Activity

1. johnml1135

   added

   Hermit Crab

   performance

   on Oct 30, 2026
2. github-project-automation

   added this to Servalon Oct 30, 2026
3. github-project-automation

   moved this to 🆕 New in Servalon Oct 30, 2026
4. johnml1135 commented on Nov 7, 2026

   

   johnml1135

   on Nov 7, 2026

   CollaboratorAuthor

   More actions

   Archived from docs/hc-optimization-ledger @ 2eba2f5 (docs/optimizations/shape-sharing-on-clone.md), which is being deleted:

   Copy-on-write Shape sharing in Word.Clone

   Status: retained on perf/hc-optimization-archive, not merged (Word.CloneForEngine, EnsureOwnShape,
   ResetShape, Shape.CreateEmptyLike). Portable: partly. The observation (three of four clone sites
   never use the copied shape) is engine-independent; the mechanism (shared reference to an immutable
   structure) is how a Rust port would naturally write it already.

   The observation

   A CPU profile of the heaviest Sena word (60 s sampled, memoized path) put Word cloning at 36% of wall,
   of which Shape.CopyTo was 28%: every clone deep-copied every ShapeNode, every node's feature
   structure (SimpleFeatureValue.CloneImpl 22%), every annotation, and built a node-mapping dictionary.
   The callers: Word.ReplayOnto 15% (memo replay; never mutates the shape), Word.ExpandAlternatives 9%
   (immediately replaces _shape), AnalysisAffixProcessAllomorphRuleSpec.ApplyRhs 5% (GenerateShape
   immediately Clear()s the copy), non-head clones 2.5%, LexicalLookup (the RootAllomorph setter
   replaces _shape).

   The change

   • Internal engine clone: if the source shape is frozen, share the reference and mark _shapeShared; otherwise
     deep-copy it. Public Word.Clone() always returns a deep, independently mutable shape.
   • EnsureOwnShape(): clone once, on demand, before in-place mutation. Called in AnalysisStratumRule and
     SynthesisStratumRule before phonological rules are applied to a fresh clone (rewrite specs insert
     nodes and hold node references obtained by matching, so the un-share must precede the match).
   • ResetShape(): replace with a new empty shape of the same configuration. Used by the four
     GenerateShape/rebuild sites (analysis affix, analysis compounding subrule, synthesis affix, synthesis
     compounding) instead of clone-then-clear.
   • ExpandAlternatives marks the adopted shape as shared; the RootAllomorph setter and constructors mark
     their fresh clones as owned.

   The safety net is that a shared shape is frozen: any mutation through a missed site throws
   (Shape.CheckFrozen, frozen node feature structures; a missing guard on Annotation.Range was added).
   During development the first full test run threw at exactly the two stratum sites, which is how they
   were found; no other site surfaced across 122 HermitCrab tests, 464 fixture words and 45 Sena words.

   Measured

   Profile after the change: Word clone 36% → 9% of wall, Shape.CopyTo 28.6% → 13% (the remainder is
   the stratum's EnsureOwnShape and part copying inside GenerateShape), ReplayOnto 21.7% → 6.5%. Peak
   working set on the heaviest word 11 GB → 6 GB. Wall-time attribution is unreliable on this machine (see
   the round document); counts are unchanged by construction.

   WordShapeSharingTests covers public-clone independence, copy-on-write, reset paths, and missed-mutation
   guards. The final conformance-inclusive run passed 603 tests with one expected host-privilege skip.
5. Enkidu93

   removed this from Servalon Nov 10, 2026
6. johnml1135 commented on Nov 12, 2026

   

   johnml1135

   on Nov 12, 2026

   CollaboratorAuthor

   More actions

   Reference numbers from the PanGloss port, in case this is revived. In PanGloss, equal shapes and syntactic feature structures made by separate rule applications in one parse are interned to one allocation (sillsdev/PanGloss#249). That is broader than sharing on clone: at a 200M work cap the search held 379,394 words over only 160 distinct shapes and 42 distinct syntactic feature structures, and the private copies were 690 MB of 1,152 MB. Word mindeshi (Mbugwe), peak resident set, one thread:

   work cap	before	shape + feature-structure sharing
   50M	199 MB	82 MB
   200M	1,152 MB	389 MB
   1G	5,074 MB	1,408 MB

   Results were identical on four grammars. Cost: Swahili -0.55%, Amharic +0.51% instructions. These are Rust numbers, not C# ones. If the clone-sharing here leaves equal-but-separate shapes from different rule applications, interning is the follow-on idea. Remaining per-word growth is tracked in #639 (finish each stratum as it completes) and #640 (store folded alternatives as differences).
7. johnml1135 commented on Nov 12, 2026

