Thomas Camlong
c73b139dbd
PHASE2-FINDINGS.md had grown past 1,200 lines mixing server memory, client rendering, methodology and a long list of rejected experiments, which made it unusable as a reference. Split under docs/perf/: README.md headline results and the change-by-change ratio table SERVER-MEMORY.md what the container's RAM is made of, layer by layer CLIENT-PERFORMANCE.md dashboard cost in Chrome: DOM, fibers, images, paint REJECTED.md ideas implemented, measured and abandoned, with numbers METHODOLOGY.md the tools, and the measurement traps that gave wrong answers PHASE2-FINDINGS.md is now an index so existing links keep working. METHODOLOGY.md is the part worth keeping: every wrong number in this investigation came from an instrument rather than the application, and each trap is written down - snapshots cannot explain peaks, V8 truncates snapshot strings to ~1 KiB, snapshot objects are not live handles, measuring without a forced GC measures V8's growth, taking a CDP snapshot added 134 MiB to the process being measured, per-component React attribution silently returns nothing on a production build, a fixed port can measure a leftover container, and a server-side change that moves a client-side metric means the workload changed.
Homarr performance investigation
Measured work on Homarr's memory and rendering cost, split by subject. Every figure here comes from
a run of scripts/benchmarks/stress-restore.mts against a restored real backup — 4 boards, 51
widgets, 11 integrations, driven through an actual login — not from a synthetic or empty container.
| document | subject |
|---|---|
| SERVER-MEMORY.md | What the container's RAM is made of, and the changes that reduced it |
| CLIENT-PERFORMANCE.md | Dashboard rendering cost in Chrome: DOM, fibers, images, paint |
| REJECTED.md | Ideas that were implemented and measured, then abandoned — with the numbers |
| METHODOLOGY.md | The tools, and the measurement traps that produced wrong answers |
Headline result
Server RAM, merge-base 06724cfeb vs this branch, two alternated runs per side, reported as cgroup
anonymous memory because that is what counts against a container limit and cannot be reclaimed:
| stage | before | after | change |
|---|---|---|---|
| peak | 376.6 MiB | 279.3 MiB | −26% |
| boot idle | 130.5 MiB | 62.4 MiB | −52% |
| after sign-in | 247.4 MiB | 142.9 MiB | −42% |
| dashboard loaded | 293.4 MiB | 170.4 MiB | −42% |
| 4 tabs, 2 boards | 366.5 MiB | 258.2 MiB | −30% |
| after 2 min idle soak | 323.7 MiB | 200.8 MiB | −38% |
No before/after range overlaps at any stage. The cost is +18% CPU (solid, non-overlapping).
The changes that did it, by ratio
| change | size | effect |
|---|---|---|
--max-semi-space-size=4 |
1 line | −19% idle, −21% peak |
preloadEntriesOnStart: false |
1 line | −52% anon at idle, 310 fewer files even at full load |
| Umami: stop fetching 100k raw events per refresh | 1 file | −26% peak, −93% transient buffers |
serverExternalPackages (mysql2/pg + 8 more) |
1 block | −21 MiB idle |
| ssh2's 16 MiB WASM heap (docker-modem patch) | 1 patch | −15 to −18 MiB external |
content-visibility on media-release cards |
4 lines | −59% image requests, −41% layout objects |
What is left, and why it is hard
One lever has real headroom and it pays twice. Roughly 41 MiB of retained bundle source and ~38 MiB of V8 compiler zone memory both scale with how much JavaScript the server loads and compiles, and the largest single contributor is a 16.10 MiB subtree reached by every API route — the tRPC router importing every integration eagerly. Reducing that reduces both.
Everything else has been measured and closed: see REJECTED.md.