为什么C++中的STDIN读行比Python慢得多?

Why is reading lines from stdin much slower in C++ than Python?

我想用Python和C++来比较STDIN中的字符串输入的读取行,并震惊地看到我的C++代码运行速度比等效的Python代码慢了一个数量级。因为我的C++是生疏的,我还不是一个专家,请告诉我,如果我做错了什么,或者如果我误解了什么。

(TLDR回答:包括声明:cin.sync_with_stdio(false)或仅使用fgets代替。

TLDR结果:一直向下滚动到问题的底部,然后查看表格。)

C++代码:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
#include <iostream>
#include <time.h>

using namespace std;

int main() {
    string input_line;
    long line_count = 0;
    time_t start = time(NULL);
    int sec;
    int lps;

    while (cin) {
        getline(cin, input_line);
        if (!cin.eof())
            line_count++;
    };

    sec = (int) time(NULL) - start;
    cerr <<"Read" << line_count <<" lines in" << sec <<" seconds.";
    if (sec > 0) {
        lps = line_count / sec;
        cerr <<" LPS:" << lps << endl;
    } else
        cerr << endl;
    return 0;
}

// Compiled with:
// g++ -O3 -o readline_test_cpp foo.cpp

python等价物:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
#!/usr/bin/env python
import time
import sys

count = 0
start = time.time()

for line in  sys.stdin:
    count += 1

delta_sec = int(time.time() - start_time)
if delta_sec >= 0:
    lines_per_sec = int(round(count/delta_sec))
    print("Read {0} lines in {1} seconds. LPS: {2}".format(count, delta_sec,
       lines_per_sec))

以下是我的结果:

1
2
3
4
5
$ cat test_lines | ./readline_test_cpp
Read 5570000 lines in 9 seconds. LPS: 618889

$cat test_lines | ./readline_test.py
Read 5570000 lines in 1 seconds. LPS: 5570000

我需要注意的是,我在Mac OS X v10.6.8(雪豹)和Linux 2.6.32(Red Hat Linux 6.2)下都尝试过。前者是一个MacBookPro,后者是一个非常强大的服务器,而不是说这太中肯了。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
$ for i in {1..5}; do echo"Test run $i at `date`"; echo -n"CPP:"; cat test_lines | ./readline_test_cpp ; echo -n"Python:"; cat test_lines | ./readline_test.py ; done
Test run 1 at Mon Feb 20 21:29:28 EST 2012
CPP:   Read 5570001 lines in 9 seconds. LPS: 618889
Python:Read 5570000 lines in 1 seconds. LPS: 5570000
Test run 2 at Mon Feb 20 21:29:39 EST 2012
CPP:   Read 5570001 lines in 9 seconds. LPS: 618889
Python:Read 5570000 lines in 1 seconds. LPS: 5570000
Test run 3 at Mon Feb 20 21:29:50 EST 2012
CPP:   Read 5570001 lines in 9 seconds. LPS: 618889
Python:Read 5570000 lines in 1 seconds. LPS: 5570000
Test run 4 at Mon Feb 20 21:30:01 EST 2012
CPP:   Read 5570001 lines in 9 seconds. LPS: 618889
Python:Read 5570000 lines in 1 seconds. LPS: 5570000
Test run 5 at Mon Feb 20 21:30:11 EST 2012
CPP:   Read 5570001 lines in 10 seconds. LPS: 557000
Python:Read 5570000 lines in  1 seconds. LPS: 5570000

微小基准附录和摘要

为了完整起见,我想我会用原来的(同步)C++代码更新同一个文件上的读取速度。同样,这是用于快速磁盘上的100M行文件。以下是与几种解决方案/方法的比较:

1
2
3
4
5
6
Implementation      Lines per second
python (default)           3,571,428
cin (default/naive)          819,672
cin (no sync)             12,500,000
fgets                     14,285,714
wc (not fair comparison)  54,644,808


