[nigrosimone]

A performance PR here comes with numbers measured by its author, on his machine, with his method, and the review can only take them or leave them. This adds a workflow that measures the PR itself, on the runner, against its base branch.

It builds both, starts servers from each build and alternates the load between them scenario by scenario, so both arms see the same machine in the same minute and only the ratio is read. Each arm runs twice, the second time from a copy of its build: base against base and head against head is the noise of that run, and a row is marked only when it moved more than that. Every server reports its own cpu time too, so the table shows how busy it was and what a request cost it. http_load_test from uSockets for the http rows and load_test for the ws rows, both built from the submodule, the server pinned to a core whose hyperthread stays idle.

Scenarios: hello world, headers in and out, the JSON POST of http_load_test through onData, a microcached route, ws echo at 20 B and 4 KB. Same design as brianc/node-postgres#3775.

The table is posted as a comment on the PR by a second workflow on workflow_run, so also on a PR from a fork, where the token of the run cannot comment. workflow_run uses the file on master, so the comment starts after the merge. About 15 minutes per run. Same code on both arms on a hosted runner (nigrosimone#4):

Scenario	Base req/s	Head req/s	Head / base	Rounds	Noise	Busy	CPU per request
http/hello-world	107869	108234	1.004x (+0.4%)	1.00 to 1.01	±1.0%	100% / 100%	9.26 / 9.23 us (0.996x)
http/headers	92553	92745	1.001x (+0.1%)	1.00 to 1.00	±0.7%	100% / 100%	10.79 / 10.77 us (0.998x)
http/json-post	84469	85234	1.009x (+0.9%)	1.00 to 1.01	±1.0%	100% / 100%	11.83 / 11.72 us (0.991x)
http/cached	117905	117662	0.999x (-0.1%)	0.99 to 1.00	±2.0%	100% / 100%	8.48 / 8.50 us (1.000x)
ws/echo-20b	111423	111237	0.997x (-0.3%)	0.99 to 1.00	±1.4%	100% / 100%	8.97 / 8.98 us (1.003x)
ws/echo-4kb	96290	98132	1.012x (+1.2%)	1.00 to 1.02	±2.6%	100% / 100%	10.37 / 10.18 us (0.988x)

Node v26.10.0, AMD EPYC 7763, 4 cores (server on cpu 0, load on 2,3), 100 connections.

see https://github.com/uNetworking/uWebSockets.js/actions/runs/36669803181/jobs/89233603871/summary_raw

[uNetworkingAB]

I would love it more if it used uSockets http_load_test instead of wrk

[nigrosimone]

Done: the http rows use http_load_test from uSockets now, built from the submodule without ssl. It always sends GET / (or its POST), so each scenario gets its own servers with the route on "/". The cached route runs no JS, there http_load_test counts the responses. The table in the description is a new same-code run.

[uNetworkingAB]

Okay either way but this is just "make examples" in uSockets folder otherwise