Improving Network Efficiency with Topology Aware Routing
Talkat KubeCon + CloudNativeCon North America 2020 Virtual · on Thursday, 2020-11-19· at 16:50· Intrado Virtual Event Platform,·
summary
Kubernetes adds zone and region labels on EndpointSlices and updates kube-proxy to watch only local slices, using an algorithm that subsets endpoints by zone once a service has roughly three times the zone count (example: ~9 endpoints for 3 zones) and redistributes when overload exceeds about 25%. Simulations run against 39 million inputs with the kubernetes topology simulator score the approach 87 vs 73, keep 84% in-zone traffic, show ~1.7% average overload and cost ~37% more endpoint slices.
AI-written summary · about
chapters (22)
- 0:00Intro
- 0:15Outline
- 1:00Disclaimers
- 1:31How it all works
- 2:07iptables
- 3:14Key Issues
- 3:53Constraints
- 5:18TopologyKeys
- 6:02Examples
- 6:34Complexity
- 7:28Difficult to Implement
- 8:40The Ideal
- 9:33Goals
- 11:15Standard Topology Labels
- 12:13Simulating Algorithms
- 12:51The Algorithm
- 19:24Delivering EndpointSlices
- 19:47Summary
- 20:38Evaluation Criteria
- 21:57Simulation Results
- 24:18In the next few months
- 25:12Longer term
related talks
31mTopology-Aware Service Routing for Kubernetes is Coming Soon!
34mSpreading Apps, Controlling Traffic & Optimising Costs in Kubernetes
22mSpreading Apps, Controlling Traffic & Managing Costs in Kubernetes
43mScale Kubernetes Service Endpoints 100x
51mHow Kubernetes Networking Really Works: A Packet’s Journey Across Pods and Nodes
19mSIG-network: Updates and Future Directions
36mNetwork-aware Scheduling in Kubernetes
32mThe Easy--Don't Drive Yourself Crazy--Way to Kubernetes Networking