Workshop: HPC Systems Professionals Workshop (HPCSYSPROS26)
Authors: Sadaf R. Alam (University of Bristol, Bristol Centre for Supercomputing); Alex Joseph Lovell-Troy (Los Alamos National Laboratory); and Mark Klein (Swiss National Supercomputing Centre)
Abstract: This experience report describes how reproducibility, security, and safe day-2 operations are handled in OpenCHAMI, an open, API-first, composable HPC control plane developed collaboratively under a Linux Foundation foundation and deployed in production across multiple sites. Modern centres run Slurm-scheduled HPC alongside Kubernetes-based AI, and maintenance windows test the management plane hardest. Rather than a controlled benchmarking study, we contribute a deployed open system, transferable operational practices, and an honest account of what is validated today versus roadmapped. We describe a composable microservices architecture with snapshot-based, predictably degrading state; a layered image-build pipeline with one-shot cloud-init convergence that removes configuration drift; and a WireGuard-based ephemeral bootstrapping model that keeps no long-lived credentials in images. We report author-reported outcomes, including a boot-to-scheduler-ready reduction from roughly eight minutes to ninety seconds, and evidence that the same control plane transfers across sites, architectures, and workload classes.
Back to HPC Systems Professionals Workshop (HPCSYSPROS26) Archive Listing
Back to Full Workshop Archive Listing