Scale your Batch / Big Data / AI Workloads Beyond the Kubernetes Scheduler
summary
Benchmarks run on a 100-node simulated cluster show the default kube-scheduler fails to finish 1,000 jobs × 10 pods (only ~51% complete, massive rescheduling churn) and the co-scheduling plugin deadlocks high-parallel jobs (100 jobs × 100 pods yields ~81% failure). Using Kueue (CRDs + controller) with the suspend field and resource flavors yields 100% success and finishes near the theoretical 10–15 minute ideal, avoids overcommit, integrates without replacing kube-scheduler, and triggers autoscaling via CA provisioning requests.
AI-written summary · about
chapters (8)
- 0:00Scaling beyond Kubernetes Scheduler
- 1:49Online vs batch workloads
- 4:19Kubernetes Scheduler limits
- 8:28Benchmarking batch workloads
- 10:13Custom scheduler options
- 15:17Q and alternative solutions
- 18:32Why Q for OpenShift AI
- 20:31Q&A
related talks
29mScale Your Batch / Big Data / AI Workloads Beyond the K8s...
24mKubernetes Batch Processing at Scale - A Scheduling Perspective
30mDeep Dive: Scheduling SIG
25mKube-scheduler-evaluator: Evaluating Kubernetes Schedulers at Hyperscaler…
36mA Toolkit for Simulating Kubernetes Scheduling at Scale
01h 29mTutorial: Unleash the Full Potential Of Kubernetes Scheduler: Configuration, Extension And Operation In Production
29mData Processing Efficiency: Optimizing Batch Workloads on Kubernetes With Custom Schedulers
10mFair Scheduling for Deep Learning Workloads in Kubernetes