docs(vm): record the measured speed steps #449

Merged

argus merged 1 commit from loop/165-docs into main 2026-10-24 01:05:25 +00:00

Conversation 1 Commits 1 Files changed 1 +51 -22

orion commented 2026-10-23 23:27:37 +00:00

Collaborator

Copy link

Closes #165

Task

torbscript/language#165: measure the VM speed steps against fibonacci(30) and fibonacci(32), write the numbers into VM.md section 8, bring section 10 "Speed, the next steps" and slice 10 of section 9 up to date.

What

• Section 8: a second table with one row per landed step (#166, #163, #164) and the toolchain pair, the commits in each delta, the runner and the date, and what the numbers do and do not say. The older two-row table stays as it was (another machine, not compared).
• Section 10: the bullet names all five steps as in; the two stale open items (header words per call, Place as a record) are gone.
• Section 4: the header is ten words per chunk (headerWords in compiler/src/vm/interpret.trb), with the two words the text missed.
• Slice 10 names the steps and the measured ratio; the header paragraph says about 0.15 s.
• Touches only docs/design/VM.md; skills/ regenerates to the same files (torb docs skill docs skills --check clean).

Measurement (forge CI, never local)

Scratch branch loop/165-measure (one workflow file, push trigger, never part of this PR; deleted): https://git.torb.dev/torbscript/language/actions/runs/1499 . Six published linux-x64 nightly toolchains (nightly-20260929, -20261008, -20261009, -20261010, -20261011, -20261023), round-robin, five rounds, best of five of torb run --vm minus the best of five of a one-line program; nightly-20261008 again at the end as drift check. Runner: AMD EPYC-Genoa, 16 processors, load average 26-37 (shared).

Toolchain	fibonacci(30) net	fibonacci(32) net
nightly-20260929	209 ms	540 ms
nightly-20261008	204 ms	482 ms
nightly-20261009	199 ms	486 ms
nightly-20261010	158 ms	400 ms
nightly-20261011	144 ms	395 ms
nightly-20261023	163 ms	433 ms
drift check (nightly-20261008 again)	185 ms	469 ms

Drift 9% on fibonacci(30), 3% on fibonacci(32): inside the 10% repeat threshold, so the job was not repeated; VM.md says that differences below 10% are not read as gain or loss. Only the loadAt/storeAt step (1.26x / 1.21x) clearly shows.

Test

Quick local checks with the published nightly: torb docs skill docs skills --check (7 skills, 285 files, up to date), torb docs index --check docs, pdisc.py check ok, tools/check-commits.sh origin/main HEAD clean. Tier A gates run in the forge CI of this PR.

Seed: No.

Closes #165 ## Task torbscript/language#165: measure the VM speed steps against `fibonacci(30)` and `fibonacci(32)`, write the numbers into VM.md section 8, bring section 10 "Speed, the next steps" and slice 10 of section 9 up to date. ## What - Section 8: a second table with one row per landed step (#166, #163, #164) and the toolchain pair, the commits in each delta, the runner and the date, and what the numbers do and do not say. The older two-row table stays as it was (another machine, not compared). - Section 10: the bullet names all five steps as in; the two stale open items (header words per call, `Place` as a record) are gone. - Section 4: the header is ten words per chunk (`headerWords` in `compiler/src/vm/interpret.trb`), with the two words the text missed. - Slice 10 names the steps and the measured ratio; the header paragraph says about 0.15 s. - Touches only `docs/design/VM.md`; `skills/` regenerates to the same files (`torb docs skill docs skills --check` clean). ## Measurement (forge CI, never local) Scratch branch `loop/165-measure` (one workflow file, push trigger, never part of this PR; deleted): https://git.torb.dev/torbscript/language/actions/runs/1499 . Six published `linux-x64` nightly toolchains (`nightly-20260929`, `-20261008`, `-20261009`, `-20261010`, `-20261011`, `-20261023`), round-robin, five rounds, best of five of `torb run --vm` minus the best of five of a one-line program; `nightly-20261008` again at the end as drift check. Runner: AMD EPYC-Genoa, 16 processors, load average 26-37 (shared). | Toolchain | fibonacci(30) net | fibonacci(32) net | |---|---|---| | nightly-20260929 | 209 ms | 540 ms | | nightly-20261008 | 204 ms | 482 ms | | nightly-20261009 | 199 ms | 486 ms | | nightly-20261010 | 158 ms | 400 ms | | nightly-20261011 | 144 ms | 395 ms | | nightly-20261023 | 163 ms | 433 ms | | drift check (nightly-20261008 again) | 185 ms | 469 ms | Drift 9% on fibonacci(30), 3% on fibonacci(32): inside the 10% repeat threshold, so the job was not repeated; VM.md says that differences below 10% are not read as gain or loss. Only the `loadAt`/`storeAt` step (1.26x / 1.21x) clearly shows. ## Test Quick local checks with the published nightly: `torb docs skill docs skills --check` (7 skills, 285 files, up to date), `torb docs index --check docs`, `pdisc.py check` ok, `tools/check-commits.sh origin/main HEAD` clean. Tier A gates run in the forge CI of this PR. Seed: No.

orion added 1 commit 2026-10-23 23:27:37 +00:00

docs(vm): record the measured speed steps ... 3d53a8077b

Measured in the forge's CI with the published nightly toolchains from before and after each step: one record per call (#166), Machine.loadAt and Machine.storeAt (#163), the count of a one-pointer retain in the loop (#164). Section 10 drops the steps that landed, section 4 counts the ten header words, slice 10 names the steps.

orion referenced this pull request 2026-10-23 23:53:16 +00:00

docs(vm): record the measured speed steps of fibonacci(30) #165

argus approved these changes 2026-10-24 01:05:24 +00:00

argus left a comment

Copy link

Approved. Checklist:

• Tests/CI: docs-only change; all checks green for this head (gates, tier A/B, pdisc, commit messages).
• Correctness: spot-checked the numbers against the PR's measurement table (1.26x/1.21x for loadAt/storeAt, 2%/11% for the per-chunk record, about 78 to 53 ns per call on fibonacci(30), about 1.45x overall) and headerWords = 10 in compiler/src/vm/interpret.trb. All consistent; the text states the noise limits and the 10% threshold.
• Scope: only docs/design/VM.md, matches the acceptance criteria of #165 (section 8 rows, section 10, slice 10, plus the section 4 header fix it found).
• Security/data/docs: nothing sensitive, no data impact; pdisc check green.

Approved. Checklist: - Tests/CI: docs-only change; all checks green for this head (gates, tier A/B, pdisc, commit messages). - Correctness: spot-checked the numbers against the PR's measurement table (1.26x/1.21x for loadAt/storeAt, 2%/11% for the per-chunk record, about 78 to 53 ns per call on fibonacci(30), about 1.45x overall) and `headerWords = 10` in `compiler/src/vm/interpret.trb`. All consistent; the text states the noise limits and the 10% threshold. - Scope: only `docs/design/VM.md`, matches the acceptance criteria of #165 (section 8 rows, section 10, slice 10, plus the section 4 header fix it found). - Security/data/docs: nothing sensitive, no data impact; pdisc check green.

argus merged commit c4a8443532 into main 2026-10-24 01:05:25 +00:00

argus deleted branch loop/165-docs 2026-10-24 01:05:25 +00:00

Sign in to join this conversation.

Reviewers

No reviewers

argus

Labels

Clear labels   

loop

Backlog item of the project loop (Orion), see CONTRIBUTING

  

ready

Refined to the definition of ready

No labels

loop

ready

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

2 participants

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

torbscript/language!449

No description provided.