   

   johnml1135

   on Nov 12, 2026

   CollaboratorAuthor

   More actions

   WIN alloc -35..-47% vs master, wall -23..-37% (Amharic -37%, Mbugwe -23%, Aweti -3%), CPU -5..-24%, mbugwe 250-word peak RSS 2.40->1.88 GB | parse output identical | not pushed, no PR (owner)

   machine #593: internal Word clones share a frozen Shape (PanGloss #232 idea)

   Ready to post as a comment on sillsdev/machine #593. Branch shape-share, commit 8c96a4d7, off origin/master 0cbf85f. Not pushed. Per owner, no PR body written and no further code.

   Change (12 files, +153/-17; 76 of those lines are a new test file)

   • Revived the idea from the archived HermitCrab optimization round record: what was tried, and a faster package that is not mergeable #490 code (perf/hc-optimization-archive: CloneForEngine, EnsureOwnShape, ResetShape), reimplemented minimally on master (the archive is on a memoized, older base and bundles HermitCrab engine: share frozen syntactic feature structures on internal Word clones (about 15% of a heavy Sena profile, never isolated) #595 syntactic-FS sharing, which is NOT included here).
   • Word.CloneForEngine() (internal): if the source shape is frozen, share it (_shapeShared), else deep-copy. Public Word.Clone() and the protected copy ctor stay deep. Non-head words are engine-cloned too.
   • EnsureOwnShape(): copy once before in-place edits; used in SynthesisStratumRule before the phonological rules. ResetShape() + Shape.CreateEmptyLike(): replaces clone-then-Clear() in the four rule-spec sites (AnalysisAffixProcessAllomorph, AnalysisCompoundingSubrule, SynthesisAffixProcessAllomorph, SynthesisCompounding).
   • Engine clones now at: Morpher lexical lookup (2), AnalysisAffixTemplateRule, SynthesisStratumRule (2), SynthesisAffixTemplatesRule (2), AnalysisCompoundingRule (1), Word.ExpandAlternatives. AnalysisStratumRule stays a deep clone (phonological rules mutate and hold node references, so sharing would be undone at once).
   • Safety net: a shared shape is frozen, so a missed mutation throws rather than corrupting. None surfaced in tests or word lists.

   Identity (sorted parses per word: surface, shape string, morph list with morpheme id/gloss/allomorph index, syntactic FS, realizational FS, MPR features, partial flag)

   Compared master vs shape-share vs shape-share+#633, byte-for-byte (cmp): all IDENTICAL.

   list	words	with parses
   Aweti	all 206	157
   Amharic	all 673	190
   Mbugwe	first 250	239
   Sena	154-word sample (every 40th) + first 100	69 + 34
   Limits: Mbugwe line 283 (chaanjiyɛ, the known >16 GB word) excluded; master hit the 6 GB heap limit (OOM) at word 304 on a 400-word run, so Mbugwe was cut to 250. Sena is a sample, not 6146 words.
   Conformance: fieldworks-parser-conformance run through machine's hermitcrab-conformance-runner branch (merged onto master and onto shape-share, HC_CONFORMANCE_SUITE), with the runner's PR633_DUMP hook logging every word's signature: 1414 word results, sorted, identical on both. The 4 conformance gate tests that fail (CoverageReport_*, CheckedInCoverageTablesAreUpToDate, EveryConformanceFixturePassesSelfCheck, EveryFixtureContributesAtLeastOneCoverageRow) fail the same four on unmodified master+runner (suite version vs the pin), unrelated.

   Tests

   HermitCrab.Tests 181 -> 186 passed (5 new WordShapeSharingTests: public clone stays deep, engine clone shares frozen/copies unfrozen, shared shape rejects mutation, EnsureOwnShape, ResetShape). SIL.Machine.Tests 910 passed, 3 skipped (preexisting), 0 failed. dotnet csharpier check . clean. Not run red-without-change for the new tests (they call the new methods).

