Tako Performance

Repeatable benchmarks for Tako, nginx, Caddy, HAProxy, and Envoy.

See RESULTS.md for the baseline, high-load follow-up, raw result links, findings, and profiling targets.

This repo contains:

• A small Go HTTP application used by every proxy candidate.
• Proxy configs for single-upstream runs, with load-balanced mode available for larger testbeds.
• Remote setup and restore scripts for benchmark hosts.
• A VM-local high-load harness with CPU/RAM/RPS/failure graph rendering.
• Raw benchmark output and Markdown reports.

The first baseline target was an Ubuntu 24.04 VM. The public report intentionally omits exact hostnames, IP addresses, and private network details.

Run

BENCH_VM=<ssh-host> ./scripts/sync-to-vm.sh
BENCH_VM=<ssh-host> BENCH_IP=<target-ip> CONCURRENCY_LIST="100 500" ./scripts/run-http-benchmarks.sh
BENCH_VM=<ssh-host> \
TAKO_SERVER_BIN=/opt/tako-performance/bin/<tako-server-release> \
SOURCE_IPS="127.0.0.2,127.0.0.3" \
PROXIES="nginx haproxy envoy tako caddy" \
CONCURRENCY_LIST="1000 2500 5000" \
REQUEST_TIMEOUT=60s \
COOLDOWN_SECONDS=10 \
./scripts/run-vm-local-http-benchmarks.sh
./scripts/render-metrics-graphs.sh results/<timestamp>/http-vm-local
BENCH_VM=<ssh-host> \
TAKO_SERVER_BIN=/opt/tako-performance/bin/<tako-server-release> \
COOLDOWN_SECONDS=10 \
./scripts/run-vm-local-tako-feature-benchmarks.sh

HTTP scripts default to MODES=single. Use MODES="single lb" only on a larger testbed where four upstream app processes have enough CPU capacity. HTTP scripts also default to REQUEST_TIMEOUT=60s so high-concurrency rows measure overload latency instead of a short client timeout.

HAProxy and Envoy cases require the matching system binaries on the benchmark VM. All proxy cases use the same local app, HTTPS route, certificate, and high-water request/connection protection where the proxy exposes a comparable knob.