Beyond Kubernetes: Adapting To Spe...
summary
Panelists describe concrete work: open sourcing KAITO/Kai and Skyhook for runtime tuning, a proposed gang/coscheduling and topology‑aware scheduler plugin ecosystem, and Envy Sentinel for graded node health detection. They report fractional‑GPU support and caching/sharding patterns to avoid full dataset duplication, production customers running AI at thousands of GPUs (training clusters of ~2,000–3,000 nodes) and batch jobs that burst to tens of thousands of pods. Expect tooling around Kueue, Volcano/Q, DCGM exporter, and tighter GKE/AKS/EKS integrations.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 7:28Kubernetes for AI workloads
- 13:30Customer challenges with AI
- 17:22Balancing leadership and community
- 20:33Community pain points
- 27:15Beyond GPU workloads
- 30:05Edge GPUs for research
- 32:14Future of accelerators
- 37:39Closing remarks
related talks
16mAI + Kubernetes: What Beginners...
33mFrom spreadsheet scheduling to Kubernetes: building an on-premise ML platform
40mPanel: Kubernetes and GPU Trends for AI in Financial Services
09m⚡ Lightning Talk: Advanced GPU-Orchestrated Workflows and HPC Integrations on K8s for Distributed AI/ML at Scale
18mKeynote: Adapting Kubernetes for Machine Learning Workflows
10mLightning Talk: Mind the Topology: Smarter Scheduling for AI Workloads on Kubernetes
27mHarnessing Kubernetes for Scalable AI/ML Workloads
42mThe Challenges Managing a Kubernetes-Based Machine Learning Infrastructure