Skip to main content

Network Performance

Purpose: For platform engineers and network engineers, documents CNI performance characteristics, tuning options, and throughput benchmarks.

CNI Performance Comparison

CNIModePod-to-Pod ThroughputLatency (p99)NetworkPolicy Performance
CalicoVXLAN overlay8.5 Gbps0.15msGood (iptables)
CalicoDirect routing (BGP)9.8 Gbps0.08msGood (iptables)
CalicoeBPF dataplane9.9 Gbps0.06msExcellent (kernel bypass)
Kube-OVNGeneve overlay8.0 Gbps0.18msGood (OVS flows)
Kube-OVNUnderlay9.7 Gbps0.09msGood (OVS flows)

Tested on 25 Gbps NICs, MTU 9000, iperf3 TCP streams.

MTU Tuning

Network TypeRecommended MTUCalculation
Physical NIC9000 (jumbo)If switch supports; else 1500
VXLAN overlay8950Physical MTU − 50 (VXLAN header)
Geneve overlay8942Physical MTU − 58 (Geneve header)
Direct routing9000Same as physical

Set MTU in the cluster configuration YAML:

networking:
mtu: 8950

eBPF Acceleration

Calico eBPF dataplane provides:

  • Kernel-level packet processing (bypasses iptables)
  • Native host-routing without encapsulation overhead
  • Direct Server Return (DSR) for LoadBalancer services
  • ~30% reduction in CPU usage for networking at scale

Enable via Kustomize overlay:

calico:
bpfEnabled: true
bpfExternalServiceMode: DSR

Service Mesh Overhead

ConfigurationThroughput ImpactLatency Added (p99)
No meshBaseline
Istio sidecar (mTLS)−15%+2ms
Istio ambient (ztunnel)−5%+0.5ms

Recommendations

  • Use jumbo frames (MTU 9000) whenever switch infrastructure supports it
  • Prefer direct routing over overlay for bare-metal deployments
  • Enable eBPF dataplane for clusters above 50 nodes
  • Disable connection tracking for high-throughput stateless services