Scaling To the Stars: Simulating Massive Clusters With KWOK
Lightningat KubeCon + CloudNativeCon Europe 2025 · on Thursday, 2025-04-03· at 17:10· Level 0 | ICC Auditorium·
summary
NVIDIA uses KWOK to simulate very large clusters on a laptop and validate rollout logic before touching production. A watchdog script for daemonsets initially fails at scale and causes OOMs on clusters of up to 1,250 nodes and ~18,000 pods; running the same tests under >2,000 nodes with KWOK reveals hotspots. Fixes include caching to cut requests by 40% fewer queries, batching list API calls and shrinking per-node memory, after which rollouts succeed at production scale.
AI-written summary · about
related talks
30mDeploy 1 thousand pods in my JUST laptop (w/ kwok)
31mLoad Testing Kubernetes: How to Optimize Your Cluster Resource Allocation in Production [I]
28mScaling in Kubernetes Safely on On-Prem KaaS Across 1,300+ Clusters and 40,000+ Nodes
37mOne Cluster to Rule Them All - Multi-Tenancy for 1k Users
38mScaling Kube-Apiserver to Thousands of Tenants – Lessons Learned From kcp
15mDemo: Resource(Cost) optimization on planet scale Kubernetes clusters.
30mScaling Kubernetes to Thousands of Nodes Across Multiple Clusters, Calmly
29mScale Your Batch / Big Data / AI Workloads Beyond the K8s...