默认情况下,cin与stdio同步,从而避免任何输入缓冲。如果将此项添加到主项的顶部,您将看到更好的性能:

1
std::ios_base::sync_with_stdio(false);

通常,当一个输入流被缓冲时,而不是一次读取一个字符,该流将以更大的块读取。这减少了系统调用的数量,而系统调用通常比较昂贵。然而,由于基于FILE*stdioiostreams通常具有单独的实现,因此具有单独的缓冲区,如果两者一起使用,这可能会导致问题。例如:

1
2
3
4
int myvalue1;
cin >> myvalue1;
int myvalue2;
scanf("%d",&myvalue2);

如果cin读取的输入比实际需要的多,那么第二个整数值将不适用于scanf函数,该函数有自己的独立缓冲区。这将导致意想不到的结果。

为了避免这种情况,默认情况下,流与stdio同步。实现这一点的一种常见方法是让cin根据需要使用stdio函数一次读取每个字符。不幸的是,这会带来很多开销。对于少量的输入,这不是一个大问题,但是当您阅读数百万行时,性能损失是显著的。

幸运的是,库设计人员决定,如果您知道自己在做什么,也应该能够禁用此功能以获得更好的性能,因此他们提供了sync_with_stdio方法。


出于好奇,我看了看引擎盖下面发生了什么,并在每个测试中使用了DTruss/Strace。

C++

1
2
./a.out < in
Saw 6512403 lines in 8 seconds.  Crunch speed: 814050

系统调用sudo dtruss -c ./a.out < in

1
2
3
4
5
6
7
8
9
CALL                                        COUNT
__mac_syscall                                   1
<snip>
open                                            6
pread                                           8
mprotect                                       17
mmap                                           22
stat64                                         30
read_nocancel                               25958

Python

1
2
./a.py < in
Read 6512402 lines in 1 seconds. LPS: 6512402

系统调用sudo dtruss -c ./a.py < in

1
2
3
4
5
6
7
8
CALL                                        COUNT
__mac_syscall                                   1
<snip>
open                                            5
pread                                           8
mprotect                                       17
mmap                                           21
stat64                                         29


我落后了几年,但是:好的。

在原始日志的"编辑4/5/6"中,您使用的是结构:好的。

1
$ /usr/bin/time cat big_file | program_to_benchmark

这在两个不同的方面是错误的:好的。

  • 实际上,您是在计时"cat"的执行,而不是您的基准。"time"显示的"user"和"sys"CPU使用率是"cat"的使用率,而不是基准程序的使用率。更糟糕的是,"实时"时间也不一定准确。根据"cat"和本地操作系统中管道的实现情况,"cat"可能会在读卡器进程完成工作之前编写最终的巨大缓冲区并退出。好的。

  • 使用"cat"是不必要的,实际上会适得其反;你在添加运动部件。如果您使用的是一个足够旧的系统(即只有一个CPU,在某些代的计算机中,I/O速度比CPU快),那么仅仅运行"cat"这个事实就可以大大改变结果。您还需要服从输入和输出缓冲以及其他处理'cat'。(如果我是Randal Schwartz,这很可能会给你赢得"无用的猫使用"奖。好的。

  • 更好的结构是:好的。

    1
    $ /usr/bin/time program_to_benchmark < big_file

    在这个语句中,它是一个shell,它打开一个大的文件,并将其作为一个已经打开的文件描述符传递给您的程序(实际上是传递给"time",然后作为子进程执行您的程序)。100%的文件读取是您要测试的程序的严格责任。这可以让您真正了解它的性能,而不会出现虚假的复杂情况。好的。

    我将提到两个可能的,但实际上是错误的,"修复"也可以考虑(但我用不同的方式对它们进行"编号",因为这些不是原始帖子中出错的内容):好的。

    a.您可以通过只对程序进行计时来"修复"此问题:好的。

    1
    $ cat big_file | /usr/bin/time program_to_benchmark

    b.或通过对整个管道进行计时:好的。

    1
    $ /usr/bin/time sh -c 'cat big_file | program_to_benchmark'

