Kube-scheduler-evaluator: Evaluating Kubernetes Schedulers at Hyperscaler…
summary
An open-source framework defines large-scale workloads in Go, emulates a cluster as a single binary, and advances an internal virtual time so a 160-day Alibaba trace with 1,523 nodes, 6,000 GPUs, 8,000 tasks finishes in about 25 seconds. Metrics stream to VictoriaMetrics and Grafana, scenarios run against emulated controllers or external kwok/kind/real clusters, and the tool is designed to fit local development, CI, and pre-release benchmarking for scheduler outcome comparison.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 0:24Scheduler evaluation gap
- 4:43Existing tools
- 6:28KubeScheduler Evaluator features
- 11:21Single binary execution
- 13:10Virtual time for speed
- 14:18Programmable scenarios in Go
- 16:03Demo with Alibaba data
- 20:40How to try KubeScheduler Evaluator
- 21:21Integration into development
- 23:32Key takeaways
- 25:19Q&A
related talks
36mA Toolkit for Simulating Kubernetes Scheduling at Scale
25mScale your Batch / Big Data / AI Workloads Beyond the Kubernetes Scheduler
27mUnderstanding Kube-Scheduler-Simulator
29mScale Your Batch / Big Data / AI Workloads Beyond the K8s...
32mBuilding a Kubernetes Scheduler using Custom Metrics
29mSIG Scheduling Deep Dive
01h 29mTutorial: Unleash the Full Potential Of Kubernetes Scheduler: Configuration, Extension And Operation In Production
38mBeyond Kubernetes: Adapting To Spe...