Kubernetes as a container platform for the Edge
summary
Experimental results on Grid5000 evaluate a vanilla Kubernetes with 1 master + 99 workers under increasing WAN RTT from 1–400 ms using cluster-loader to create pods and namespaces. Pod startup splits into create|schedule|run phases: create stays local but schedule→run and run→watch slow with WAN RTT and multiple control-plane round trips. A surprising DNS pattern appears when CoreDNS replicas load-balance to the distant master causing ~100 ms responses instead of local 1–1.5 ms. KubeFed, KubeEdge (local SQLite cache) and Submariner are examined as partial fixes.
AI-written summary · about
chapters (11)
- 0:00Intro
- 0:24What we do?
- 4:09What we present today?
- 6:58Experimental Protocol: Monitoring VICTORIA
- 7:50Pod startup in K8s
- 9:04Pod startup results
- 12:49DNS results
- 14:41Lessons
- 16:04Kubefed: Cluster federation
- 17:58Submariner: Inter-cluster connectivity
- 18:44Takeaway
related talks
23mMigrating 5G Core to the Edge, Wisely and Confidently with Kubernetes
22mKubeEdge: From Fixed Location To Movable Edge, Latest Updates And Future
36mEdge Computing and Kubernetes BOF
01h 16mIntro + Deep Dive: Kubernetes IoT Edge WG
34mManaging Edge Nodes With “KubeEdge”
33mKubeEdge: From Fixed Location to Movable Edge, Latest Updates and...
34mKubernetes IoT and Edge Working Group
35mKubernetes, Lessons Learnt