Google CPU Profiler使用指南及小工具

转自：http://www.searchtb.com/2012/12/google-cpu-profiler.html

(一) gperftools-2.0 编译
wget https://gperftools.googlecode.com/files/gperftools-2.0.tar.gz
tar –xzf gperftools-2.0.tar.gz

cd gperftools-2.0

./configure –prefix=/home/wuzhu/tools/gperftools –enable-frame-pointers

make && make install

注：

编译时打开了 –enable-frame-pointers ，这要求被测试的程序在编译时要加上gcc编译选项，否则某些多线程程序可能会 core:

CCFLAGS=-fno-omit-frame-pointer

(二) google cpu profiler 基本使用

嵌入 google cpu profiler 代码与编译

1) 在我们要测试的程序源码中先 include 头文件，然后在要 profile 的代码前后加上ProfilerStart() 和 ProfilerStop()

其接口声明如下：

2) 在编译时加上 -fno-omit-frame-pointer 和 链接库 -ltcmalloc_and_profiler

CCFLAGS=-fno-omit-frame-pointer

-ltcmalloc_and_profiler

3) 执行程序，生成 profile 数据文件

4) 分析生成的数据文件

% pprof /bin/ls ls.prof

Enters “interactive” mode

% pprof –text /bin/ls ls.prof

Outputs one line per procedure

% pprof –gv /bin/ls ls.prof

Displays annotated call-graph via ‘gv’

% pprof –gv –focus=Mutex /bin/ls ls.prof

Restricts to code paths including a .*Mutex.* entry

% pprof –gv –focus=Mutex –ignore=string /bin/ls ls.prof

Code paths including Mutex but not string

% pprof –list=getdir /bin/ls ls.prof

(Per-line) annotated source listing for getdir()

% pprof –disasm=getdir /bin/ls ls.prof

(Per-PC) annotated disassembly for getdir()

% pprof –text localhost:1234

Outputs one line per procedure for localhost:1234

% pprof –callgrind /bin/ls ls.prof

Outputs the call information in callgrind format

示例一：

1) cpu_profiler_example.cpp

2) Makefile

GPerfTools=/home/wuzhu/tools/gperftools

CCFLAGS=-fno-omit-frame-pointer -g

ALL_BINS=cpu_profiler_example

all:$(ALL_BINS)

cpu_profiler_example :cpu_profiler_example.o

g++ $(CCFLAGS) -o $@ $^ -L./ -L$(GPerfTools)/lib -Wl,-Bdynamic -ltcmalloc_and_profiler

.cpp.o:

g++ $(CCFLAGS) -c -I./ -I$(GPerfTools)/include -fPIC -o $@ $<

clean:

rm -f $(ALL_BINS) *.o

3) 执行

[wuzhu@search041142.sqa.cm4 cpu_profiler]$ ./cpu_profiler_example

loopop: 255

PROFILE: interrupts/evictions/bytes = 5848/3861/185584

会产生性能数据文件: cpu_profiler_example_29502.prof

4). 分析性能数据

pprof –text cpu_profiler_example cpu_profiler_example_29502.prof

Using local file cpu_profiler_example.

Using local file cpu_profiler_example_29502.prof.

Removing killpg from all stack traces.

Removing main from all stack traces.

Removing __libc_start_main from all stack traces.

Total: 5962 samples

5894 98.9% 98.9% 5894 98.9% loopop

38 0.6% 99.5% 38 0.6% memcpy

28 0.5% 100.0% 28 0.5% memset

2 0.0% 100.0% 2 0.0% _init

注：

更详细的使用方法请见 google cpu profiler wiki:
http://gperftools.googlecode.com/svn/trunk/doc/cpuprofile.html
(三) Google CPU Profiler支持对动态链接库进行性能分析

下面，通过示例2来说明 Google Cpu Profiler 是支持对动态链接库进行性能分析的。

1) 源代码

2) Makefile

GPerfTools=/home/wuzhu/tools/gperftools

CCFLAGS=-fno-omit-frame-pointer -g

ALL_BINS=libTestProfiler.so main_dynamic_link

all:$(ALL_BINS)

main_dynamic_link :main_dynamic_link.o

g++ $(CCFLAGS) -o $@ $^ -L./ -L$(GPerfTools)/lib -Wl,-Bdynamic -lTestProfiler -ltcmalloc_and_profiler

libTestProfiler.so:TestProfiler.o

g++ -shared $(CCFLAGS) -o $@ $^

.cpp.o:

g++ $(CCFLAGS) -c -I./ -I$(GPerfTools)/include -fPIC -o $@ $<

clean:

rm -f $(ALL_BINS) *.o *.prof

3) 执行

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:`pwd`

./main_dynamic_link

result: 127

PROFILE: interrupts/evictions/bytes = 5991/3915/188368

产生性能数据文件: main_dynamic_link_24327.prof

4) 分析性能数据

pprof –text main_dynamic_link main_dynamic_link_24327.prof

Using local file main_dynamic_link.

Using local file main_dynamic_link_24327.prof.

Removing killpg from all stack traces.

Removing main from all stack traces.

Removing __libc_start_main from all stack traces.

