CERN: A Cloud Native Scientific Computing Platform for NextGen AI

[mrbobbytables]

CERN’s "NextGen AI" Scientific Computing Platform is a purely cloud-native architecture built to support data acquisition and advanced machine learning workloads for the High-Luminosity Large Hadron Collider. Transitioning away from legacy HPC schedulers like SLURM, the platform relies entirely on Kubernetes and a suite of CNCF projects—such as ArgoCD, Bootc, Longhorn, Kyverno, and Kueue—to provide advanced scheduling, quota management, and heterogeneous hardware support (including CPUs, multi-vendor GPUs, and FPGAs) over a shared, on-premises resource pool. To bridge the gap between modern cloud-native environments and the traditional workflow needs of physicists, the platform integrates critical "glue" features like ContainerSSH for backward-compatible shell access and automated Kyverno mutating policies to dynamically configure volumes, identity, and hardware resources without requiring users to write complex YAML.

Link: https://architecture.cncf.io/architectures/cern-scientific-computing/