benchmark

Compares zoxy against HAProxy, nginx, Envoy and Pingora: every proxy gets the identical linearly-growing open-loop load ramp until it stops keeping up, and the output is one self-contained HTML report overlaying latency, CPU, memory and achieved req/s against offered load.

It runs unattended every night on a throwaway six-VM fleet with no public addresses, and the summary is posted to Discord and published to Pages.

             0 ──────── linear ramp ────────► MAX_RATE (100k)
                                                  ┌─► backend0 ─┐
        zrk ───────► proxy-under-test ────────────┼─► backend1  │  nginx origin
 (open loop, CO-     (pinned core, 4 GiB cap,     ├─► backend2  │  pool: canned
  corrected)          ONE at a time)              └─► backend3 ─┘  64B..100k
      │                       │                    round-robin     bodies, 2
      │                       │                                    cpus each
      │                       └── cAdvisor :8081 ── sampled at 1Hz
      │                            (cpu + working-set, into the run's
      └── per-1s NDJSON             own artifacts — no Prometheus)
        + HdrHistogram
             │
             └─► bench report ── report.json + report.html

Every proxy is an HTTP/1.1 reverse proxy doing the same job: parse each request, pick an origin from a four-node pool, forward it over a pooled keep-alive upstream, stream the response back. The pool is what makes the job the one a production proxy actually does — a single origin measures forwarding, and forwarding alone; balancing across a set is the other half, and it is the half where proxies differ.

Caddy and Traefik are not in the line-up. Both were measured in earlier revisions and both came in below Envoy on every profile they ran. They were removed to cut nightly run time and cost: the suite benches one proxy at a time on a fleet that exists only for the run, so each name costs a full ramp per profile in wall-clock and VM-hours, and neither was ever the line a reader was reading zoxy against. git log restores either one — details in docs/METHODOLOGY.md.

Docs

docs/METHODOLOGY.md	what is measured, the fairness rules that make the numbers comparable, and what travels with them
docs/RUNNING.md	profiles, triggering a run, the make targets, CI, and why make local is not a result
docs/SETUP.md	the one-time cloud setup CI cannot do for you
docs/LAYOUT.md	what lives where in this repo

Quick start

make build     # the bench binary (static musl)
make test      # unit tests
make smoke     # a ~90s end-to-end run of the whole suite, 2 proxies

The nightly runs itself; use workflow_dispatch to trigger one by hand. make local runs the whole suite on your own docker daemon — good for working on the harness, not a benchmark result. See docs/RUNNING.md.