{
  "schemaVersion": "0.11.0",
  "canonical": "https://www.pystone.net/notes/in-flight-load-local-speedup-tail-latency-literature/",
  "atlas": "https://www.pystone.net/?node=in-flight-load-local-speedup-tail-latency-literature#knowledge-atlas",
  "markdown": "https://www.pystone.net/notes/in-flight-load-local-speedup-tail-latency-literature.md",
  "context": "https://www.pystone.net/notes/in-flight-load-local-speedup-tail-latency-literature.context.json",
  "knowledgeVersion": "224c990773de.5fa8af6e39fa",
  "build": {
    "siteCommit": "224c990773de166d23a886306577dd90379529ce",
    "notesCommit": "5fa8af6e39fa3891d1b9b4832bfa6c4e0ecaaf0a",
    "builtAt": "1970-01-01T00:00:00.000Z",
    "version": "224c990773de.5fa8af6e39fa"
  },
  "id": "note:in-flight-load-local-speedup-tail-latency-literature",
  "slug": "in-flight-load-local-speedup-tail-latency-literature",
  "title": "在途量、局部加速与尾延迟：性能成本拓扑文献摘录",
  "type": "note",
  "visibility": "public",
  "idStability": "rename-stable",
  "author": {
    "name": "Perrin Yong",
    "profile": "https://www.pystone.net/profile/"
  },
  "publisher": {
    "name": "Perrin Yong",
    "profile": "https://www.pystone.net/profile/"
  },
  "aliases": [],
  "summary": "在途量、局部加速与尾延迟：性能成本拓扑文献摘录",
  "contentRole": "unspecified",
  "isMoc": false,
  "mocRecognition": "none",
  "generated": false,
  "attribution": "unspecified",
  "domain": "99-资料来源、摘录与归档",
  "tags": [],
  "mocs": [],
  "contentHash": "dc7d1592b15517e0be1a997076876f7c416e0d0e91104534b49f289d57d2167f",
  "assets": [],
  "headings": [
    {
      "depth": 1,
      "text": "在途量、局部加速与尾延迟：性能成本拓扑文献摘录",
      "anchor": "在途量局部加速与尾延迟性能成本拓扑文献摘录",
      "citation": "https://www.pystone.net/notes/in-flight-load-local-speedup-tail-latency-literature/#%E5%9C%A8%E9%80%94%E9%87%8F%E5%B1%80%E9%83%A8%E5%8A%A0%E9%80%9F%E4%B8%8E%E5%B0%BE%E5%BB%B6%E8%BF%9F%E6%80%A7%E8%83%BD%E6%88%90%E6%9C%AC%E6%8B%93%E6%89%91%E6%96%87%E7%8C%AE%E6%91%98%E5%BD%95"
    },
    {
      "depth": 2,
      "text": "Little（1961）：平均在途量、到达率与停留时间",
      "anchor": "little1961平均在途量到达率与停留时间",
      "citation": "https://www.pystone.net/notes/in-flight-load-local-speedup-tail-latency-literature/#little1961%E5%B9%B3%E5%9D%87%E5%9C%A8%E9%80%94%E9%87%8F%E5%88%B0%E8%BE%BE%E7%8E%87%E4%B8%8E%E5%81%9C%E7%95%99%E6%97%B6%E9%97%B4"
    },
    {
      "depth": 2,
      "text": "Amdahl（1967）：未改善部分限制整体加速",
      "anchor": "amdahl1967未改善部分限制整体加速",
      "citation": "https://www.pystone.net/notes/in-flight-load-local-speedup-tail-latency-literature/#amdahl1967%E6%9C%AA%E6%94%B9%E5%96%84%E9%83%A8%E5%88%86%E9%99%90%E5%88%B6%E6%95%B4%E4%BD%93%E5%8A%A0%E9%80%9F"
    },
    {
      "depth": 2,
      "text": "Dean 与 Barroso（2013）：规模化系统中的尾延迟",
      "anchor": "dean-与-barroso2013规模化系统中的尾延迟",
      "citation": "https://www.pystone.net/notes/in-flight-load-local-speedup-tail-latency-literature/#dean-%E4%B8%8E-barroso2013%E8%A7%84%E6%A8%A1%E5%8C%96%E7%B3%BB%E7%BB%9F%E4%B8%AD%E7%9A%84%E5%B0%BE%E5%BB%B6%E8%BF%9F"
