CNCF-Hosted Co-Located Events Europe 2023

Deploying and managing Kubernetes clusters on High Performance Computing (HPC) hardware and infrastructure can be a challenge, to which diskless nodes and proprietary networks highly contribute. The Swiss National Supercomputing Centre (CSCS) is in the process of implementing a Kubernetes Infrastructure, spawning from standard resources to HPC. The goal is to deploy and manage multiple clusters centrally, regardless of the nature of the resources at disposal. Rancher has been appointed as the facilitator for such task. The integration with Harvester, a Cloud-Native Hyperconverged Infrastructure, enables the creation in different VLANs of clusters on-demand having custom node and cluster configurations. A GitOps approach, leveraging ArgoCD app-of-apps pattern, largely eases the deployment of applications on these clusters, resulting in several crucial aspects, e.g. re-usage of deployment code, enforcement of the declared state and self-healing. Currently, Slurm scheduler is used to manage workloads on HPC. But CSCS is looking for Kubernetes oriented complementary or alternative solutions. Ceph storage is widely used by the clusters, in all its flavors: Rados Block Device (RBD) and CephFS for persistent volumes, and S3 as a backup solution.