性能分析之valgrind及相关工具 _valgrind

【性能分析之valgrind及相关工具】要须心地收汗马，孔孟行世目杲杲。这篇文章主要讲述性能分析之valgrind及相关工具相关的知识，希望能为你提供帮助。
今天在重新整理资料的时候，发现这部分的工具没有操作性比较强的。所以就重新一个个命令行参数的熟悉一下。

其实对C/C++的应用来说，查看函数时间调用和内存消耗调用的有很多工具，这个工具算是做得比较不错的一个。建议直接用yum安装，有的说可以安装更新的版本，像我这种不喜欢各种找依赖条件的，就直接yum了。

它的使用也非常简单，比如下面这个例子，可以查看内存越界和泄露的情况。只要用valgrind调用一下被测试的程序即可。

memcheck

[root@7dgroupSample10]# vi test5.c [root@7dgroupSample10]# gcc -Wall -o test5 test5.c [root@7dgroupSample10]# valgrind --tool=memcheck --leak-check=full ./test5 ==318==Memcheck, a memory error detector ==318==Copyright (C) 2002-2013, and GNU GPLd, by Julian Seward et al. ==318==Using Valgrind-3.10.0 and LibVEX; rerun with -h for copyright info ==318==Command: ./test5 ==318== ==318== Invalid write of size 4 //内存越界 ==318== at 0x40054E: f (in/root/GDB/Sample10/test5) ==318== by 0x40055E: main (in/root/GDB/Sample10/test5) ==318== Address 0x51f7068 is 0bytes after a block of size 40 allocd ==318== at 0x4C29BFD: malloc(in /usr/lib64/valgrind/vgpreload_memcheck-amd64-linux.so) ==318== by 0x400541: f (in/root/GDB/Sample10/test5) ==318== by 0x40055E: main (in/root/GDB/Sample10/test5) ==318== ==318== ==318==HEAP SUMMARY: ==318== in use at exit: 40 bytes in 1 blocks ==318== total heap usage: 1 allocs, 0 frees, 40bytes allocated ==318== ==318== 40 bytes in 1 blocks are definitely lost in loss record 1 of 1 //内存泄露 ==318== at 0x4C29BFD: malloc(in /usr/lib64/valgrind/vgpreload_memcheck-amd64-linux.so) ==318== by 0x400541: f (in/root/GDB/Sample10/test5) ==318== by 0x40055E: main (in/root/GDB/Sample10/test5) ==318== ==318==LEAK SUMMARY: ==318== definitely lost: 40 bytes in 1 blocks ==318== indirectly lost: 0 bytes in 0 blocks ==318== possibly lost: 0 bytes in 0 blocks ==318== still reachable: 0 bytes in 0 blocks ==318== suppressed: 0 bytes in 0 blocks ==318== ==318==For counts of detected and suppressed errors, rerun with: -v ==318==ERROR SUMMARY: 2 errors from 2 contexts (suppressed: 1 from 1) [root@7dgroupSample10]#

再看一下它的其他工具：
cachegrind： Cache分析器，可以查看引用次数，命中率等信息。

[root@7dgroupSample10]# valgrind --tool=cachegrind ./test5 ==1655==Cachegrind, a cache and branch-prediction profiler ==1655==Copyright (C) 2002-2013, and GNU GPLd, by Nicholas Nethercote et al. ==1655==Using Valgrind-3.10.0 and LibVEX; rerun with -h for copyright info ==1655==Command: ./test5 ==1655== --1655--warning: L3 cache found, using its data for the LL simulation. --1655--warning: specified LL cache: line_size 64 assoc 20 total_size 31,457,280 --1655--warning: simulated LL cache: line_size 64 assoc 30 total_size 31,457,280 ==1655== ==1655== I refs: 152,928 ==1655== I1 misses: 765 ==1655== LLi misses: 759 ==1655== I1 miss rate: 0.50% ==1655== LLi miss rate: 0.49% ==1655== ==1655== D refs: 51,260 (39,547 rd + 11,713 wr) ==1655== D1 misses: 2,861 ( 2,321 rd + 540 wr) ==1655== LLd misses: 2,384 ( 1,891 rd + 493 wr) ==1655== D1 miss rate: 5.5% ( 5.8% + 4.6% ) ==1655== LLd miss rate: 4.6%( 4.7% + 4.2% ) ==1655== ==1655== LL refs: 3,626 ( 3,086 rd + 540 wr) ==1655== LL misses: 3,143 ( 2,650 rd + 493 wr) ==1655== LL miss rate: 1.5%( 1.3% + 4.2% ) [root@7dgroupSample10]#

callgrind: 查看函数调用关系及时间消耗

[root@7dgroupSample10]# valgrind --tool=callgrind /usr/sbin/nginx ==1713==Callgrind, a call-graph generating cache profiler ==1713==Copyright (C) 2002-2013, and GNU GPLd, by Josef Weidendorfer et al. ==1713==Using Valgrind-3.10.0 and LibVEX; rerun with -h for copyright info ==1713==Command: /usr/sbin/nginx ==1713== ==1713==For interactive control, run callgrind_control -h. ==1713== ==1713==Events : Ir ==1713==Collected : 8916001 ==1713== ==1713==I refs: 8,916,001 [root@7dgroupSample10]# ll 总用量 884 -rw-------.1 root root 5047 4月 17 21:12 cachegrind.out.1655 -rw-------.1 root root 129128 4月 17 21:15callgrind.out.1702 -rw-------. 1 root root 128813 4月 17 21:16 callgrind.out.1713

这个结果可以生成调用图。如下所示：

还可以用kcachegrind（有linux和windows版本）打开out文件。如下所示：

在这个图的下面还有一个百分比图例，直接双击可以一层层drill down下去，直到一个具体的函数。也可以直接看到相应的源码（前提是有ELF文件）。

这个工具有它的局限性，就是它不能attach进程。如果要查看运行中的进程的内存消耗和函数消耗情况，可以使用其他工具。有感兴趣的以后我再写如何查看运行中的C/C++剖析工具。
这是个对C/C++进程的工具。java的工具就更多了，以后再整理。

工具用好了事半功倍，但是前提是理解原理。