std::shared_ptr vs std::make_shared:意外的缓存未命中和分支预测

std::shared_ptr vs std::make_shared: unexpected cache misses and branch prediction

我正在尝试测量 std::shared_ptrstd::make_shared 创建的指针的效率。

我有下一个测试代码:

#include <iostream>
#include <memory>
#include <vector>


struct TestClass {
    TestClass(int _i) : i(_i) {}
    int i = 1;
};

void sum(const std::vector<std::shared_ptr<TestClass>>& v) {
    unsigned long long s = 0u;
    for(size_t i = 0; i < v.size() - 1; ++i) {
        s += v[i]->i * v[i + 1]->i;
    }
    std::cout << s << '\n';
}

void test_shared_ptr(size_t n) {
    std::cout << __FUNCTION__ << "\n";
    std::vector<std::shared_ptr<TestClass>> v;
    v.reserve(n);
    for(size_t i = 0u; i < n; ++i) {
        v.push_back(std::shared_ptr<TestClass>(new TestClass(i)));
    }
    sum(v);
}

void test_make_shared(size_t n) {
    std::cout << __FUNCTION__ << "\n";
    std::vector<std::shared_ptr<TestClass>> v;
    v.reserve(n);
    for(size_t i = 0u; i < n; ++i) {
        v.push_back(std::make_shared<TestClass>(i));
    }
    sum(v);
}

int main(int argc, char *argv[]) {
    size_t n = (argc == 3 ) ? atoi(argv[2]) : 100;
    if(atoi(argv[1]) == 1) {
        test_shared_ptr(n);
    } else {
        test_make_shared(n);
    }
    return 0;
}

编译为g++ -W -Wall -O2 -g -std=c++14 main.cpp -o cache_misses.bin

我 运行 使用 std::shared_ptr 构造函数进行测试并使用 valgrind 检查结果:

valgrind --tool=cachegrind --branch-sim=yes ./cache_misses.bin 1 100000
==2005== Cachegrind, a cache and branch-prediction profiler
==2005== Copyright (C) 2002-2017, and GNU GPL'd, by Nicholas Nethercote et al.
==2005== Using Valgrind-3.13.0 and LibVEX; rerun with -h for copyright info
==2005== Command: ./cache_misses.bin 1 100000
==2005==
--2005-- warning: L3 cache found, using its data for the LL simulation.
--2005-- warning: specified LL cache: line_size 64  assoc 12  total_size 9,437,184
--2005-- warning: simulated LL cache: line_size 64  assoc 18  total_size 9,437,184
test_shared_ptr
18107093611968
==2005==
==2005== I   refs:      74,188,102
==2005== I1  misses:         1,806
==2005== LLi misses:         1,696
==2005== I1  miss rate:       0.00%
==2005== LLi miss rate:       0.00%
==2005==
==2005== D   refs:      26,099,141  (15,735,722 rd   + 10,363,419 wr)
==2005== D1  misses:       392,064  (   264,583 rd   +    127,481 wr)
==2005== LLd misses:       134,416  (     7,947 rd   +    126,469 wr)
==2005== D1  miss rate:        1.5% (       1.7%     +        1.2%  )
==2005== LLd miss rate:        0.5% (       0.1%     +        1.2%  )
==2005==
==2005== LL refs:          393,870  (   266,389 rd   +    127,481 wr)
==2005== LL misses:        136,112  (     9,643 rd   +    126,469 wr)
==2005== LL miss rate:         0.1% (       0.0%     +        1.2%  )
==2005==
==2005== Branches:      12,732,402  (11,526,905 cond +  1,205,497 ind)
==2005== Mispredicts:       16,055  (    15,481 cond +        574 ind)
==2005== Mispred rate:         0.1% (       0.1%     +        0.0%   )

std::make_shared:

