Instant KAI Sandboxes with vCluster
summary
Claim: vCluster runs a full Kubernetes control plane inside a pod and lets teams install independent Kai Scheduler instances for isolated scheduling. Demo shows fractional GPU allocation (examples: 20% and 10% slots scheduled with Kai), virtual cluster pods around ~300–400 MB, and create/delete cycles that finish in ≈39 seconds. Setup runs on a local kind cluster with an Nvidia GPU and demonstrates GPU sharing, independent scheduler versions, and clean rollback when experiments fail.
AI-written summary · about
chapters (14)
- 0:00Introduction and Overview
- 0:14Introduction to Kai Scheduler
- 0:38Setting Up the Environment
- 0:56Understanding Kubernetes Scheduling
- 1:37Features of Kai Scheduler
- 2:23Live Demo: Verifying GPU Access
- 2:54Introduction to V Cluster
- 4:49Technical Architecture of V Cluster
- 7:05GPU Workloads Beyond AI
- 8:36Deploying a Simple Pod
- 10:22Upgrading Schedulers in Production
- 15:07Configuring V Cluster for Kai Scheduler
- 22:53Creating and Testing Virtual Clusters
- 26:21Conclusion and Resources
related talks
37mVirtual Clusters in the Real World: Lessons from k3k’s Journey to Production
35mMulti-Tenancy on Kubernetes with Virtual Clusters
10mLightning Talk: Mind the Topology: Smarter Scheduling for AI Workloads on Kubernetes
35mRunning Isolated VirtualClusters With Kata & Cluster API
01h 47mKubernetes Virtual Clusters
23mVirtual Kubernetes Clusters: A New Approach to Multi-tenancy
26mVirtual Kubernetes Cluster
28mGPU sharing done right: Secrets, security, and scaling with Vault and vCluster