Network-aware Scheduling in Kubernetes
summary
Framework implements two controllers and three scheduler plugins plus a netperf feeder and uses Calico bandwidth CNI to treat bandwidth as an extended resource. It schedules by an AppGroup CR (explicit pod dependency and min bandwidth / max cost), a NetworkTopology CR (latency costs populated from netperf), and a TopologicalSort plugin to order pod placement. Demo shows ~30–40% lower request latency for Online Boutique and ~20% higher Redis throughput; scheduler overhead stays under one second at 10k-node scale.
AI-written summary · about
chapters (20)
- 0:00Introduction
- 1:54Background
- 4:11Framework
- 5:13Latency
- 6:16Framework overview
- 8:09Bandwidth
- 9:01Example
- 9:55Network topology cr
- 10:28Topological sorting
- 11:24Latency optimization
- 12:07Filtering and scoring
- 13:01Cluster topology
- 13:11Filtering function
- 14:02Scoring function
- 14:41Demo
- 15:34Deploying Controller
- 21:30Networkaware Performance
- 26:53Challenges Lessons Learned
- 28:51Thank You
- 29:20Questions
related talks
36mNetwork Bandwidth-Aware Kubernetes Cluster
36mNetwork Bandwidth-Aware Kubernetes Cluster
30mSched.net: A Network-Aware Kubernetes Scheduler [I]
30mEnhancing the Kubernetes Scheduler for Diverse Workloads in Large Clusters
01h 29mTutorial: Unleash the Full Potential Of Kubernetes Scheduler: Configuration, Extension And Operation In Production
47mNHD - A Topology-Aware Scheduler for K8s for Low-Latency & HPC Applications
18mLet There Be Topology-Awareness in Kube-Scheduler!
42mDeep Dive Into the Latest Kubernetes Scheduler Features