Trimaran: Real Load Aware Scheduling in Kubernetes
Talkat KubeCon + CloudNativeCon North America 2021 · on Wednesday, 2021-10-13· at 15:25· Room 408 AB + Online·
summary
Trimaran adds two scheduler scoring plugins, TargetLoadPacking and LoadVariationRiskBalancing, plus a cluster-wide LoadWatcher that pulls Prometheus, SignalFX or the Kubernetes metrics API and caches sliding-window averages and standard deviations. Tests show better packing (100-node/400-pod sim) and lower fragmentation, and scheduler-perf runs at scale (5,000 nodes / 1,000 pods and 500 nodes / 1,000 pods) keep latency within SLOs. Deployment is used at PayPal and runs as either a service or library; the plugins work but require metric fallbacks and tuning.
AI-written summary · about
chapters (18)
- 0:00Introduction
- 0:52Agenda
- 1:34Background
- 3:13Architecture
- 5:12Loadwatcher
- 7:38Target Load Packing
- 8:40Target Load Packing Algorithm
- 10:56Experiments
- 12:44Performance
- 14:38Strategy
- 15:54Load variation
- 16:46Example
- 19:31Design
- 20:55Demo
- 28:02Summary
- 29:36Licenses
- 30:51Future work
- 32:00References
related talks
42mDeep Dive Into the Latest Kubernetes Scheduler Features
34mWorking your Cluster: Smarter Scheduling Decisions for Your Work...
36mNetwork-aware Scheduling in Kubernetes
30mEnhancing the Kubernetes Scheduler for Diverse Workloads in Large Clusters
32mOptimizing Application Performance on Kubernetes
01h 29mTutorial: Unleash the Full Potential Of Kubernetes Scheduler: Configuration, Extension And Operation In Production
36mA Toolkit for Simulating Kubernetes Scheduling at Scale
35mLeveraging the Kubernetes Enqueue Scheduling Plugin for Smarter Workload Placement