关于「预算分配:把页面百分位拆到每一个调用」,《扇出尾延迟与对冲请求》给出的定量结论是什么?

结论:第六章把页面预算拆给各调用,按成本不等分可省下 0.0597 的容量成本。

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: 0.0597

class n cost k equal share optimal share
cache60 1.0 0.000100 0.000032
db30 4.0 0.000100 0.000130
search9 10.0 0.000100 0.000324
partner1 40.0 0.000100 0.001295
capacity cost, equal split14.2683
capacity cost, optimal split13.4164
saving0.0597 of the equal-split cost
check P(page meets B)0.990000
查看完整验算程序
# Chapter 6. p62: unequal split of the slow share by marginal cost.
# Model (illustrative): four classes of calls in one page, n_i calls
# each, all with deadline B = 200 ms. Holding class i to slow share e_i
# needs a rate gap g_i = ln(1/e_i)/B per server at cost k_i per unit
# gap. Minimise sum n_i k_i g_i subject to
#   sum n_i (-ln(1 - e_i)) = -ln(0.99)   (page meets B with 0.99).
# Lagrange: e_i / (1 - e_i) = k_i / (nu B); nu found by bisection.
import math

B = 200.0
CLASSES = (("cache", 60, 1.0), ("db", 30, 4.0), ("search", 9, 10.0),
           ("partner", 1, 40.0))
BUDGET = -math.log(0.99)


def used(nu):
    terms = []
    for _, n, k in CLASSES:
        x = k / (nu * B)
        e = x / (1.0 + x)
        terms.append(-n * math.log1p(-e))
    return math.fsum(terms)


def cost(eps):
    return math.fsum(n * k * math.log(1.0 / e) / B
                     for (_, n, k), e in zip(CLASSES, eps))


if __name__ == "__main__":
    lo, hi = 1e-6, 1e9
    for _ in range(300):
        mid = math.sqrt(lo * hi)
        if used(mid) > BUDGET:
            lo = mid
        else:
            hi = mid
    nu = hi
    opt = [(k / (nu * B)) / (1.0 + k / (nu * B)) for _, _, k in CLASSES]
    total_n = math.fsum(n for _, n, _ in CLASSES)
    eq = -math.expm1(math.log(0.99) / total_n)
    print("class      n    cost k   equal share   optimal share")
    for (name, n, k), e in zip(CLASSES, opt):
        print("  {:<8} {:3d}   {:6.1f}    {:.6f}      {:.6f}".format(
            name, n, k, eq, e))
    c_eq, c_opt = cost([eq] * len(CLASSES)), cost(opt)
    print("capacity cost, equal split:   {:.4f}".format(c_eq))
    print("capacity cost, optimal split: {:.4f}".format(c_opt))
    print("saving: {:.4f} of the equal-split cost".format(1.0 - c_opt / c_eq))
    chk = math.fsum(-n * math.log1p(-e) for (_, n, _), e in
                    zip(CLASSES, opt))
    print("check P(page meets B) = {:.6f}".format(math.exp(-chk)))

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 6 章「预算分配:把页面百分位拆到每一个调用」(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