Constructing Kubeflow Super Nodes with Karpenter
summary
Karpenter automatically provisions and consolidates nodes for Kubeflow notebooks and pipelines, launching instance types (spot or on‑demand, Graviton/AMD) to match pod constraints. Demo provisions a 4‑GPU instance for a notebook and uses GPU operator with MIG/time‑slicing to run four one‑GPU pods on a single physical GPU. Pipelines trigger nodes to spin up per task and then terminate them when finished; background consolidation shrinks extra‑large nodes to smaller types to cut cost (spot can save ~90%). Karpenter provisioner CRDs drive this behavior.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:27Challenges around machine learning on Kubeflow
- 3:02Challenges around cluster autoscaler
- 6:27Provisioner
- 8:54Kubenetes
- 9:37Consolidation
- 10:47Kubeflow
- 11:40Architecture
- 18:05Demo
- 25:12Wrapup
related talks
27mHarness the power of Karpenter to scale, optimize & upgrade K8s
22mUse Karpenter to scale, optimize & upgrade K8s
34mReduce Costs and Optimize on Your Cloud with Karpenter
18mKarpenter : Your trusted Companion for Kubernetes Scaling
22mKarpenter Efficient scaling of Kubernetes clusters
01h 01mJust-in-time Nodes for Any AWS EKS Cluster: Karpenter
19mJust in time Worker Nodes with Karpenter
36mGroupless Autoscaling with Karpenter