Tutorial: Build, Operate, and Use a Multi-Tenant AI Cluster Based Entirely on Open Source
Talkat KubeCon + CloudNativeCon Europe 2025 · on Wednesday, 2025-04-02· at 16:15· Level 1 | Hall Entrance N11·
summary
Open-source stack shows a concrete recipe: deploy Nvidia GPU Operator, Kueue for workload-aware queuing and quotas, scheduler plugins for gang and topology-aware placement, and app wrapper plus autopilot for automated health checks and fault recovery. Demos run on a 24‑GPU test cluster and IBM Research reports operating ~1,200 GPUs with near-100% utilization by letting teams borrow unused quota, preempting or suspending jobs, and automatically migrating workloads off failing GPUs without human intervention.
AI-written summary · about
chapters (13)
- 0:00Multi-tenant AI GPU clusters
- 2:25AI workloads and hardware needs
- 6:20Kubernetes and platform overview
- 10:32Cluster setup prerequisites
- 12:17Cluster config and team onboarding
- 18:18Queuing and quota with Kueue
- 20:03Kueue demo and preemption
- 32:14Fault detection with Autopilot
- 47:59Fault recovery with App Wrappers
- 55:01Kubeflow trainer example
- 57:32KubeRay fine-tuning example
- 1:06:28Batch inference example
- 1:09:59Key takeaways and results
related talks
21mExperience Operating Large GPU Clusters at Organizational Scale
39mScalable Multi-Node AI Workloads in Multi-Tenant AI Clouds U...
24mCluster Management for Large Scale AI and GPUs: Challenges and Opportunities
39mScalable Multi-Node AI Workloads in Multi-Tenant AI Clouds U...
36mProduction GPU Cluster with K8s for AI and DL Workloads
38mKubernetes for AI Workloads – What Works and What… Doesn’t
28mShared Yet Isolated at Scale: Building Multi-Tenant Inference Platform on…
05mSponsored Keynote: Towards Clouds of AI Clusters
