HermitCrab analysis: store folded equivalent analyses as differences, not whole Words (per-word memory)

未关闭
  • 哪里不对？
#640 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
5/5
预计耗时
一周以上
新手友好度
35/100
Issue 类型
功能
描述清晰度
描述清楚
活跃度
活跃
技术栈
csharp
领域
backend, performance

调研方向

Read AnalysisStratumRule.cs around lines 162–171 and Word.cs around lines 80–94 and 467–513 to trace how alternatives are stored and expanded. The issue says to first measure how many bytes folded alternatives hold in C# on a heavy word; the proposed storage change is recorded for later and is not scheduled. Done means alternatives can be rebuilt from their differences while synthesis results remain exact.

由索引模型根据 Issue 内容生成。

描述

Hermit Crab performance

HermitCrab stores each folded equivalent analysis as a whole Word. When two analyses have the same merge key, AnalysisStratumRule.Apply does canonicalWord.Alternatives.Add(mruleOutWord) (AnalysisStratumRule.cs:167-171), keeping the full word: its own Shape, two feature structures, _allomorphs dictionary, _mruleApps list and non-head list, all copied in the Word(Word) constructor (Word.cs:80-94). What actually differs from the canonical is small: the morphological rule trail, the allomorph trail and the syntactic feature structure.

Recorded for later, not scheduled. Related to #593 and #595 (sharing the shape and feature structures inside each word); this one is about not keeping a full word per alternative at all.

No C# memory measurement exists for this mechanism. The numbers are PanGloss ones; the C# structure is the same (Word.Alternatives, Word.cs:467) and C# already rebuilds alternatives from a difference on demand in Word.ExpandAlternatives (Word.cs:472-513), which only applies trail and feature-structure changes on top of the source.

Idea: store a folded alternative as its differences (rule trail, allomorph trail, syntactic feature structure) and rebuild the Word when synthesis expands it. Keep the word whole whenever any other field differs, so results stay exact.

Reference implementation: sillsdev/PanGloss#250 (stacked on #249). Results are identical to before (status, signature, work steps on four grammars); the change is storage only.

Evidence (PanGloss, Rust)

Word mindeshi, Mbugwe, peak resident set, one thread. #250 on top of shape and feature-structure sharing: 50M cap 82 to 59 MB, 200M cap 389 to 267 MB, 1G cap 1,408 to 774 MB. At a 200M cap the search held 94,010 canonical words and 285,384 alternatives, three quarters of all words. An alternative was a 336-byte inline Word plus about 256 bytes of heap for trails and a map; in that search the alternatives differed from their canonical only by trail and syntactic features. Instructions did not regress (Swahili -3.48%, Cebuano -2.71%, Amharic +0.13%).

Notes and risks
  • C# alternatives are created from mruleOutWord, which is itself a clone-derived word, so #593 shrinks them but does not remove the per-alternative object, dictionary and lists.
  • ExpandAlternatives reads the canonical at expansion time, and the canonical's syntactic feature structure is widened by later arrivals, so a delta must be rebuilt from the canonical at expansion, not fold, time.
  • The Source chain (AnalysisStratumRule.cs:162) keeps earlier strata's words reachable from alternatives; a delta form must not break ExpandAlternatives.
  • Related: #512 (cull candidates at emission) attacks the same retained set from the other side.

Next: recorded for later, not scheduled. Measure how many bytes folded alternatives hold in C# on a heavy word before building anything.

主要语言
C#
星标
29
派生
16
平均合并
4 天 11 小时
30 天内合并 PR
14

环境准备

  • 没有 Dockerfile 或 Docker Compose 文件
  • 有 Pull Request 模板
  • 没有贡献指南

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

sillsdev/machine 的其他 Issue

查看 sillsdev/machine 的全部 Issue

相似的 Issue

更多 C# Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。