性能分析工具

软件的性能是软件质量的重要考察点,不论是在线服务程序还是离线程序,甚至是终端应用,性能都是用户体验的关键。这里说的性能重大的范畴来讲包括了性能和稳定性两个方面,我们在做软件测试的时候也是要重点测试版本的性能表现和稳定性的。对于软件测试过程中发现的性能问题,如何定位有很多的方法。基本的方法可能是开发者对代码进行review,或者是使用一些工具对代码进行性能分析。常见的性能分析tuning工具有哪些呢?下面两篇文章做了详细的总结:

在我工作中主要是关于Linux C++程序代码的性能分析,gprof是可用于Linux C++代码性能profiling的工具之一,本文主要讲讲我对gprof的学习和使用过程。

Gprof的基本原理

gprof能够让你知道你的代码哪些地方是比较耗时的,哪些函数是被调用次数很多的,并且能够让你一目了然的看到函数与函数之间的调用关系。gprof是gcc/g++编译器支持的一种性能诊断工具。只要在编译时加上-pg选项,编译器就会在编译程序时在每个函数的开头加一个mcount函数调用,在每一个函数调用之前都会先调用这个mcount函数,在mcount中会保存函数的调用关系图和函数的调用时间和被调次数等信息。最终在程序退出时保存在gmon.out文件中,需要注意的是程序必须是正常退出或者通过exit调用退出,因为只要在exit()被调用时才会触发程序写gmon.out文件。

那么,gprof的使用方法主要以下三步:

  • 会用-pg参数编译程序
  • 运行程序,并正常退出
  • 查看gmon.out文件
 

Gprof使用实例

  1. #include<iostream>
  2. using namespace std;
  3. int add(int a, int b)
  4. {
  5. return a+b;
  6. }
  7. int sub(int a, int b)
  8. {
  9. return a-b;
  10. }
  11. int call ()
  12. {
  13. std::cout << add(1,2) << std::endl;
  14. std::cout << sub(2,4) << std::endl;
  15. }
  16. int main()
  17. {
  18. int a=1, b=2;
  19. cout << add(a,b) << endl;
  20. for (int i=0; i<10000; i++)
  21. call();
  22. return 0;
  23. }

使用g++编译并加上-pg参数:

  1. g++ -o hello hello_grof.cpp -pg -g

得到可执行文件,我们可以使用readelf查看一下它的符号表里和没有-pg时编译的有啥不同:readelf -r ./hello和readelf -r ./hello_normal得出的结果对比。

左边为有-pg参数编译的结果。可以看出多了三个函数符号_mcount, __monstartup, _mcleanup都是和gprof相关的调用。

 
使用gdb调试hello程序,在mcount函数中打断点也可以看到其调用关系,在add函数执行前先调用mcount函数:
 
接下来运行程序./hello,会在当前目录下生成gmon.out文件。使用gprof查看文件信息:
  1. gprof -b ./hello gmon.out

得到如下输出:

