Coordinate Workloads Colocation: QoS-Oriented Scheduling Enhancement on K8s
summary
Kubernetes add-on Koordinator implements QoS-aware colocation by adding sub-QoS classes, a colocation profile webhook, a code-manager recommender for resource profiling, and a node-side codelet that measures CPU scheduling and memory allocation latency. It reclaims idle requested-but-unused resources and uses an oversubscription + kill/reservation policy plus kernel-level knobs (CPU bursting, cgroups tweaks) to protect 99th-percentile latency apps while boosting batch utilization; authors report measurable utilization uplift after deployment and open-source code on GitHub.
AI-written summary · about
chapters (21)
- 0:00Intro
- 0:25Contents & Speaker
- 2:30Resource Efficiency
- 4:41Resource Model in K8s
- 6:30Why Noisy Neighbors
- 8:37Key Features
- 9:31Architecture
- 12:13Key Design - Koordinator Priority
- 14:17Key Design - Koordinator QoS
- 15:00QoS and Priority
- 16:22Spark on Kubernetes
- 17:35Spark on K8s with Koordinator
- 19:05Resource Overcommitment
- 21:54Load-aware Scheduling
- 23:35Fine-grained CPU Management
- 25:36koordlet CPU Suppress
- 27:47koordlet Satisfaction Eviction
- 29:05Various QoS Management
- 29:57Various Scheduling Features
- 31:08Koordinator Milestone
- 31:31Getting Involved
related talks
06m⚡ Lightning Talk: Colocating Data Workloads and Web Services on Kubernetes to Improve Resource Utilization
28mMultitenancy and Fairness at Scale with Kueue: A Case Study
40mLimits, Requests, QoS... Rethinking What You Thought You Knew About Scheduling in Kubernetes
20mA full-scenario colocation of workloads based on Kubernetes
45mLimits, Requests, QoS, PriorityClasses: Demystifying Scheduling in K8s!
37mCo-Location of CPU and GPU Workloads with High Resource Efficiency
11mK8s Resources Ensuring Performance and Optimization
36mCapacity Scheduling for Elastic Resource Sharing in Kubernetes