That’s Just My Cup of Tea: Configuring Cilium for Performance and Scale
summary
Cilium defaults favor compatibility but can be tuned for big gains: enabling Cilium EndpointSlices (stable since 1.16+) moves agents to watch slices and yields 87% API-server latency reduction and ~60% faster pod network SLA in 1K-node/20K-pod tests; enabling eBPF host routing or netkit devices removes veth overhead and brings container throughput/latency close to host networking; netkit requires a recent kernel (netkit in kernel 6.8) and some features need NIC/kernel support; use cilium connectivity-perf CLI and Hubble metric filters to measure effects.
AI-written summary · about
chapters (14)
- 0:00Cilium configuration intro
- 1:48AI workloads driving scale
- 3:56Control plane pressure
- 5:42Cilium endpoint slices
- 7:51Network policy security identities
- 9:36Hubble metrics filtering
- 10:59Data plane performance
- 13:48eBPF host routing
- 16:15Netkit devices overhead
- 18:25Performance features disabled
- 19:51Testing connectivity perf
- 21:55Demo: Cilium connectivity perf
- 23:40Scalability improvements
- 24:43Closing remarks
related talks
20mTackling Container Networking at Scale
16mCilium Wars: Scaling Up Your Network with Jedi-like Cilium Handling Techniques - Yash Tandon
24mRewiring Cilium: Operator-Driven Scale and Security With CiliumEndpointSlice
25mMore Churn No Problem: Lessons Learned Running Cilium in Production
31mUnleashing the Tides of Kubernetes Networking by Removing kube-proxy
26mSupercharge Kubernetes Networking with Cilium and IPv6
28m100Gbit/S Clusters With Cilium: Building Tomorrows Networking
36mCilium: Welcome, Vision and Updates