关于「分片与分层:同样的槽位,不同的摆法」,《缓存命中率与车氏近似》给出的定量结论是什么?

结论:再看车氏近似一栏。前一级的命中率 0.1931、0.2613、0.3450、0.4027,与单一缓存在同样容量下的值完全一致,这一级没有问题。问题出在后一级:两级合计的近似值是 0.4319、0.4190、0.4177 与 0.4309,比模拟值系统地偏高,对半分时偏高约六个百分点。原因在于,第一级的缺失流并不是独立引用:一个对象刚刚在第一级缺失,它就被放进了第一级,接下来一段时间里它的请求都会在第一级命中,不会再到达第二级;也就是说,到达第二级的请求在时间上是「打散」过的,同一对象的请求被拉开了间隔。车氏近似把它当作独立引用,就高估了第二级的命中机会。

数据口径:本书用标准库程序推翻这条直觉:在齐普夫请求下,最近最少使用(LRU)缓存的命中率随容量增长得很慢,示意参数下容量放大一百二十八倍,命中率只放大 6.80 倍;而车氏近似只凭请求概率与容量,就能事先算出 LRU 命中率,与逐条模拟的最大差距为 0.0013。 书中参数为示意;全部计算由书中程序生成;本书给出只读的容量建议,不访问任何服务器、不写任何配置;「排盘」仅作一种计算结果的名称,不做任何命理推断;属于非个人化经营建议。

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: 0.1931, 0.2613, 0.3450, 0.4027, 0.4177, 0.4190

Two LRU levels with 2000 slots in all; one cache: 0.4483
C1C2Che L1Che L2Che allsim all
25017500.19310.23880.43190.4277
50015000.26130.15770.41900.3968
100010000.34500.07270.41770.3562
15005000.40270.02810.43090.4048
查看完整验算程序
# Chapter 6. p62: two LRU levels in tandem, the setting of Che (2002).
# Model: level 1 has C1 slots and sees all requests; its misses go to
# level 2 with C2 slots. Following Che, Tung and Wang, the miss stream
# of level 1 is treated as independent references with rates
# q_i = p_i (1 - h1_i); level 2 is solved by the same Che equation on
# q normalised. Overall hit = H1 + sum q_i h2_i. A single LRU of
# C1 + C2 slots is the yardstick. Seeded trace of 250000 Zipf requests
# (a = 0.8, N = 20000, illustrative), 50000 warm-up.
import math
import random
from collections import OrderedDict

N, A, L, WARM, TOTAL = 20000, 0.8, 250000, 50000, 2000


def zipf(n, a):
    w = [1.0 / (i ** a) for i in range(1, n + 1)]
    s = math.fsum(w)
    return [x / s for x in w]


def occupancy(p, t):
    return math.fsum(-math.expm1(-x * t) for x in p)


def che_t(p, c, tol=1e-9):
    lo, hi = 0.0, 1.0
    while occupancy(p, hi) < c:
        hi = 2.0 * hi
    while hi - lo > tol * hi:
        mid = 0.5 * (lo + hi)
        if occupancy(p, mid) < c:
            lo = mid
        else:
            hi = mid
    return 0.5 * (lo + hi)


def che_two(p, c1, c2):
    t1 = che_t(p, c1)
    h1 = [-math.expm1(-x * t1) for x in p]
    q = [x * (1.0 - h) for x, h in zip(p, h1)]
    m = math.fsum(q)
    qn = [x / m for x in q]
    t2 = che_t(qn, c2)
    l1 = math.fsum(x * h for x, h in zip(p, h1))
    l2 = math.fsum(x * -math.expm1(-y * t2) for x, y in zip(q, qn))
    return l1, l2


def sim_two(trace, c1, c2):
    a, b = OrderedDict(), OrderedDict()
    h1 = h2 = 0
    for t, x in enumerate(trace):
        if x in a:
            a.move_to_end(x)
            if t >= WARM:
                h1 += 1
            continue
        a[x] = None
        if len(a) > c1:
            a.popitem(last=False)
        if x in b:
            b.move_to_end(x)
            if t >= WARM:
                h2 += 1
        else:
            b[x] = None
            if len(b) > c2:
                b.popitem(last=False)
    n = len(trace) - WARM
    return h1 / n, h2 / n


if __name__ == "__main__":
    p = zipf(N, A)
    rng = random.Random(20261113)
    trace = rng.choices(range(N), weights=p, k=L)
    t0 = che_t(p, TOTAL)
    single = math.fsum(x * -math.expm1(-x * t0) for x in p)
    print("Two LRU levels with {} slots in all; one cache: {:.4f}"
          .format(TOTAL, single))
    print("    C1    C2   Che L1   Che L2   Che all   sim all")
    for c1 in (250, 500, 1000, 1500):
        c2 = TOTAL - c1
        l1, l2 = che_two(p, c1, c2)
        s1, s2 = sim_two(trace, c1, c2)
        print("  {:4d}  {:4d}   {:.4f}   {:.4f}    {:.4f}    {:.4f}".format(
            c1, c2, l1, l2, l1 + l2, s1 + s2))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《缓存命中率与车氏近似、排盘结果缓存容量的最优边界母典》第 6 章「分片与分层:同样的槽位,不同的摆法」(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-327 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026327,
  author    = {Fan, Yafu},
  title     = {Cache Hit Ratios and the Che Approximation: Optimal Capacity for Caching Computed Charts},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论