Benchmark in real Chromium too, and only call a row slower when the samples say so

npm run perf -- --browser times the same workloads in Chromium as a desktop
and as a phone (390x844 at 3x, CPU slowed 4x), one browser per version. Each
run collects garbage first and ends by forcing layout and the map canvas, so
deferred work counts. The jsdom DOM is about 30x slower than a browser's and
has no layout, so the interface rows need this.

Shared machines swing single runs by 30% on identical code, CPU time
included. A row now counts as slower or faster only when it moved 15% and a
Mann-Whitney U test gives p < 0.01 (15 runs a side by default). Identical
code: no flags in three full runs; a routing slowdown planted in a copy is
flagged in jsdom and on the phone.

The panel workload renders both selections every run instead of alternating,
which split its samples in two. CI drives the runner's Chrome through
playwright-core, no browser download.
This commit is contained in:
Frederic Jung 2026-10-26 12:04:57 +01:00
parent c88bad3eef
commit c76dd4d556
No known key found for this signature in database
11 changed files with 236 additions and 42 deletions

View file

@ -14,12 +14,28 @@ have huge late-game networks, and most of them are on phones.
```bash
git show origin/main:site/index.html > /tmp/base.html
npm run perf -- --base /tmp/base.html # about a minute
npm run perf -- --base /tmp/base.html # headless only, a few minutes
npm run perf -- --base /tmp/base.html --browser # plus real Chromium, desktop and phone (what CI runs)
```
`scripts/perf-bench.js` loads `tests/fixtures/big-network.json` (about 100 POPs, 150 links,
950 PNIs, 830 cache boxes, 110 IX ports, 170 customers) into the base and into this checkout, in
one process, and times them in alternating turns. It prints a Markdown table (median of 7 runs):
950 PNIs, 830 cache boxes, 110 IX ports, 170 customers) into the base and into this checkout and
times them in alternating turns, 15 runs each. It prints one table per environment:
- **Headless (jsdom):** our JavaScript. The most stable numbers for routing and simulation, but jsdom's
DOM is about 30× slower than a browser's and it has no layout, so its panel and dashboard rows say
little about what players feel.
- **Chromium, desktop** and **Chromium, phone** (`--browser`, `scripts/perf-browser.js`): the real DOM,
layout and canvas. The phone is 390×844 at 3× pixel density with the CPU slowed down 4×, about a
mid-range Android. Each version runs in its own browser. Use these rows for anything in the interface,
and to catch forced layouts (reading `scrollTop`, `offsetHeight` and friends right after changing the
DOM), which jsdom cannot see at all.
Browsers leave layout and canvas drawing for the next frame, so each timed run ends by forcing both;
without that, DOM and map work would look free. For the same reason, a profile that puts a lot of
time on a line reading `scrollTop` is usually showing layout that has to happen anyway, not waste:
check with the benchmark before "fixing" it.
Each table has these rows:
- Routing: `compute()`
- One game day: `dayTick()`
@ -28,7 +44,11 @@ one process, and times them in alternating turns. It prints a Markdown table (me
- NOC dashboard tabs
- Map at three zoom levels
and says whether both versions route the save's traffic identically.
and it says whether both versions route the save's traffic identically.
`--browser` needs Chromium: a local `playwright` with its browser works as is; otherwise
`playwright-core` (a dev dependency, installed by `npm ci`) drives the Chrome in `PERF_CHROME` or
`--chrome PATH`. CI uses the Chrome that ships with GitHub's Ubuntu runners.
Without `--base` it just times this checkout. `--runs N`, `--md FILE`, `--json FILE` and
`--summary FILE` are there for CI and for digging.
@ -57,15 +77,21 @@ incidents and failures).
## Read the table
- Below about 2 ms, differences are noise. The verdicts already ignore them.
- A busy machine can swing a single row by ±20% on identical code. When a verdict looks surprising
for what the change touches, rerun with `--runs 15` before reporting it, and report both runs.
- Shared machines (CI runners, cloud sandboxes) swing a single run by ±30% on identical code, and no
timing method gets around it: CPU time is as noisy as wall time there. So a row only gets a verdict
when it moved by 15% or more **and** the samples say it is real: p < 0.01 in a Mann–Whitney U test
(`scripts/perf-stats.js`), and more than 2 ms (8 ms on the phone). Otherwise it shows ≈, whatever
the percentage. The `p` column says how sure each row is.
- A big percentage with ≈ means "could be noise". If it matters for the change, run again (or with
`--runs 25`) and report both runs; do not report a single noisy number as a finding.
- ⚠️ slower (15% or more) or 🔴 slower (50% or more) needs a reason in the PR, or a fix. A new
feature may cost time; a refactor or a fix should not.
- "Routing results differ" is expected when a change is meant to change gameplay (new cities,
balance, routing rules), and a bug when it only promises to be faster.
- It is headless (jsdom with a recording canvas): it measures our JavaScript, not the browser's
painting. For drawing changes, also look at the game on a phone.
- Headless Chromium draws the canvas in software, real phones usually on the GPU, so map rows in the
Chromium tables overstate drawing cost. For drawing changes, also look at the game on a phone.
- Absolute numbers depend on the machine. Compare base and head from the same run, never numbers
from different runs or machines.
## Keeping the fixture useful