Skip to content

API: change default aux dir to .tlacache - #20

Merged
damiendoligez merged 1 commit into
tlaplus:masterfrom
johnyf:auxdir
Feb 16, 2021
Merged

API: change default aux dir to .tlacache#20
damiendoligez merged 1 commit into
tlaplus:masterfrom
johnyf:auxdir

Conversation

@johnyf

@johnyf johnyf commented Dec 7, 2020

Copy link
Copy Markdown
Contributor

This change makes .tlacache the default auxiliary directory, following #16.

In the future, a command-line option could be added for specifying the auxiliary directory.

@johnyf johnyf mentioned this pull request Dec 21, 2020

@damiendoligez damiendoligez left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Looks good to me.

@damiendoligez
damiendoligez merged commit 7556ec5 into tlaplus:master Feb 16, 2021
@johnyf

johnyf commented Feb 16, 2021

Copy link
Copy Markdown
Contributor Author

Thank you for reviewing and merging this pull request.

@johnyf
johnyf deleted the auxdir branch March 4, 2021 16:18
wkirschenmann pushed a commit to wkirschenmann/tlapm that referenced this pull request Aug 19, 2026
Real-solver FfiGrpc runs, fixed RSS sampler, binaries chosen just
before and after each expected transition. Records the full curve:
base 1.6 verdicts/s truncated; tlaplus#17 unlocks x8 then OOMs at 13.9 GB
(as does tlaplus#19); tlaplus#20 is what makes the run finish (5.1 GB); tlaplus#21 buys
x3.5 throughput; phase 4 flattens RSS to 439 MB at unchanged wall.
End to end: unfinishable -> 3.5 min at 439 MB, 48 verdicts/s
sustained, above the 43 v/s chunked-run target.

Co-Authored-By: Claude <noreply@anthropic.com>
wkirschenmann pushed a commit to wkirschenmann/tlapm that referenced this pull request Aug 19, 2026
The corpus that motivated the effort, measured across the transition
binaries with real solvers: before the expand_defs unlock every binary
OOMs early; tlaplus#20 collapses 77 -> 11.6 v/s as its heap walks to 13.6 GB
and dies 380 verdicts short; tlaplus#21/tlaplus#24 finish with ~1 GB of margin on a
15 GB container while phase 4 does the same run at 1.39 GB flat and
~20 % faster in every quartile -- the user's measured GC coupling,
reproduced and then removed. On the user's 7.7 GB machine only the
phase-4 binary can finish: from '17 hours then OOM at 58 %' to 7.5
minutes.

Co-Authored-By: Claude <noreply@anthropic.com>
wkirschenmann pushed a commit to wkirschenmann/tlapm that referenced this pull request Aug 21, 2026
…dencies

The ranked table did not have the granularity it implied, and the
"recommended path" section restated it instead of adding to it.

**Granularity.** The table now says explicitly that one row is one pull
request, with a `commits` column: eight rows are groups that must land
together (tlaplus#3 is 2 commits, tlaplus#15 is 3, tlaplus#17 is 2, tlaplus#18 is 3, tlaplus#19 is 6, tlaplus#20 is
5), twelve are a single commit. Item 17 also notes that the branch holds
an add-then-revert pair there -- the prover-slot-division hypothesis,
refuted by measurement -- which must not be replayed.

**Two rows were missing entirely**, and the question exposed it: the
measurement instrumentation (17 commits, 22 files, +876/-92) and the
harness (5 commits, 9 files, +698) buy no performance, but T1/T2 *are*
the checker and P1/P2/P3 *are* the probes -- an optimization landed
before them ships unverifiable. They are now rows 0 and 0b, marked as
prerequisites rather than ranked.

**The step ladder is gone.** Its six invented steps were the same items
regrouped, with diffstat totals that restated the table; the batching
advice it carried was already one sentence in the summary. Replaced by
what it should have been -- dependencies:

  * hard, in code: tlaplus#3's two commits in order; tlaplus#11 requires tlaplus#15 (the memo
    attaches to the expansion fold's shared states); tlaplus#15's three commits
    in order; tlaplus#19 requires tlaplus#5 (it patches the tree tlaplus#5 rebuilt); #0/#0b
    before any protocol;
  * hard by measurement, and easy to miss: tlaplus#3 must not ship without tlaplus#15
    on INSTANCE-heavy specs -- alone it takes preparation 458 -> 842 s
    while cutting peak memory 13.9 -> 5.1 GB;
  * soft: tlaplus#6+#12 for the x49 elaboration clock, tlaplus#10+tlaplus#16 for the memory
    unlock, tlaplus#18/tlaplus#19 adjacent but independent.

Plus three review-coherent batches, stated as such rather than as a path.
The protocol section is renamed to say what it is.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CUUoeEmuL3jsYhUb3UrhJH
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Development

Successfully merging this pull request may close these issues.

2 participants