  1. Flat profile:
  2. Each sample counts as 0.01 seconds.
  3. no time accumulated
  4. %   cumulative   self              self     total
  5. time   seconds   seconds    calls  Ts/call  Ts/call  name
  6. 0.00      0.00     0.00    10001     0.00     0.00  add(int, int)
  7. 0.00      0.00     0.00    10000     0.00     0.00  sub(int, int)
  8. 0.00      0.00     0.00    10000     0.00     0.00  call()
  9. 0.00      0.00     0.00        1     0.00     0.00  global constructors keyed to _Z3addii
  10. 0.00      0.00     0.00        1     0.00     0.00  __static_initialization_and_destruction_0(int, int)
  11. Call graph
  12. granularity: each sample hit covers 2 byte(s) no time propagated
  13. index % time    self  children    called     name
  14. 0.00    0.00       1/10001       main [7]
  15. 0.00    0.00   10000/10001       call() [10]
  16. [8]      0.0    0.00    0.00   10001         add(int, int) [8]
  17. -----------------------------------------------
  18. 0.00    0.00   10000/10000       call() [10]
  19. [9]      0.0    0.00    0.00   10000         sub(int, int) [9]
  20. -----------------------------------------------
  21. 0.00    0.00   10000/10000       main [7]
  22. [10]     0.0    0.00    0.00   10000         call() [10]
  23. 0.00    0.00   10000/10001       add(int, int) [8]
  24. 0.00    0.00   10000/10000       sub(int, int) [9]
  25. -----------------------------------------------
  26. 0.00    0.00       1/1           __do_global_ctors_aux [13]
  27. [11]     0.0    0.00    0.00       1         global constructors keyed to _Z3addii [11]
  28. 0.00    0.00       1/1           __static_initialization_and_destruction_0(int, int) [12]
  29. -----------------------------------------------
  30. 0.00    0.00       1/1           global constructors keyed to _Z3addii [11]
  31. [12]     0.0    0.00    0.00       1         __static_initialization_and_destruction_0(int, int) [12]
  32. -----------------------------------------------
  33. Index by function name
  34. [11] global constructors keyed to _Z3addii (hello_grof.cpp) [9] sub(int, int) [10] call()
  35. [8] add(int, int)          [12] __static_initialization_and_destruction_0(int, int) (hello_grof.cpp)

可以使用运行命令:

  1. gprof -b ./hello  gmon.out | gprof2doc.py > ~WWW/hello.dot

生成dot格式的调用关系图文件,可以使用windows版的GVEdit for Graphviz软件查看调用关系图:

 
附上一张比较复杂的程序调用关系图:
对于调用的关系和调用热点一目了然。
 

Gprof输出解读

这部分内容可将gprof -b ./hello中的-b参数去掉,可以显示字段的详细含义描述:

  1. 14  %         the percentage of the total running time of the
  2. 15 time       program used by this function.
  3. 16
  4. 17 cumulative a running sum of the number of seconds accounted
  5. 18  seconds   for by this function and those listed above it.
  6. 19
  7. 20  self      the number of seconds accounted for by this
  8. 21 seconds    function alone.  This is the major sort for this
  9. 22            listing.
  10. 23
  11. 24 calls      the number of times this function was invoked, if
  12. 25            this function is profiled, else blank.
  13. 26
  14. 27  self      the average number of milliseconds spent in this
  15. 28 ms/call    function per call, if this function is profiled,
  16. 29        else blank.
  17. 30
  18. 31  total     the average number of milliseconds spent in this
  19. 32 ms/call    function and its descendents per call, if this
  20. 33        function is profiled, else blank.
  21. 34
  22. 35 name       the name of the function.  This is the minor sort
  23. 36            for this listing. The index shows the location of
  24. 37        the function in the gprof listing. If the index is
  25. 38        in parenthesis it shows where it would appear in
  26. 39        the gprof listing if it were to be printed.

总结

gprof是常见的性能分析工具,在此罗列一下它的一些不足,也是从网上看的:

  • 1、对多线程支持不好,不准确
  • 2、必须退出exit()才行
  • 3、它只能分析应用程序在运行过程中所消耗掉的用户时间,无法得到程序内核空间的运行时间。对内核态的调用分析无能为力。如果程序系统调用比率比较大,就不适合。

Linux C++程序进行性能分析工具gprof使用入门的更多相关文章

  1. [转]程序进行性能分析工具gprof使用入门

    性能分析工具 软件的性能是软件质量的重要考察点,不论是在线服务程序还是离线程序,甚至是终端应用,性能都是用户体验的关键.这里说的性能重大的范畴来讲包括了性能和稳定性两个方面,我们在做软件测试的时候也是 ...

  2. Linux使用sar进行性能分析

    转:https://blog.csdn.net/xusensen/article/details/54606401#sar%E7%AE%80%E4%BB%8B Linux使用sar进行性能分析 Lin ...