    这些错误的原因与2相同:他们仍然不必要地使用"cat"。我提到它们有几个原因:好的。

    • 对于那些对posix shell的I/O重定向功能不太满意的人来说,它们更为"自然"。好的。

    • 在某些情况下,可能需要'cat'(例如:要读取的文件需要某种访问权限,而您不想将该权限授予要基准化的程序:'sudo cat/dev/sda/usr/bin/time my_compression_test--no output`)好的。

    • 在实践中,在现代机器上,管道中添加的"cat"可能没有实际意义。好的。

    但我迟疑地说了最后一句话。如果我们在"编辑5"中检查最后一个结果--好的。

    1
    2
    $ /usr/bin/time cat temp_big_file | wc -l
    0.01user 1.34system 0:01.83elapsed 74%CPU ...

    --这表明"cat"在测试期间消耗了74%的CPU;实际上,1.34/1.83大约是74%。也许是一段:好的。

    1
    $ /usr/bin/time wc -l < temp_big_file

    只剩下0.49秒!可能不是这样:"cat"必须支付read()系统调用(或等效调用)的费用,该调用从"disk"(实际上是缓冲区缓存)传输文件,以及将文件传输到"wc"的管道写入。正确的测试仍然需要执行那些read()调用;只有对pipe的写操作和从pipe调用中读取的操作才能被保存,而且这些操作应该非常便宜。好的。

    不过,我预测你可以测量'cat file_wc-l'和'wc-l事实上,我在Linux 3.13(Ubuntu 14.04)系统上对一个1.5GB的垃圾文件进行了一些快速测试,获得了这些结果(这些实际上是"3中最好的"结果;当然,在启动缓存之后):好的。

    1
    2
    3
    4
    5
    6
    $ time wc -l < /tmp/junk
    real 0.280s user 0.156s sys 0.124s (total cpu 0.280s)
    $ time cat /tmp/junk | wc -l
    real 0.407s user 0.157s sys 0.618s (total cpu 0.775s)
    $ time sh -c 'cat /tmp/junk | wc -l'
    real 0.411s user 0.118s sys 0.660s (total cpu 0.778s)

    注意,这两个管道结果声称花费的CPU时间(user+sys)比实时时间多。这是因为我使用的是shell(bash)的内置"time"命令,它了解管道;我在一台多核机器上,管道中的独立进程可以使用独立的内核,从而比实时更快地累积CPU时间。使用/usr/bin/time,我发现CPU时间比实时时间要短——这表明它只能在命令行上对传递给它的单个管道元素计时。此外,shell的输出给出毫秒,而/usr/bin/time只给出百分之一秒。好的。

    因此,在"wc-l"的效率水平上,"cat"有着巨大的区别:409/283=1.453或45.3%的实时性,775/280=2.768,或高达177%的CPU使用率!随机地,它就在时间测试箱里。好的。

    我应该补充一点,在这些测试风格之间,至少还有一个显著的区别,我不能说这是有益的还是错误的;您必须自己决定:好的。

    当你运行"cat big_file/usr/bin/time my_program"时,你的程序正以"cat"发送的速度从管道接收输入,并以不大于"cat"写入的块的形式接收输入。好的。

    当您运行`/usr/bin/time my_program当然,如果同一个程序在两种情况下的执行方式明显不同,这是一个有趣的基准测试结果。它表明,实际上,程序或其I/O库正在做一些有趣的事情,比如使用mmap()。因此,在实践中,两种方式都运行基准可能是好的;也许通过一些小因素对"cat"结果进行折扣,以"原谅"运行"cat"本身的成本。好的。好啊。


    我在我的电脑上用Mac电脑上的G++复制了原始结果。

    在EDCOX1的12个循环之前,将下面的语句添加到C++版本中,使其与Python版本内联:

