wrk2 open-loop bench results

Generated by scripts/wrk2_bench.sh on 2026-05-15 21:14:56Z. CPU: 12th Gen Intel(R) Core(TM) i9-12900HX — Kernel: Linux 6.19.6-arch1-1 x86_64

Methodology

wrk2 is an open-loop HTTP load driver: it issues requests at a fixed target rate regardless of whether previous responses have arrived, and corrects the latency histogram for Coordinated Omission. Closed-loop tools (wrk, ab, vegeta, our own bench.escript) hide the tail because the client backs off when the server stalls.

For each scenario, each server is driven at four rates: 50 %, 75 %, 90 %, and 95 % of its measured peak from bench_results.md. The 90 % point typically sits around the elbow; 95 % usually saturates. The achieved column is the actual rate wrk2 sustained; when achieved / target < 0.99 the row is tagged (saturated).

Each measurement runs wrk2 with both --latency (CO-corrected, "Recorded Latency") and --u_latency (uncorrected, "Service Time"). Two tables per scenario:

• CO-corrected — what every issued request would have seen. This is the honest tail.
• Uncorrected — what a closed-loop tool would have reported. Shown so the size of the coordinated-omission gap is visible directly. If the corrected and uncorrected p99 are close, the server is keeping up; if they diverge, that's a saturation signal even if achieved is at target.

Each row is the median of 3 runs at 30s each, -t8 -c50. Per-run logs live under /tmp/wrk2/<scenario>/<server>/<rate-pct>/run-N.log.

Scenario set is the h1-runnable subset of ?MAIN_SCENARIOS in scripts/bench.escript. pipelined_h1 and websocket_msg_throughput are excluded because wrk2's -c50 keep-alive shape can't reproduce pipelining or WebSocket frames; varied_paths_router is excluded because bench.escript's --standalone mode rejects multi-path workloads (no stable single path to expose to an external wrk2 driver); the h2-only scenarios are excluded because wrk2 is HTTP/1.1.

hello

CO-corrected (Recorded Latency) — honest tail.

Uncorrected (Service Time) — what closed-loop tools report. Compare to the corrected table for the size of the CO gap.

json

CO-corrected (Recorded Latency) — honest tail.

Uncorrected (Service Time) — what closed-loop tools report. Compare to the corrected table for the size of the CO gap.

echo

CO-corrected (Recorded Latency) — honest tail.

Uncorrected (Service Time) — what closed-loop tools report. Compare to the corrected table for the size of the CO gap.

headers_heavy

CO-corrected (Recorded Latency) — honest tail.

Uncorrected (Service Time) — what closed-loop tools report. Compare to the corrected table for the size of the CO gap.

large_response

CO-corrected (Recorded Latency) — honest tail.

Uncorrected (Service Time) — what closed-loop tools report. Compare to the corrected table for the size of the CO gap.

multi_request_body

CO-corrected (Recorded Latency) — honest tail.

Uncorrected (Service Time) — what closed-loop tools report. Compare to the corrected table for the size of the CO gap.

post_4kb_form

CO-corrected (Recorded Latency) — honest tail.

Uncorrected (Service Time) — what closed-loop tools report. Compare to the corrected table for the size of the CO gap.

large_post_streaming

CO-corrected (Recorded Latency) — honest tail.

Uncorrected (Service Time) — what closed-loop tools report. Compare to the corrected table for the size of the CO gap.

gzip_response

CO-corrected (Recorded Latency) — honest tail.

Uncorrected (Service Time) — what closed-loop tools report. Compare to the corrected table for the size of the CO gap.