Thousands of Virtual Kubelets: 1-to-1 Mapping a Supercomputer To Kubernetes With Supernetes
Talkat KubeCon + CloudNativeCon Europe 2025 · on Friday, 2025-04-04· at 13:45· Level 1 | Hall Entrance S10 | Room B·
summary
Supernetes performs a 1-to-1 mapping of Slurm nodes to v1/Nodes and Slurm tasks to v1/Pods, runs on LUMI, and tolerates strict HPC limits (no root, no namespaces, tight firewalls). Architecture uses an mTLS gRPC reverse tunnel from an agent on the login node to a controller that spawns thousands of virtual kubelets as Kubernetes nodes; FluxCD is patched for reconciliation (FluxCD patch). Demo shows ~103 virtual nodes and live sbatch↔pod sync working, with a 10s polling delay and unresolved networking, multi‑tenancy and HA issues.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:03Cloud vs HPC differences
- 3:10Kubernetes and Slurm similarities
- 6:24Bridging HPC and cloud
- 7:07Superetes architecture
- 10:18Live demo of Superetes
- 14:09HPC-to-cloud bridges
- 16:12HPC challenges
- 20:03Future HPC and cloud
- 23:36Conclusion
- 24:40Q&A
related talks
05mHow To Run Kubernetes Pods On My Slurm-Based HPC Center
30mBuilding a Global Supercomputer with Virtual Kubelet
31mVirtual What? Virtual Kubelet vs. Virtual Node
32mPods Everywhere! InterLink: A Virtual Kubelet Abstraction Streamlining HPC Resource Exploitation
28mKubeFlux: An HPC Scheduler Plugin for Kubernetes
36mIntro: Virtual Kubelet
36mIntroduction to Virtual Kubelet – Featuring Titus by Netflix
35mMulti-Tenancy on Kubernetes with Virtual Clusters
