关于「排队尾分布:利用率把尾巴拉长,对冲也是负载」,《扇出尾延迟与对冲请求》给出的定量结论是什么?

结论:先看利用率的作用。利用率 0.30 时,单个调用的平均响应时间是 7.1 毫秒,第九十九百分位是 32.9 毫秒;利用率 0.70 时,分别升到 16.7 与 76.8 毫秒;利用率 0.90 时,升到 50.0 与 230.3 毫秒;利用率 0.95 时,平均是 100.0 毫秒,第九十九百分位是 460.5 毫秒。从 0.5 到 0.9,每一个百分位都乘以 5.0,因为所有百分位都与「一减 rho」成反比。利用率最后那几个百分点最贵:从 0.90 到 0.95 只多用了五个百分点,尾部却翻了一倍。

数据口径:书中服务时间、慢请求比例与成本均为示意参数;全部计算由书中程序生成;本书属于非个人化经营建议。 书中以标注框区分已发表的证据、可运行的 Python 计算(22 段程序)、原典与作者的假说。

English version

对照数据表

下表数字由本章 Python 程序原样打印,可重新运行复现。 程序只用 Python 标准库与结论吻合的打印值: 0.30, 0.5, 0.70, 0.9, 0.90, 0.95

M/M/1, service mean 5 ms, page = 100 parallel calls
rhomean Tcall p99page p50page p99ratio
0.307.132.935.565.82.00
0.5010.046.149.892.12.00
0.7016.776.882.9153.42.00
0.8025.0115.1124.4230.12.00
0.9050.0230.3248.8460.32.00
0.95100.0460.5497.5920.52.00
page p99 / call p99 does not depend on rho: 1.999
going from rho 0.5 to 0.9 multiplies every quantile by 5.0
查看完整验算程序
# Chapter 5. p51: M/M/1 response-time tail and the fan-out page.
# Model: Poisson arrivals at rate lam, exponential service with mean
# 5 ms (mu = 0.2 per ms, illustrative). In an M/M/1 queue under FIFO the
# response time is exponential with rate mu - lam, so
#   P(T > t) = exp(-(mu - lam) t),  call p99 = ln(100) / (mu - lam),
# and a page of N = 100 independent calls has
#   page p99 = -ln(1 - 0.99^(1/100)) / (mu - lam).
import math

MU = 0.2


def call_q(rho, q):
    return -math.log(1.0 - q) / (MU * (1.0 - rho))


def page_q(rho, q, n=100):
    return -math.log(1.0 - q ** (1.0 / n)) / (MU * (1.0 - rho))


if __name__ == "__main__":
    print("M/M/1, service mean 5 ms, page = 100 parallel calls")
    print("   rho   mean T   call p99   page p50   page p99   ratio")
    for rho in (0.3, 0.5, 0.7, 0.8, 0.9, 0.95):
        mean = 1.0 / (MU * (1.0 - rho))
        c99, g50, g99 = call_q(rho, 0.99), page_q(rho, 0.5), page_q(rho, 0.99)
        fmt = "  {:4.2f}  {:7.1f}  {:9.1f}  {:9.1f}  {:9.1f}   {:5.2f}"
        print(fmt.format(rho, mean, c99, g50, g99, g99 / c99))
    print("page p99 / call p99 does not depend on rho: {:.3f}".format(
        page_q(0.5, 0.99) / call_q(0.5, 0.99)))
    print("going from rho 0.5 to 0.9 multiplies every quantile by {:.1f}"
          .format((1.0 - 0.5) / (1.0 - 0.9)))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《扇出尾延迟乘法律与对冲请求、单页多调用时百分位延迟的放大母典》第 5 章「排队尾分布:利用率把尾巴拉长,对冲也是负载」(版本 1.0.0). 5D Library. DOI 待发布

DOI 待发布 · 专著页正式发布后上线(本站暂无专著页) · 5DL-B-326 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026326,
  author    = {Fan, Yafu},
  title     = {The Multiplicative Law of Fan-Out Tail Latency and Hedged Requests: How Percentile Latency Amplifies When One Page Makes Many Calls},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论