docs(reprint): the ledger carries four runs and the cost is 4736 s - #323
Open
the-homeless-god wants to merge 3 commits into
Open
the-homeless-god wants to merge 3 commits into
the-homeless-god wants to merge 3 commits into
Conversation
Task 8570 said the print takes ten hours against a goal of two. The number came from the last RECORDED row of docs/reprint-ledger.tsv, 17 September 2026, 36839 s. Runs after it were many and not one was written down: the ledger is filled by hand and has no instrument of completeness (task 6704). Four missing runs are written in now. 2026-10-04, print, landed, 4736 s (1 h 18 min 56 s), on gpu. Start 18:24:13 UTC from the first line of /home/u/b-print-0725-print.log, end 19:43:09 UTC from the file timestamp (stat -c %y); the log was read, not touched. Output 7 files, 44661951 bytes, from the line that says so. Peak 50.6 GiB, the largest of 143 pulses of the print itself, not /usr/bin/time; in the judging stage the peak was 50.0 GiB. The judging chain took 2508501226319 steps of 4000000000000, 62 percent, and the print said that itself: it prints a warning when a stage passes half the cap. The copy had no git, so the fingerprint says the seed was printed from commit "no-git" - there is no hash for this run. 2026-10-04, print, killed, 4989 s (1 h 23 min 09 s), on dev. SIGTERM from the harness at 19 percent of the judging stage, peak 41.6 GiB. Start 16:53:13 UTC from the log header of /home/b/projects/flang-cluster/print-batch.log, end from the timestamp of its last line. 2026-10-04, print WITHOUT THE JUDGE, landed, 6890 s (114 min 50 s), 7 files, rebuild of the printed sources 1 min 17 s. A FOREIGN measurement: the number came in an agent report of the same day, no log was named, nothing here measured it. Peak, bytes and cap are zeroes with the reason spelled out. 2026-09-25, print, landed, 128700 s (35 h 45 min), from the header of scripts/bootstrap-reprint.sh and .github/workflows/reprint.yml. Peak and bytes were never recorded and the tree names no log for that run - the same case as the row of 30 August (task 6704). The ledger header now carries two warnings. Rows BEFORE 1 October were taken by a binary WITHOUT the kernel call memo (ADR-0051, task 5709) and are not comparable: 36839 s against 4736 s is 7.8 times, and that is the memo, not the tree. And steps per second is NOT a measure of speed - a memoized call is charged its steps without being executed, so between two neighbouring pulses of one and the same print the counter ran 85 million and then 2.4 billion steps per second. Cost is written as seconds of a landed run; a percentage of the cap is spend, not speed. Task 8570 is rewritten to that and renamed. The goal of two hours is met, point 3 of its acceptance is done, and what is left is named: the step cap, which the judging stage ate 62 percent of. It is raised from 4 to 8 trillion by commit 3f9b9a77b, which on 4 October lives in a branch and is not on dev. Points 1, 2 and 4 are marked not done or not measured, with the reason. The three places that convert runs to one machine through steps per second are named with line numbers and called an unfit measure; they are not edited here. scripts/seed/reprint-freshness.fscript told the human the cost was "from 10 h 14 min to 35 h 45 min, peak 44.9 GiB", quoting this ledger. It now says from 1 h 18 min to 35 h 45 min, peak 50.6 GiB. Checked: prose-numbers-guard --plan Check, 224 of 224 marks agree with the tree; task-numbers-guard --plan Check, no doubled numbers; flang check of reprint-freshness.fscript clean; the freshness plan still reads the ledger and still reports the last byte-for-byte check of 8 September, which my rows do not touch. flang/scripts/tasks.fscript --plan "task board whole" exhausts its own step limit on the untouched tree as well, so it proves nothing either way. NOT MEASURED: the reprint itself was not run from this worktree - a print is running on gpu and a second one on the same machine is not started beside it; and sh scripts/bootstrap-reprint.sh --bystro was not called for the same reason.
A stale number in the header of scripts/bootstrap-reprint.sh changed
someone's work on 4 October 2026: an executor read there that a landed run
costs between 6 h 23 min and 35 h 45 min and did NOT run the reprint. A
second one read "8 h 34 min" from docs/reprint-ledger.tsv. That is what a
stale number costs - not a blemish in prose but work cancelled.
The measured cost of the landed run of 4 October is 4736 s, 1 h 18 min 56 s
(gpu, /home/u/b-print-0725-print.log, 18:24:13 to 19:43:09, 7 files and
44661951 bytes out, peak 50.6 GiB over the pulses of the print itself).
The header is rewritten with ZERO growth of lines, because the growth gate
refuses added lines in shell: 14 lines replaced by 14, the file still 3592
lines. What the script now says:
the price of --check 8 h 34 min 02 s, 30 August -> 1 h 18 min 56 s
(4736 s), 4 October
the refusal of --bystro "the price is 8 h 34 min" -> "1 h 19 min"
the print directory "39 MB and half a day" -> "45 MB and an hour
and a bit", 44661951 bytes, peak 50.6 GiB
"the print takes six six hours -> an hour and a bit
hours", five places
Both shells parse it: bash -n and sh -n are clean.
