Workshops Archive

HPC Infrastructure Management at Scale


Workshop: HPC Systems Professionals Workshop (HPCSYSPROS26)

Authors: Jeremiah Hoyle (Accenture); Nishit Patel (Shell); Daniel Rix, Cory Kim, Carl Dukatz, and Chris Young (Accenture); Michael Gujral (Shell); Adam Hough (Microsoft Corporation); and Fokko Masselink, John Thiesfeld, and Ronald Cogswell (Shell)

Abstract: HPC infrastructure management at scale depends on more than golden images: operators must build, version, and deploy the right operating-system image to the right node role safely, repeatably, and without interrupting running jobs. The HPC OS project replaces a Satellite-driven provisioning model, where behavior depended on Web UI host parameters and Ansible overrides, with a central patching source for on-premises clusters. Layered RHEL, Rocky Linux, and Ubuntu CPU/GPU images are built with Packer and Ansible, stored as versioned artifacts, tracked with release metadata, and deployed through MAAS, orchestration, and post-deploy Ansible. In scale testing across 300 CPU/GPU nodes, the most stable operating point was 100 concurrent deployments, driven by AWS SSM and Shell-specific tradeoffs rather than MAAS; Satellite stabilized at about 60 nodes. We present lessons for replacing build-centered automation with a reproducible image build and deployment platform for large HPC environments.


Back to HPC Systems Professionals Workshop (HPCSYSPROS26) Archive Listing

Back to Full Workshop Archive Listing