The Mac scaling run
Robert’s decision of 15 Sep 2026, 19:30: the suite and step 30’s measurements on the M3 Max (14 cores, 10 performance and 4 efficiency, 96 GB) at MO_CORES 1, 4, 10, and 14, in parallel with the rounds. The question is scaling on all fourteen cores against one on the same machine, and where the curve stops; it is not a comparison with the VM. The commands are one script; the table goes beside step 30’s in interpreter-step-30’s Result.
Before
The Mac has not run the suite since 14 Sep morning, and neither interpreter-step-29b’s MAP_NORESERVE reservation nor step 30’s threads have run there. So the suite comes first, and a failure there is the first result.
The commands
From the repo root on the Mac, on main after 16 Sep, with Zig 0.16 on PATH:
git pull
mo-wiki/plans/interpreter-step-30-suite/mac-run.sh
It builds, runs zig build test (silent on success, minutes cold), builds the queue and the ledger as native binaries, then per core count: mo-bench best of five, the queue’s 100k creates and lease-and-ack pairs at 1 and 32 workers five times (control-run-7-suite/measure.py, the same client as the rounds), and the ledger’s 100k transfers from 32 clients five times (ledger_load.py, step 30’s driver with its memory read made to work on a Mac). Results land in ../step30-mac/, one text file per run and driver.log with one line per result; driver.log is what to paste back. Fewer core counts: mac-run.sh 1 14. About an hour and a half in all.
What to read
| row | the VM, 1 core / 4 cores | what the Mac answers |
|---|---|---|
| queue pairs a second, 32 workers | 342 / 423 (fsync-bound) | whether the Mac’s disk moves it, and whether cores do |
| ledger transfers a second | 1,990 / 1,897 | whether 10 and 14 lift it or the fsync pool is the ceiling |
bench http-1k, kv-10k-get, replay-1m | in step 30’s table | the curve from 1 to 14, and the efficiency cores’ share |
Result (15 Sep 2026, 22:46 to 23:20 local)
Superseded by interpreter-step-34 (16 Sep): placement with the starter and the crossing made cheap, 100,000 cross-scheduler asks from 4 s to 0.14, the queue’s pairs level across cores, MO_CORES=1 no longer needed on the Mac. Run by Fable alone on the disk, driver.log in ../step30-mac/ (not in git; the table is in interpreter-step-30’s Result). The suite green in two minutes warm. Every row is fastest at one core: the queue 2,908 pairs a second at 32 workers at 1 core against 2,416 at 14, creates 7,868 against 5,476, the ledger 5,106 transfers a second against 4,073, echo-1k 19 ms against 104, kv-10k-get 383 ms against 1,236, replay-1m unchanged. The Mac’s disk moves the fsync bound eight times up; cores move nothing up and the cross-scheduler ask moves every crossing row down. Decision row: MO_CORES=1 for the rounds’ Mac rows until placement is fixed.