    1
    2
    3
    std::ios_base::sync_with_stdio(false);
    char buffer[1048576];
    std::cin.rdbuf()->pubsetbuf(buffer, sizeof(buffer));

    同步与stdio将速度提高到2秒,设置更大的缓冲区将速度降低到1秒。


    如果您不关心文件加载时间或加载小文本文件,那么getline、流运算符scanf非常方便。但是,如果性能是您关心的,那么您应该将整个文件缓冲到内存中(假设它适合)。

    下面是一个例子:

    1
    2
    3
    4
    5
    6
    7
    8
    9
    10
    11
    12
    13
    14
    //open file in binary mode
    std::fstream file( filename, std::ios::in|::std::ios::binary );
    if( !file ) return NULL;

    //read the size...
    file.seekg(0, std::ios::end);
    size_t length = (size_t)file.tellg();
    file.seekg(0, std::ios::beg);

    //read into memory buffer, then close it.
    char *filebuf = new char[length+1];
    file.read(filebuf, length);
    filebuf[length] = '\\0'; //make it null-terminated
    file.close();

    如果需要,可以将流环绕在该缓冲区周围,以便更方便地访问,如下所示:

    1
    std::istrstream header(&filebuf[0], length);

    此外,如果您控制文件,请考虑使用平面二进制数据格式而不是文本。读写更可靠,因为你不必处理空白区的所有含糊不清之处。它也更小,解析速度更快。


    顺便说一下,C++版本的行计数比Python版本的计数大的原因是,当试图读取EOF时,EOF标记只被设置。所以正确的循环应该是:

    1
    2
    3
    4
    5
    6
    while (cin) {
        getline(cin, input_line);

        if (!cin.eof())
            line_count++;
    };


    以下代码对我来说比这里发布的其他代码快:(Visual Studio 2013,64位500 MB文件,行长一致,格式为[0,1000])。

    1
    2
    3
    4
    5
    6
    7
    const int buffer_size = 500 * 1024;  // Too large/small buffer is not good.
    std::vector<char> buffer(buffer_size);
    int size;
    while ((size = fread(buffer.data(), sizeof(char), buffer_size, stdin)) > 0) {
        line_count += count_if(buffer.begin(), buffer.begin() + size, [](char ch) { return ch == '\
    '
    ; });
    }

    它比我所有的python尝试都快了一倍多。


    在第二个示例中(使用scanf())原因可能是scanf("%s")解析字符串并查找任何空格字符(space、tab、newline)。

    另外,是的,cpython做了一些缓存以避免硬盘读取。


    答案的第一个要素是:很慢。该死的慢。我使用scanf获得了巨大的性能提升,如下所示,但它仍然比python慢两倍。

    1
    2
    3
    4
    5
    6
    7
    8
    9
    10
    11
    12
    13
    14
    15
    16
    17
    18
    19
    20
    21
    22
    23
    24
    25
    26
    27
    28
    29
    #include <iostream>
    #include <time.h>
    #include <cstdio>

    using namespace std;

    int main() {
        char buffer[10000];
        long line_count = 0;
        time_t start = time(NULL);
        int sec;
        int lps;

        int read = 1;
        while(read > 0) {
            read = scanf("%s", buffer);
            line_count++;
        };
        sec = (int) time(NULL) - start;
        line_count--;
        cerr <<"Saw" << line_count <<" lines in" << sec <<" seconds." ;
        if (sec > 0) {
            lps = line_count / sec;
            cerr <<"  Crunch speed:" << lps << endl;
        }
        else
            cerr << endl;
        return 0;
    }


    好吧,我看到在你的第二个解决方案中,你从cin转到scanf,这是我第一个建议(cin是slooooooooooooow)。现在,如果您从EDCOX1的1度切换到EDCOX1(16),您将看到性能的另一个提升:EDCOX1×16Ω是用于字符串输入的最快的C++函数。

    顺便说一句,不知道同步的事,很好。但你还是应该试试fgets