Performance.

Measured, not claimed.

All benchmarks are lies: these numbers measure one scenario, on one machine, under artificial conditions. They show relative differences between implementations, not absolute performance. If it matters to you, run ze-perf on your own hardware with your own workload.

BGP benchmark

ze-perf establishes two BGP sessions with a device under test: a sender injects 100,000 routes, a receiver times their arrival. The same harness runs unmodified against BIRD, FRR, GoBGP, RustyBGP, freeRtr, rustbgpd, and OpenBGPd, all in Docker on the same host.

Go carries an estimated 10-15% CPU overhead against C/Rust implementations. That number comes from code-path analysis, not a measured benchmark -- Ze has not been benchmarked at DFZ scale (1M+ prefixes).

Convergence 62ms to propagate 100,000 routes (2026-06-05 run)
Throughput 1,612,903 routes/sec sustained during propagation
Withdrawal 596ms from withdrawal sent to receiver idle
Full results All DUTs, all runs, full methodology
ze-perf 
# build ze-perf and all DUT images, then run
$ python3 test/perf/run.py --build --test

# test specific DUTs only
$ python3 test/perf/run.py --build --test ze bird

# regenerate docs/performance.md from results
$ bin/ze-perf report --doc test/perf/results/*.json
Prerequisites

Docker (Colima on macOS). ze-perf works against any BGP implementation, not just Ze -- point it at your own DUT.

Benchmarking guide architecture, flags, JSON output

Performance evidence in the labs.

Where else the numbers show up.

BGP Protocol Interop

Daemon Docker
  • Same DUTs as the benchmark: FRR, BIRD, GoBGP
  • Correctness first; the perf numbers above measure the same sessions

VPP Dataplane Evidence

Daemon Docker
  • Ze programs FIB, traffic, and firewall into a real VPP daemon
  • Dataplane throughput backs the numbers in the VPP guide

VLAN QoS Wire-Level Proof

Daemon AF_PACKET
  • 802.1p PCP tagging verified on the wire
  • Not just kernel-state acceptance -- captured packets