关于「串行比例与相干开销:吞吐为什么有峰值」,《并发拓扑耦合与黑板复用》给出的定量结论是什么?

结论:第一章说明峰值从何而来,示意参数下峰值在 6.78 个窗口附近。

数据口径:书中参数为示意,数据为合成,全部计算由书中程序生成;本书是非个人化经营建议,不提供任何加速比承诺。 第一章说明峰值从何而来,示意参数下峰值在 6.78 个窗口附近。

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: 6.78

A. speed-up by number of nodes, s = sigma0.08, kappa = 0.02
N Amdahl Gustafson USL USL/N
11.00 1.00 1.00 1.000
21.85 1.92 1.79 0.893
43.23 3.76 2.70 0.676
64.29 5.60 3.00 0.500
74.73 6.52 3.02 0.431
85.13 7.44 2.99 0.373
126.38 11.12 2.65 0.221
167.27 14.80 2.29 0.143
248.45 22.16 1.73 0.072
329.20 29.52 1.37 0.043
B. N*6.78; best integer N = 7, C = 3.017
Amdahl ceiling 1/s12.5; USL at N = 32 is 0.455 of its peak
C. N* and peak C over a grid (illustrative)
sigma kappa N* C(N*)
0.020.005 14.00 6.45
0.020.020 7.00 3.57
0.020.050 4.43 2.42
0.080.005 13.56 4.75
0.080.020 6.78 3.02
0.080.050 4.29 2.18
0.150.005 13.04 3.63
0.150.020 6.52 2.56
0.150.050 4.12 1.95
查看完整验算程序
# Chapter 1. p11: Amdahl, Gustafson and the universal scalability law.
# Amdahl: S(N) = 1 / (s + (1 - s) / N), a fixed job split N ways.
# Gustafson: S(N) = N - s (N - 1), the job grows with N.
# USL (Gunther): C(N) = N / (1 + sigma (N - 1) + kappa N (N - 1)),
# peak at N* = sqrt((1 - sigma) / kappa). sigma = 0.08 and
# kappa = 0.02 are illustrative values for writing windows, not data.
import math

SIGMA = 0.08
KAPPA = 0.02


def amdahl(n, s):
    return 1.0 / (s + (1.0 - s) / n)


def gustafson(n, s):
    return n - s * (n - 1)


def usl(n, sigma, kappa):
    return n / (1.0 + sigma * (n - 1) + kappa * n * (n - 1))


def peak(sigma, kappa):
    return math.sqrt((1.0 - sigma) / kappa)


if __name__ == "__main__":
    print("A. speed-up by number of nodes, s = sigma = {:.2f},"
          " kappa = {:.2f}".format(SIGMA, KAPPA))
    print("     N   Amdahl  Gustafson   USL   USL/N")
    for n in (1, 2, 4, 6, 7, 8, 12, 16, 24, 32):
        c = usl(n, SIGMA, KAPPA)
        print("  {:>4}   {:6.2f}   {:7.2f}   {:5.2f}  {:5.3f}".format(
            n, amdahl(n, SIGMA), gustafson(n, SIGMA), c, c / n))
    ns = peak(SIGMA, KAPPA)
    best = max(range(1, 65), key=lambda k: usl(k, SIGMA, KAPPA))
    print("B. N* = {:.2f}; best integer N = {}, C = {:.3f}".format(
        ns, best, usl(best, SIGMA, KAPPA)))
    print("   Amdahl ceiling 1/s = {:.1f}; USL at N = 32 is {:.3f}"
          " of its peak".format(1.0 / SIGMA,
                                usl(32, SIGMA, KAPPA)
                                / usl(best, SIGMA, KAPPA)))
    print("C. N* and peak C over a grid (illustrative)")
    print("   sigma  kappa    N*     C(N*)")
    for s in (0.02, 0.08, 0.15):
        for k in (0.005, 0.02, 0.05):
            p = peak(s, k)
            print("   {:.2f}   {:.3f}   {:5.2f}   {:5.2f}".format(
                s, k, p, usl(p, s, k)))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《多核并发操作拓扑耦合、共享黑板无锁通信与跨节点涌现增益母典》第 1 章「串行比例与相干开销:吞吐为什么有峰值」(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-411 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026411,
  author    = {Fan, Yafu},
  title     = {Topological Coupling of Concurrent Multi-Core Operations: Lock-Free Blackboard Signalling and Cross-Node Emergent Gains},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论