Comprehensive Performance Benchmark on Various Well-Known CNI Plugins
summary
Eight Kubernetes clusters are benchmarked—one per CNI—using sock perf, netperf and a gang‑style load tool from Panel Networks to measure TCP/UDP throughput, latency, CPU cost (UDP RR), pod network‑launch time and network‑policy impact. Flannel, Weave (fast datapath) and Romana hit throughput closest to a host‑network baseline; Calico and several others suffer ~15–20% overlay throughput loss and ~25% added latency on average. CPU overhead ranges ~10–40% (Weave ~2x, Cilium ~3x in these defaults). Most CNIs add ~0.5–1s launch time (Flannel >10s here). Network policies add negligible response delay up to 200 rules.
AI-written summary · about
chapters (21)
- 0:00Intro
- 0:45Agenda
- 1:17Kubernetes Network Model
- 2:01What CNI do?
- 2:36CNI Plugins
- 3:23Flannel
- 4:53AWS CNI
- 5:31Romana
- 6:09Goals
- 7:03Environment
- 8:15Tools
- 11:57TCP latency
- 13:55TCP throughput vs packet sizes
- 16:04CPU Overhead
- 19:14Network launch time
- 22:50Network Policy Calico
- 23:12Network Policy (Weave)
- 23:23Network Policy Cilium
- 25:19Takeaways
- 29:13References
- 30:39Steps
related talks
35mHow Container Networking Affects Database Performance
30mCloud Native NFV Performance Benchmarks
31mA Practical, Research‑Backed Introduction To Cloud‑Native Kernel‑Bypass Networking
34mTroubleshooting Low Latency Application on CNF Deployment
34mUnderstanding the Kubernetes Networking Layer
22mEmpowering Edge Software Stack on Arm with High Performance Container Networking Tech
26mTuning Your Cloud: Improve Global Network Performance for Application
45mGoldie of Gold and the Three CNIs