Enhancing the Kubernetes Scheduler for Diverse Workloads in Large Clusters
Talkat KubeCon + CloudNativeCon North America 2020 Virtual · on Thursday, 2020-11-19· at 15:45· Intrado Virtual Event Platform,·
summary
Apple extends the Kubernetes scheduling framework with concrete plugins and measurements: a static IP prefilter+filter plugin that checks node IP maps and reduces webhook overhead, a coscheduling (gang) plugin implementing sort/reserve/permit with timeouts, and a per-profile percent-of-nodes-to-score tuning knob. Benchmarks show webhook extenders consume ~50% of scheduling time versus ~4% for the plugin, and scoring fewer nodes can yield up to 3x throughput differences on affinity-heavy workloads in clusters of thousands to ten‑thousands of nodes.
AI-written summary · about
chapters (16)
- 0:00Intro
- 0:58Vanilla Scheduler is Becoming Insufficient
- 2:45Emerging Scheduling Requirements
- 3:00Scheduling Framework
- 5:20Comparison of Different Methods Extending the Vanilla scheduler
- 9:15StaticIP Scheduler Plugin Custom PreFilter and Filter plugins
- 11:15Compared with Scheduler Webhook Extenders
- 12:42Case Study Gang scheduling for batch jobs
- 13:57Lightweight Coscheduling Plugins
- 20:04Case Study Scalable scheduling
- 21:44Custom Scheduler Parameters percentage OfNodesToScore
- 24:02Group Scoring
- 25:05Scalable Scheduling Challenges
- 26:01Custom Scheduling with Multiple Profiles and Plugins An example policy file
- 27:37Summary
- 28:37Acknowledgements
related talks
36mNetwork-aware Scheduling in Kubernetes
42mDeep Dive Into the Latest Kubernetes Scheduler Features
30mDeep Dive: Scheduling SIG
01h 29mTutorial: Unleash the Full Potential Of Kubernetes Scheduler: Configuration, Extension And Operation In Production
32mA Tale of Two Plugins: Safely Extending the Kubernetes Scheduler with WebAssembly
32mCoordinate Workloads Colocation: QoS-Oriented Scheduling Enhancement on K8s
34mLatest Kubernetes Scalability Improvements
25mScale your Batch / Big Data / AI Workloads Beyond the Kubernetes Scheduler