NOT TOUCHED ON PURPOSE: the MEASURED_COST and MAX_STEPS block, where
"35 h 45 min" also stands. Commit 3f9b9a77b rewrites those very seven lines
to raise the step cap, and two edits of one line would collide in the merge.
Task 8570 says so, and says that the stale number stays there until that
commit lands.
Task 8570 also gains two things measured today. The second print of the day,
running under the raised cap of eight trillion, prints NO "step budget
running out" warning at all - the raise did exactly what it was for; its
seconds are not written into the ledger because the run is still going and a
price must not be written from a ceiling. And the three places that convert
runs to one machine through steps per second are named with a reason: the
conversion is an unfit measure, since a memoized call is charged its steps
without being executed.
An awk over docs/reprint-ledger.tsv for the last "print landed" row gave 6890 s - the run WITHOUT the judge, a foreign measurement whose time of day is not known. The refusal of scripts/bootstrap-reprint.sh --bystro quotes exactly that row as "the price of the last landed run", so it would have named the price of a print that never judged anything. The three rows of 4 October are ordered so the run without the judge stands first (no time of day to place it by) and the two runs with known starts keep their order: 16:53:13 UTC killed, 18:24:13 UTC landed. The last landed row is now 4736 s, the number the script header and the task name. Task 8570 says why the order is what it is.
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Task 8570 claimed the reprint takes ten hours against a goal of two. The claim
was false by a number, and the number was a gap in the ledger: 36839 s is the
last RECORDED row of
docs/reprint-ledger.tsv(17 September 2026), not thecost. Runs after it were many and none was written down.
Four rows written in
scripts/bootstrap-reprint.sh(35 h 45 min); peak and bytes never recorded, no log named in the tree (task 6704)/home/u/b-print-0725-print.log, read only: start 18:24:13 UTC from the first line, end 19:43:09 UTC fromstat -c %y/home/b/projects/flang-cluster/print-batch.log: SIGTERM from the harness at 19 % of the judging stageThe landed run of 4 October printed 7 files and 44661951 bytes, and its judging
chain ate 2508501226319 steps of 4000000000000 (62 %) - the print said that
itself. The copy had no git, so the fingerprint says commit
no-git; there isno hash for the run.
Two warnings added to the ledger header
by a binary WITHOUT the kernel call memo (ADR-0051, task 5709). 36839 s
against 4736 s is 7.8 times, and that is the memo, not the tree.
its steps without being executed, so between two neighbouring pulses of one
and the same print the counter ran 85 million and then 2.4 billion steps per
second. Cost is written as seconds of a landed run; a percentage of the cap
is spend, not speed.
Task 8570 rewritten and renamed
8570-the-print-is-outrun-because-it-is-slow.md->8570-the-print-runs-out-of-steps-not-of-hours.md. The goal of two hours ismet (point 3 of its own acceptance: a row under 7200 s). What is left is named:
the step cap, 62 % of which the judging stage ate, raised from 4 to 8 trillion
by
3f9b9a77b(on dev it is not yet). The instrument that measures the cost byitself is named too: the print's own warning at the half-cap threshold,
flang/src/emit/c/flang_repl.c:2324. The three places that convert runs to onemachine through steps per second are named with line numbers and called an
unfit measure; they are not edited here.
scripts/seed/reprint-freshness.fscriptquoted this ledger as "from 10 h 14 minto 35 h 45 min, peak 44.9 GiB" and now says from 1 h 18 min, peak 50.6 GiB.
The script header too, with zero growth of lines
A stale number there changed someone's work on 4 October: an executor read in the
header of
scripts/bootstrap-reprint.shthat a landed run costs between6 h 23 min and 35 h 45 min and did not run the reprint. The growth gate
refuses added lines in shell, so 14 lines are replaced by 14 and the file is
still 3592 lines (
bash -nandsh -nclean):--check--bystroNot touched on purpose: the
MEASURED_COST/MAX_STEPSblock, where"35 h 45 min" also stands. Commit
3f9b9a77brewrites those very seven lines toraise the cap, and two edits of one line would collide in the merge. Task 8570
says the stale number stays there until that commit lands.
The second print of 4 October, already under the raised cap of eight trillion,
prints no "step budget running out" warning at all - the raise did what it
was for. Its seconds are not in the ledger: the run is still going, and a price
must not be written from a ceiling.
Checked
All 21 pre-push guards green in 63 s (
task-ledger,prose-numbers-guard224/224,
no-growth,lint-growth,tree-inventoryamong them).flang checkof
reprint-freshness.fscriptclean; the freshness plan still reads the ledgerand still reports the last byte-for-byte check of 8 September, which these rows
do not touch.
NOT MEASURED: the reprint itself was not run from this worktree - a print is
running on gpu and a second one is not started beside it; and
sh scripts/bootstrap-reprint.sh --bystrowas not called for the same reason.flang/scripts/tasks.fscript --plan 'task board whole'exhausts its own steplimit on the untouched tree as well, so it proves nothing either way.