Paper summary
Network plugin performance at 10 and 100 GbE
Kubernetes network plugins trade throughput, latency and CPU cost differently across workloads and link speeds.
A Kubernetes Container Network Interface (CNI) plugin determines how traffic travels between pods. Packet processing adds CPU work, and the cost can change with application traffic and network speed.
Benchmark setup
The comparison covers Flannel, Calico, Cilium and Kube-router with Memcached, Nginx and Kafka on 10 and 100 GbE. It repeats the experiments five times and measures throughput, latency and CPU use. The CPU measurements distinguish user time, system time and software interrupt handling (softirqs).
Memcached at 100 GbE
On Memcached at 100 GbE, Kube-router delivered 23%, 46% and 51% higher network throughput than Flannel, Calico and Cilium, respectively. Its latency was 18%, 30% and 34% lower in the same order.
Underlay and overlay plugins
Kube-router avoids overlay encapsulation in the tested underlay setup but requires more routing management. Among the tested overlays, the paper finds Flannel favorable for CPU-intensive workloads and Cilium favorable for network-intensive workloads.