Working your Cluster: Smarter Scheduling Decisions for Your Work...
summary
An extender called Telemetry Aware Scheduler uses real node telemetry to filter and score nodes (not just CPU/memory), supporting multi‑metric rules and actions: don’t schedule, prioritize by metric (e.g., lowest temperature), label nodes for affinity and eviction. It integrates with Prometheus node exporter, Prometheus and the custom metrics adapter, can tag a faulty GPU like disable-card=gpu1, and in a demo with five pods it labels, evicts and reschedules pods successfully; scheduling relies on the metrics pipeline so outages can delay decisions.
AI-written summary · about
chapters (7)
- 0:00Start
- 3:01Why Resource Dates and Scheduling Matter
- 3:13Telemetry Aware Scheduling
- 5:18Telemetry Aware Scheduler
- 7:06Metrics Pipeline
- 8:05Telemetry Aware Policy
- 11:25Disk Schedule Strategy
related talks
32mBuilding a Kubernetes Scheduler using Custom Metrics
22mKeep HPC Running - an SRE's Guide to Supporting GPUs on Kubernetes
31mKubernetes Advanced Scheduling
30mScore-Driven Multi-Cluster Management: An Evaluation Framework for…
29mPower Efficiency Aware Kubernetes Scheduler
05mProject Lightning Talk: Scheduling AI Workload Among Multiple Clusters
35mSIG Scheduling Deep Dive
30mSched.net: A Network-Aware Kubernetes Scheduler [I]