valgrind --tool=cachegrind --branch-sim=yes ./cache_misses.bin 2 100000
==2014== Cachegrind, a cache and branch-prediction profiler
==2014== Copyright (C) 2002-2017, and GNU GPL'd, by Nicholas Nethercote et al.
==2014== Using Valgrind-3.13.0 and LibVEX; rerun with -h for copyright info
==2014== Command: ./cache_misses.bin 2 100000
==2014==
--2014-- warning: L3 cache found, using its data for the LL simulation.
--2014-- warning: specified LL cache: line_size 64  assoc 12  total_size 9,437,184
--2014-- warning: simulated LL cache: line_size 64  assoc 18  total_size 9,437,184
test_make_shared
18107093611968
==2014==
==2014== I   refs:      41,283,983
==2014== I1  misses:         1,805
==2014== LLi misses:         1,696
==2014== I1  miss rate:       0.00%
==2014== LLi miss rate:       0.00%
==2014==
==2014== D   refs:      14,997,474  (8,834,690 rd   + 6,162,784 wr)
==2014== D1  misses:       241,781  (  164,368 rd   +    77,413 wr)
==2014== LLd misses:        84,413  (    7,943 rd   +    76,470 wr)
==2014== D1  miss rate:        1.6% (      1.9%     +       1.3%  )
==2014== LLd miss rate:        0.6% (      0.1%     +       1.2%  )
==2014==
==2014== LL refs:          243,586  (  166,173 rd   +    77,413 wr)
==2014== LL misses:         86,109  (    9,639 rd   +    76,470 wr)
==2014== LL miss rate:         0.2% (      0.0%     +       1.2%  )
==2014==
==2014== Branches:       7,031,695  (6,426,222 cond +   605,473 ind)
==2014== Mispredicts:      216,010  (   15,442 cond +   200,568 ind)
==2014== Mispred rate:         3.1% (      0.2%     +      33.1%   )

您可能会看到,当我使用 std::make_shared 时,缓存未命中率和分支预测错误率更高。 我希望 std::make_shared 更有效,因为存储的对象和控制块都位于同一个内存块中。或者至少性能应该是一样的。

我错过了什么?

环境详细信息:

$ g++ --version
g++ (Ubuntu 7.4.0-1ubuntu1~18.04.1) 7.4.0
Copyright (C) 2017 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

cachegrind不就是模拟,不是测量吗? https://valgrind.org/docs/manual/cg-manual.html#branch-sim条件分支是使用 16384 个 2 位饱和计数器的数组预测的。 并且它应该代表 2004 年的典型 desktop/server。

使用 2 位饱和计数器的简单分支预测在现代标准下是一个笑话,即使在 2004 年 CPUs 中,对于高性能 CPUs 也过于简单;根据 https://danluu.com/branch-prediction/. See also https://agner.org/optimize/,Pentium II/III 有一个 2 级自适应 local/global 预测器,每个本地历史条目有 4 位; Agner 的微架构 PDF 在开头有一章是关于分支预测的。

Intel 自 Haswell 使用 IT-TAGE 以来,现代 AMD 也使用先进的分支预测技术。

如果你有几个分支在 valgrind 的模拟中碰巧彼此别名,我不会感到惊讶,导致对运行频率较低的分支的错误预测。


您是否尝试过使用真正的硬件性能计数器?例如在 Linux:
perf stat -d ./cache_misses.bin 2 100000 应该为您提供更真实的真实硬件画面,包括真实的 L1d 未命中率和分支预测未命中率。 branchesbranch-misses 等性能事件映射到某些特定的硬件计数器,具体取决于 CPU 微体系结构。 perf list 将显示可用的计数器。

我经常在我的 Skylake CPU 上使用 taskset -c 3 perf stat -etask-clock:u,context-switches,cpu-migrations,page-faults,cycles:u,instructions:u,branches:u,branch-misses:u,uops_issued.any:u,uops_executed.thread:u -r 2 ./program_under_test

(实际上我通常会忽略分支未命中,因为我经常调整没有不可预测分支的 SIMD 循环,并且可以编程为计算不同事件的硬件计数器数量有限。)