   A/B (interleaved, order randomised per round, n=15, new process per sample, 1 warm + 2 timed passes, getrusage CPU and ru_maxrss per child, load avg 10-20 on 20 cores, nice, GC heap hard limit 6 GB)

   Arms: B = master; M = master+#633 (alloc-medium e6852c4); S = master+shape-share; SM = both (clean merge). Medians.

   workload	arm	alloc MB/pass	wall ms	CPU s	peak RSS MB	GC pause ms
   Amharic 40	B	2747	2428	11.33	121	572
   	M	2251	2081	10.02	118	447
   	S	1471	1556	8.69	115	344
   	SM	1239	1344	8.13	114	272
   Aweti 15	B	492	739	7.57	248	102
   	M	439	698	7.31	220	98
   	S	305	752	7.29	243	85
   	SM	285	645	6.95	214	84
   Mbugwe 12	B	1419	1597	11.55	231	264
   	M	1216	1440	10.99	229	241
   	S	794	1221	10.21	223	206
   	SM	708	1123	9.63	224	188

   Paired ratios (median; wins = rounds where the new arm is lower):

   compare	workload	alloc	wall	CPU	peak RSS
   S / B (vs master)	Amharic	0.535 (15/15)	0.630 (14/15)	0.761 (14/15)	0.959 (13/15)
   	Aweti	0.622 (15/15)	0.968 (9/15)	0.947 (9/15)	0.988 (13/15)
   	Mbugwe	0.560 (15/15)	0.770 (14/15)	0.904 (13/15)	0.961 (14/15)
   S / M (vs #633 alone)	Amharic	0.653 (15/15)	0.759 (13/15)	0.850 (14/15)	0.975 (10/15)
   	Aweti	0.697 (15/15)	1.093 (6/15)	0.945 (10/15)	1.105 (0/15)
   	Mbugwe	0.653 (15/15)	0.829 (13/15)	0.920 (12/15)	0.965 (10/15)
   SM / M (shape-share on top of #633)	Amharic	0.551 (15/15)	0.655 (15/15)	0.818 (15/15)	0.983 (11/15)
   	Aweti	0.649 (15/15)	0.923 (12/15)	0.900 (11/15)	0.964 (15/15)
   	Mbugwe	0.583 (15/15)	0.750 (15/15)	0.909 (12/15)	0.983 (11/15)
   SM / B (both vs master)	Amharic	0.451	0.569	0.709	0.958
   	Aweti	0.580	0.826	0.898	0.861
   	Mbugwe	0.499	0.681	0.839	0.965

   Peak memory on a long run (word-list dump, 250 Mbugwe words, one process, /usr/bin/time max RSS): master 2.40 GB, shape-share 1.88 GB (-22%), both 2.60 GB (not lower; single run, noisy, not repeated). Aweti dump (206 words): 516 -> 417 MB (S), 497 MB (SM). Wall for the same dumps: Mbugwe 717 s -> 442 s (-38%), Amharic 37 s -> 24 s, Aweti 29 s -> 24 s. These are single runs under load.

   Reading it

   • Allocation halves (-38 to -47%) everywhere and shape sharing is the bigger lever than HermitCrab: cut FeatureStruct clone and VariableBindings allocation #633 (S alone beats M alone on alloc, wall and CPU for Amharic/Mbugwe).
   • Wall and CPU wins are solid on Amharic (-37% / -24%) and Mbugwe (-23% / -10%), smaller and noisy on Aweti (small workload; S/M wall is slightly worse, 6/15). CPU is less noisy than wall under this load.
   • Peak RSS of the short benchmarks is dominated by the loaded grammar (110-250 MB), so it barely moves (-1 to -4%). The memory effect the HermitCrab optimization round record: what was tried, and a faster package that is not mergeable #490 note reported (11 -> 6 GB on the heaviest Sena word) is NOT reproduced here: I did not find that word (the heavy Mbugwe word chaanjiyɛ is a different, unbounded-growth case that sharing does not cure). The only memory evidence is the 250-word Mbugwe run (-22%).
   • Stacking with HermitCrab: cut FeatureStruct clone and VariableBindings allocation #633 is additive (SM/B alloc 0.45-0.58).

   Risks / not done

   • Sena heavy-word memory claim (11 -> 6 GB) not re-measured; no perf counters (not installed), so CPU is getrusage time. Identity on Sena and Mbugwe is a subset. Concurrency (MaxDegreeOfParallelism > 1) not exercised; shared frozen shapes are read-only so it should be safe, untested.
   • Not implemented: lazy un-share in AnalysisStratumRule (clones there are still deep), HermitCrab engine: share frozen syntactic feature structures on internal Word clones (about 15% of a heavy Sena profile, never isolated) #595 syntactic-FS sharing.
   • Temporary helpers left: branch/worktrees shape-share-on633, ss-conf-base, ss-conf-shape under ~/work/machine.worktrees (local only, delete with the branches conf-base-tmp, conf-shape-tmp). Bench harness and raw data: lanes/machine-shape-share-bench (ab-results.txt, dumps/, ab2.py, sum3.py).