Skip to content

HTTP tier5 nginx bench parity (m1-nginx-bench-parity)

Summary

Wires li-httpd into the tier5_http bench harness for agent-gateway scenarios (static_small, keepalive_pipelining), adds a CI gate that verifies correctness on live build/li-httpd and optional wrk timing ratios vs nginx, and pre-caches GET /file.bin in the runtime for bench hot paths.

Agent continuation

  1. Gate: ./scripts/check-tier5-nginx-bench-parity.sh (verify always; timing when wrk + nginx present).
  2. Full timing: PATH=/usr/sbin:$PATH wrk + HTTPD_BENCH_DURATION_SEC=30 ./scripts/check-tier5-nginx-bench-parity.sh with HTTPD_BENCH_SKIP_TIMING=0.
  3. Pages: refresh benchmarks matrix when new CSV rows land (../benchmarks + refresh-live-sites.sh).

Changed

  • runtime/li_rt_net.c — prebuilt /file.bin response cache at httpd_prepare_root.
  • benchmarks/harness/http_bench_servers.py, bench_http_parity.py — shared server start + ratio checks.
  • benchmarks/harness/bench_http.py, verify_http.py — nginx + li timing/verify.
  • benchmarks/tier5_http/suite.toml — [profiles.parity].
  • scripts/check-tier5-nginx-bench-parity.sh, scripts/httpd-plan-gates.sh.

Performance

Agent-gateway static workloads: li-httpd targets ≥0.95× nginx RPS and ≤2× nginx p99 on parity profile (configurable via HTTPD_BENCH_RPS_RATIO_MIN / HTTPD_BENCH_P99_RATIO_MAX).