  3. xDebug + webgrind 对 php 程序进行性能分析

    环境 macOs Sierra php 7.0.8 MAMP Pro 集成环境 背景 最近有一个需要在微信朋友圈上线的 h5,本人做了一个抽奖的接口,也没多想,直接上 php ci(CodeIgnit ...

  4. linux环境下 C++性能测试工具 gprof + kprof + gprof2dot

    1.gprof 很有名了,google下很多教程 g++ -pg -g  -o test test.cc ./test   //会生成gmon.out gprof ./test > prof.l ...

  5. Linux下程序包管理工具RPM

    实验环境: CentOS release 6.6 (Final)  一台 IP地址:172.16.249.230 RPM 是 Red Hat Package Manager 的缩写,本意是Red Ha ...

  6. 使用xhprof对php7程序进行性能分析

    Xhprof是facebook开源出来的一个php轻量级的性能分析工具,跟Xdebug类似,但性能开销更低,还可以用在生产环境中,也可以由程序开关来控制是否进行profile. 对于还在使用php5的 ...

  7. linux后台程序开发常用工具

    linux开发工具: 1.编辑工具:1)sourceInsight2)Notepad++3)UltraEdit4)Altova XMLSpy 2.linux服务器访问工具:1)FileZilla2)X ...

  8. 实验-使用VisualVM或JConsole进行对程序进行性能分析

    参考资料: 性能分析神器VisualVM java可视化监控工具 完成下列任务: 1.分析内存堆 使用+进行频繁的字符串拼接 2.CPU性能分析 3.线程分析 编程比较以下几个方法所创建的线程 Exe ...

  9. ubuntu上编译和使用easy_profiler对C++程序进行性能分析

    本文首发于个人博客https://kezunlin.me/post/91b7cf13/,欢迎阅读最新内容! tutorial to compile and use esay profiler with ...

随机推荐

  1. 前端复制功能的若干 -- document.execCommand()

    最近涨停科技公司实习,由于backend基础太弱...强行前端了一把..搞了两周才搞下页面里copy的功能,期间有些琐碎,恐忘,记录在此. 目前copy主流有四种方式:ZeroClipboard,Cl ...

  2. [算法] dijkstra单源无负权最小路径算法

    #include <stdio.h>#include <stdlib.h>#include <string.h> #define INF 1000000#defin ...

  3. Spark+ECLIPSE+JAVA+MAVEN windows开发环境搭建及入门实例【附详细代码】

    http://blog.csdn.net/xiefu5hh/article/details/51707529 Spark+ECLIPSE+JAVA+MAVEN windows开发环境搭建及入门实例[附 ...

  4. PHP 对MySQLI预处理的包装

    mysql 类 <?php class Mysql { private static $instance; private $link; private $query; private $stm ...

  5. DMLC深度机器学习框架MXNet的编译安装

    这篇文章将介绍MXNet的编译安装. MXNet的编译安装分为两步: 首先,从C++源码编译共享库(libmxnet.so for linux,libmxnet.dylib for osx,libmx ...

  6. CentOS 5.8 x64 安装TomCat

    简单记录一下...虽然安装很简单... 首先下载配置安装 jdk http://www.oracle.com/technetwork/java/javase/downloads/jdk-6u25-do ...

  7. Laravel Eloquent get获取空的数据问题

    在用laravel框架来获取数据,若数据不存在时,以为会返回空,其实不是的,其实是一个 collection 值,会返回如下: object(Illuminate\Database\Eloquent\ ...

  8. ice grid 完整部署过程

    待补充 一 理论准备 一个IceGrid集群有一个registry(注册表,用于定位)和多个node组成. IceGrid配置包括集群配置和应用配置: config.grid是集群配置,配置Regis ...

  9. ios 添加PCH文件

  10. URL与String转换

    NSString *str = @"www.baidu.com"; NSURL *URL = [NSURL URLWithString:str];  //string>url ...