native resource.
Back to Sovereign AI Supercomputing Cloud Workshop Archive Listing
Back to Full Workshop Archive Listing