    },
    {
      "depth": 2,
      "text": "三者如何共同约束诊断",
      "anchor": "三者如何共同约束诊断",
      "citation": "https://www.pystone.net/notes/in-flight-load-local-speedup-tail-latency-literature/#%E4%B8%89%E8%80%85%E5%A6%82%E4%BD%95%E5%85%B1%E5%90%8C%E7%BA%A6%E6%9D%9F%E8%AF%8A%E6%96%AD"
    }
  ],
  "claims": [],
  "outgoing": [
    {
      "id": "note:performance-optimization-cost-frequency-scope",
      "title": "性能优化的本质是控制成本发生的频率、时机与范围",
      "url": "https://www.pystone.net/notes/performance-optimization-cost-frequency-scope/",
      "atlas": "https://www.pystone.net/?node=performance-optimization-cost-frequency-scope#knowledge-atlas",
      "label": "性能优化的本质是控制成本发生的频率、时机与范围",
      "origin": "explicit",
      "humanReviewed": true,
      "context": "三份来源分别约束驻留、可改善比例和尾部传播，不能拼成跨系统的统一性能公式。综合迁移见性能优化的本质是控制成本发生的频率、时机与范围。",
      "citation": "https://www.pystone.net/notes/in-flight-load-local-speedup-tail-latency-literature/#%E5%9C%A8%E9%80%94%E9%87%8F%E5%B1%80%E9%83%A8%E5%8A%A0%E9%80%9F%E4%B8%8E%E5%B0%BE%E5%BB%B6%E8%BF%9F%E6%80%A7%E8%83%BD%E6%88%90%E6%9C%AC%E6%8B%93%E6%89%91%E6%96%87%E7%8C%AE%E6%91%98%E5%BD%95"
    }
  ],
  "incoming": [
    {
      "id": "note:performance-optimization-cost-frequency-scope",
      "title": "性能优化的本质是控制成本发生的频率、时机与范围",
      "url": "https://www.pystone.net/notes/performance-optimization-cost-frequency-scope/",
      "atlas": "https://www.pystone.net/?node=performance-optimization-cost-frequency-scope#knowledge-atlas",
      "label": "性能优化的本质是控制成本发生的频率、时机与范围",
      "origin": "explicit",
      "humanReviewed": true,
      "context": "Little（1961），A Proof for the Queuing Formula: L = λW在有限均值、严格平稳等条件下证明平均在途数量等于到达率乘以平均停留时间。这支持把生命周期从“总成本乘数”中拆出，单独检查对象驻留、排队和未完成工作；它不证明降低 必然改善所有体验，也不提供因果方向。",
      "citation": "https://www.pystone.net/notes/performance-optimization-cost-frequency-scope/#%E4%B8%89%E4%B8%AA%E7%8B%AC%E7%AB%8B%E7%90%86%E8%AE%BA%E7%BA%A6%E6%9D%9F"
    },
    {
      "id": "note:performance-optimization-cost-frequency-scope",
      "title": "性能优化的本质是控制成本发生的频率、时机与范围",
      "url": "https://www.pystone.net/notes/performance-optimization-cost-frequency-scope/",
      "atlas": "https://www.pystone.net/?node=performance-optimization-cost-frequency-scope#knowledge-atlas",
      "label": "性能优化的本质是控制成本发生的频率、时机与范围",
      "origin": "explicit",
      "humanReviewed": true,
      "context": "Amdahl（1967），Validity of the Single Processor Approach说明整体加速受未改善部分约束。迁移到本页时，它限制了“热点局部基准变快即整体变快”的推断：只有被改部分在实际工作中的占比和瓶颈位置足够大，局部收益才可能成为整体收益。",
      "citation": "https://www.pystone.net/notes/performance-optimization-cost-frequency-scope/#%E4%B8%89%E4%B8%AA%E7%8B%AC%E7%AB%8B%E7%90%86%E8%AE%BA%E7%BA%A6%E6%9D%9F"
    },
    {
      "id": "note:performance-optimization-cost-frequency-scope",
      "title": "性能优化的本质是控制成本发生的频率、时机与范围",
      "url": "https://www.pystone.net/notes/performance-optimization-cost-frequency-scope/",
