[
https://issues.apache.org/jira/browse/OPENNLP-1878?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Martin Wiesner updated OPENNLP-1878:
------------------------------------
Issue Type: Improvement (was: New Feature)
> Non-breaking performance follow-up for normalization and tokenization hot
> paths
> -------------------------------------------------------------------------------
>
> Key: OPENNLP-1878
> URL: https://issues.apache.org/jira/browse/OPENNLP-1878
> Project: OpenNLP
> Issue Type: Improvement
> Affects Versions: 3.0.0-M4
> Reporter: Kristian Rickert
> Assignee: Kristian Rickert
> Priority: Major
> Fix For: 3.0.0-M5
>
> Time Spent: 2h 40m
> Remaining Estimate: 0h
>
> {quote}Not ready for review yet: this branch is based on the tip of GitHub PR
> apache/opennlp#1105 (the OPENNLP-1850 "dl" rung), which is approved but not
> yet merged to main. The PR should stay open or in draft until #1105 lands, at
> which point the branch is rebased --onto main and this note is removed.{quote}
> h2. Summary
> A vetted set of non-breaking performance improvements for the OPENNLP-1850
> normalizer, tokenizer, and DL hot paths, aimed at heavy use as a search-index
> analyzer and tokenizer. Every item is output-preserving (byte-identical),
> offset-safe, thread-safe, free of regex on the text path, and ASCII-only in
> source. These were verified against the source by reading the relevant
> methods during a performance review and were deliberately deferred out of the
> OPENNLP-1850 review so as not to delay it. Land as one follow-up PR (or a
> small stack), and pair with a JMH benchmark run in a fresh JVM per case on
> realistic mixed-script text to confirm each win before and after.
> h2. Scope (ranked, tier one first)
> * Identity short-circuit on the non-aligned {{CharClass}} set folds
> ({{normalize}}, {{collapse}}, {{removeAll}}, and the static {{substitute}}).
> They currently allocate a {{StringBuilder}} and copy code point by code point
> even when nothing matches. Change to a lazy builder: scan for the first
> member, return the input unchanged when there is none, and allocate the
> builder pre-filled with the unchanged prefix only on the first hit. Biggest
> win on ASCII-heavy token streams; roughly neutral on text saturated with
> folded characters. Scope to the non-aligned path; leave the {{*Aligned}}
> variants alone unless the identity-{{Alignment}} compose is verified a true
> no-op.
> * {{Confusables.skeleton}} fast-path plus an optional bounded cache. Add a
> {{BitSet}} pre-filter of "is this code point ever a prototype key" so the
> common no-mapping case skips the two {{java.text.Normalizer}} NFD passes and
> the per-code-point boxing, and add a size-bounded, thread-safe value cache.
> Both fire only when {{CONFUSABLE_FOLD}} is configured. The two NFD passes are
> required by UTS #39 when a mapping does apply, so the floor stays at two for
> actual confusables.
> * Hoist the per-code-point volatile reads in tokenization.
> {{WordBreakProperty.ordinalOf}} does a volatile {{data()}} read per code
> point and {{ExtendedPictographic.is}} reads a volatile field per call.
> Resolve the resolved-data snapshot and the {{BitSet}} once per document in
> {{WordSegmenter.forEachSegment}} and {{WordType.of}} through package-private
> overloads, keeping the lazy first-resolution and the supplementary path.
> Removes a volatile load and a call frame per code point on the hottest
> tokenization op.
> * BMP fast-path the {{Character.codePointAt}} and {{Character.charCount}}
> pairs in {{WordSegmenter}}, {{WordType.of}}, and the {{CharClass}} cursor
> loops, so a non-surrogate {{char}} skips the JDK call and the redundant
> surrogate re-test (over 99 percent of Latin text).
> * {{DigitCharSequenceNormalizer}}: replace the per-digit {{String.valueOf}}
> allocation with a precomputed {{static final String[]}} for "0" through "9".
> Output-preserving (length still one).
> * {{Alignment.Builder}}: add a pre-size constructor {{Builder(int
> expectedLength)}} seeding the {{starts}}/{{ends}} arrays; every caller
> already knows {{text.length()}}. Kills the array-regrow chain and makes the
> defensive trim near-exact on the offset and highlighting path. Keep the
> no-arg constructor.
> * DL small cleanups: size the score and token lists in
> {{DocumentCategorizerDL}} and {{NameFinderDL}} from the known counts
> ({{ArrayList}} rather than {{LinkedList}}), and pre-size the ONNX input
> {{HashMap}}. Correctness-neutral.
> h2. Acceptance criteria
> * Every change is output-preserving (byte-identical), offset-safe, and
> thread-safe; the existing OPENNLP-1850 test suites pass unchanged.
> * A JMH benchmark, one fresh JVM per case, on realistic mixed-script text
> confirms the tier-one wins (the identity short-circuit and the
> {{Confusables}} fast-path) before and after.
> * No new dependencies, ASCII-only source, no regex on the text path.
> h2. Out of scope
> * Any behavior or output change.
> * The {{*Aligned}} {{CharClass}} variants, unless the identity-{{Alignment}}
> compose is verified a true no-op.
> * Paths already verified optimal and not to be re-investigated:
> {{CodePointSet}} BitSet membership, the compiled UAX-29 transition tables,
> the {{Alignment}} binary searches and {{andThen}}, and the
> {{mergeOverlappingSpans}} TreeMap (deliberately O(n log n) to yield document
> order).
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
