关于「只追加黑板:读路径无锁、哈希去重与订阅成本」,《并发拓扑耦合与黑板复用》给出的定量结论是什么?

结论:第四章说明只追加黑板为什么读路径不必取锁,示意参数下最优订阅比例约 0.7。

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: 0.7

A. coupling sweep, 2 min per new note read
c avoided failures read min net min/node-week
0.00 1,560 0 0.0
0.1648 912 1,944 112.2
0.2810 750 3,216 135.2
0.3921 639 4,404 148.9
0.4964 596 5,460 150.4
0.51,014 546 6,468 153.5
0.61,063 497 7,452 156.7
0.71,094 466 8,376 156.7
0.81,128 432 9,420 156.5
0.91,148 412 10,224 155.2
1.01,158 402 11,064 151.8
B. best c0.7, net 156.7 min per node-week
C. best c when reading costs more or less
1 min per note: best c0.8, net 170.7
2 min per note: best c0.7, net 156.7
4 min per note: best c0.6, net 132.9
8 min per note: best c0.3, net 100.2
查看完整验算程序
# Chapter 4. p42: reuse gain saturates, read cost grows, net gain peaks.
# 6 nodes, 26 weeks, 40 topics with Zipf-like relevance (exponent
# 0.5), 400 distinct pitfalls spread over topics by relevance. Each
# node meets 10 pitfall draws a week. Coupling c = subscribed topics /
# all topics; a node subscribes to the most relevant topics first. A
# draw is avoided if the pitfall is already on the board and its topic
# is subscribed; otherwise the node fails once and appends a note.
# Each week a node pays 1 min per subscribed topic plus 2 min for every
# new note in those topics; a failure costs 30 min. All illustrative.
import math
import random

NODES, WEEKS, TOPICS, PITS, DRAWS = 6, 26, 40, 400, 10
FAIL_MIN, TOPIC_MIN = 30.0, 1.0


def setup(rng):
    w = [1.0 / (k + 1) ** 0.5 for k in range(TOPICS)]
    topic_of = [rng.choices(range(TOPICS), w)[0] for _ in range(PITS)]
    pw = [w[t] for t in topic_of]
    draws = [[[rng.choices(range(PITS), pw)[0] for _ in range(DRAWS)]
              for _ in range(NODES)] for _ in range(WEEKS)]
    return topic_of, draws


def run(c, topic_of, draws, entry_min):
    subs = set(range(int(round(c * TOPICS))))
    board, avoided, fails, reads = set(), 0, 0, []
    for week in draws:
        new = []
        for node in week:
            for pit in node:
                if pit in board and topic_of[pit] in subs:
                    avoided += 1
                else:
                    fails += 1
                    if pit not in board:
                        new.append(pit)
        board.update(new)
        seen = 0
        for pit in new:
            if topic_of[pit] in subs:
                seen += 1
        for _ in range(NODES):
            reads.append(TOPIC_MIN * len(subs) + entry_min * seen)
    read_min = math.fsum(reads)
    net = avoided * FAIL_MIN - read_min
    return avoided, fails, read_min, net / (NODES * WEEKS)


if __name__ == "__main__":
    topic_of, draws = setup(random.Random(41142))
    grid = [k / 10 for k in range(11)]
    print("A. coupling sweep, 2 min per new note read")
    print("     c   avoided  failures  read min  net min/node-week")
    best = None
    for c in grid:
        a, f, r, net = run(c, topic_of, draws, 2.0)
        print("   {:.1f}  {:>7,}  {:>8,}  {:>8,.0f}  {:>17.1f}".format(
            c, a, f, r, net))
        if best is None or net > best[1]:
            best = (c, net)
    print("B. best c = {:.1f}, net {:.1f} min per node-week".format(*best))
    print("C. best c when reading costs more or less")
    for em in (1.0, 2.0, 4.0, 8.0):
        res = [(run(c, topic_of, draws, em)[3], c) for c in grid]
        top = max(res)
        print("   {:.0f} min per note: best c = {:.1f}, net {:.1f}".format(
            em, top[1], top[0]))

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 4 章「只追加黑板:读路径无锁、哈希去重与订阅成本」(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