      "atlas": "https://www.pystone.net/?node=performance-optimization-cost-frequency-scope#knowledge-atlas",
      "label": "性能优化的本质是控制成本发生的频率、时机与范围",
      "origin": "explicit",
      "humanReviewed": true,
      "context": "Dean 与 Barroso（2013），The Tail at Scale显示，系统规模或利用率上升时，单个组件中偶发的高延迟可能主导大规模服务的整体尾延迟。这支持把扇出、关键路径和耗时分布加入模型；具体放大幅度仍取决于相关性、超时、冗余和调度策略，不能直接把独立概率公式套到所有系统。",
      "citation": "https://www.pystone.net/notes/performance-optimization-cost-frequency-scope/#%E4%B8%89%E4%B8%AA%E7%8B%AC%E7%AB%8B%E7%90%86%E8%AE%BA%E7%BA%A6%E6%9D%9F"
    },
    {
      "id": "note:sources-excerpts-archive",
      "title": "资料来源、摘录与归档",
      "url": "https://www.pystone.net/notes/sources-excerpts-archive/",
      "atlas": "https://www.pystone.net/?node=sources-excerpts-archive#knowledge-atlas",
      "label": "资料来源、摘录与归档",
      "origin": "explicit",
      "humanReviewed": true,
      "context": "01-外部文章与转载摘录中的“在途量、局部加速与尾延迟：性能成本拓扑文献摘录”导航项",
      "citation": "https://www.pystone.net/notes/sources-excerpts-archive/#01-%E5%A4%96%E9%83%A8%E6%96%87%E7%AB%A0%E4%B8%8E%E8%BD%AC%E8%BD%BD%E6%91%98%E5%BD%95"
    }
  ],
  "contentMarkdown": "# 在途量、局部加速与尾延迟：性能成本拓扑文献摘录\n\n> [!source] 来源摘录\n> 三份来源分别约束驻留、可改善比例和尾部传播，不能拼成跨系统的统一性能公式。综合迁移见[[10-计算机、信息技术与工程/05-游戏图形与运行时/游戏性能优化/性能优化的本质是控制成本发生的频率、时机与范围|性能优化的本质是控制成本发生的频率、时机与范围]]。\n\n## Little（1961）：平均在途量、到达率与停留时间\n\n- 文献：John D. C. Little，*A Proof for the Queuing Formula: L = λW*\n- 期刊：*Operations Research*, 9(3), 383–387\n- 原文：[INFORMS](https://doi.org/10.1287/opre.9.3.383)\n\n论文证明：在相关均值有限、随机过程严格平稳且到达过程满足文中条件时，系统中的平均单位数 `L` 等于平均到达率 `λ` 乘以单位在系统中的平均时间 `W`。这是长期平均量之间的恒等关系，不依赖具体服务时间分布。\n\n它支持把对象数量、到达频率和驻留时间分开测量；不提供因果方向，也不证明只降低 `W` 就一定改善用户体验。瞬态、非平稳、有限窗口或有选择地丢弃请求时，需要重新检查条件。\n\n## Amdahl（1967）：未改善部分限制整体加速\n\n- 文献：Gene M. Amdahl，*Validity of the Single Processor Approach to Achieving Large Scale Computing Capabilities*\n- 会议：AFIPS Spring Joint Computer Conference, 483–485\n- 原文：[ACM](https://doi.org/10.1145/1465482.1465560)\n\n原文针对多处理器扩展的乐观预期，强调现实计算包含难以并行的数据管理和顺序部分；仅提高可并行部分，整体吞吐会受未改善部分限制。后来常见的“阿姆达尔定律”表达了同一上界结构。\n\n迁移到普通性能优化时，可靠含义是：必须先知道热点在真实工作负载中的占比与关键路径位置。微基准的巨大局部加速，如果作用于很小占比、被其他瓶颈遮蔽或增加协调成本，整体收益仍可能很小。原论文的具体工作负载比例不应当作今天系统的通用常数。\n\n## Dean 与 Barroso（2013）：规模化系统中的尾延迟\n\n- 文献：Jeffrey Dean、Luiz André Barroso，*The Tail at Scale*\n- 期刊：*Communications of the ACM*, 56, 74–80\n- 原文：[Google Research](https://research.google/pubs/the-tail-at-scale/)\n\n论文分析大规模在线服务：当一次请求扇出到许多组件时，单个组件偶发的高延迟会提高整条请求落入慢尾部的概率；系统规模或利用率上升后，中等规模时可忽略的抖动可能主导用户可见延迟。作者归纳了延迟来源，并讨论 hedged requests、备用副本、微分区等降低尾部影响的策略。\n\n它支持观察分位数、扇出和关键路径，而不只看平均值。具体尾部放大仍受组件相关性、超时、负载平衡、冗余和调度影响；不能把独立概率假设机械套到所有系统，冗余请求也会增加资源成本。\n\n## 三者如何共同约束诊断\n\n| 问题 | 对应来源 | 应测变量 |\n|---|---|---|\n| 为什么对象或未完成工作越来越多 | Little | 到达率、停留时间、平均在途量 |\n| 为什么热点变快而整体几乎不变 | Amdahl | 改善部分占比、未改善部分、协调开销 |\n| 为什么平均正常而用户仍感到卡顿 | Dean 与 Barroso | 高分位延迟、扇出、相关性、关键路径 |\n\n三份来源支持的是诊断顺序和变量方向，不是“成本 = 频率 × 生命周期 × 扇出”的精确普适方程。\n"
}