Total: 5991 samples

5899 98.5% 98.5% 5899 98.5% loopop

53 0.9% 99.3% 53 0.9% memcpy

38 0.6% 100.0% 38 0.6% memset

1 0.0% 100.0% 1 0.0% _init

由此证明，Google CPU Profiler支持对动态链接库的性能分析

(四) Google CPU Profiler 对用 dlopen 方式打开动态库的程序支持

运行时加载允许程序可以有选择地调用库中的函数。使用动态加载过程，程序可以先加载一个特定的库（已加载则不必），然后调用该库中的某一特定函数，这是构建支持插件的应用程序的一个普遍的方法。

示例 3 主要演示 Google CPU Profiler 是否支持对用 dlopen 方式打开动态库的程序进行性能分析。

还是以上述程序为例，对主程序代码进修改。

1) 源代码 (libTestProfiler.so 的源码同示例 2)

2) Makefile

GPerfTools=/home/wuzhu/tools/gperftools

CCFLAGS=-fno-omit-frame-pointer -g

ALL_BINS=libTestProfiler.so main_dlopen_link

all:$(ALL_BINS)

main_dlopen_link :main_dlopen_link.o

g++ $(CCFLAGS) -o $@ $^ -L./ -L$(GPerfTools)/lib -Wl,-Bdynamic -ltcmalloc_and_profiler -ldl

libTestProfiler.so:TestProfiler.o

g++ -shared $(CCFLAGS) -o $@ $^

.cpp.o:

g++ $(CCFLAGS) -c -I./ -I$(GPerfTools)/include -fPIC -o $@ $<

clean:

rm -f $(ALL_BINS) *.o *.prof

3) 运行

[wuzhu@search041142.sqa.cm4 cpu_profiler]$ ./main_dlopen_link

result: 127

do dlclose()

PROFILE: interrupts/evictions/bytes = 5984/3559/171280

产生性能分析数据：main_dlopen_link_1256.prof

4) 分析性能数据

pprof –text main_dlopen_link main_dlopen_link_13598.prof

Using local file main_dlopen_link.

Using local file main_dlopen_link_13598.prof.

Removing killpg from all stack traces.

Removing main from all stack traces.

Removing __libc_start_main from all stack traces.

Total: 5984 samples

801 13.4% 13.4% 801 13.4% 0x00002b124c0b4688

736 12.3% 25.7% 736 12.3% 0x00002b124c0b4652

735 12.3% 38.0% 735 12.3% 0x00002b124c0b4685

425 7.1% 45.1% 425 7.1% 0x00002b124c0b4671

414 6.9% 52.0% 414 6.9% 0x00002b124c0b464b

412 6.9% 58.9% 412 6.9% 0x00002b124c0b4676

401 6.7% 65.6% 401 6.7% 0x00002b124c0b4656

401 6.7% 72.3% 401 6.7% 0x00002b124c0b465c

392 6.6% 78.8% 392 6.6% 0x00002b124c0b4624

387 6.5% 85.3% 387 6.5% 0x00002b124c0b463e

382 6.4% 91.7% 382 6.4% 0x00002b124c0b467f

368 6.1% 97.8% 368 6.1% 0x00002b124c0b4639

45 0.8% 98.6% 45 0.8% memcpy

42 0.7% 99.3% 42 0.7% memset

21 0.4% 99.6% 21 0.4% 0x00002b124c0b468c

9 0.2% 99.8% 9 0.2% 0x00002b124c0b4645

3 0.1% 99.8% 3 0.1% 0x00002b124c0b467d

3 0.1% 99.9% 3 0.1% 0x00002b124c0b469c

2 0.0% 99.9% 2 0.0% 0x00002b124c0b46a6

1 0.0% 99.9% 1 0.0% 0x00002b124c0b4518

1 0.0% 99.9% 1 0.0% 0x00002b124c0b461b

1 0.0% 100.0% 1 0.0% 0x00002b124c0b4643

1 0.0% 100.0% 1 0.0% 0x00002b124c0b4695

1 0.0% 100.0% 1 0.0% 0x00002b124c0b46d4

很奇怪，这个结果显示 libTestProfiler.so 库中的符号没有正确解析，main_dlopen_link_1256.prof 文件也没有包含 libTestProfiler.so 的内存映射信息，但是我们确实在主程序已经通过dlopen将动态库装载到内存并执行成功了，为何在主程序的内存映射表中找不到动态库的信息呢？

经过一番分析和调查，终于找到原因，因为main_dlopen_link_1256.prof 文件的输出工作是在dlclose()函数执行之后调用的，而在此时主程序执行了dlclose()函数卸载了libTestProfiler.so，所以随后dump出的内存映射表当然就不会包含libTestProfiler.so的信息了。

下面，我们通过传入参数，禁止dlclose()

[wuzhu@search041142.sqa.cm4 cpu_profiler]$ ./main_dlopen_link 1

result: 127

not dlclose()

PROFILE: interrupts/evictions/bytes = 5968/3520/169408

pprof –text main_dlopen_link main_dlopen_link_18394.prof

Using local file main_dlopen_link.

Using local file main_dlopen_link_18394.prof.

Removing killpg from all stack traces.

Removing main from all stack traces.

Removing __libc_start_main from all stack traces.

Total: 5968 samples

5893 98.7% 98.7% 5893 98.7% loopop

46 0.8% 99.5% 46 0.8% memcpy

27 0.5% 100.0% 27 0.5% memset

2 0.0% 100.0% 2 0.0% _init

此时，动态库中的符号又能正常解析了。

(五) GoogleProfiler.h 小工具

该小工具主要是提供如下功能：

 自动生成性能数据文件的名称：程序名_进程号.prof

 将性能数据文件自动转换成 .callgrind 格式，方便 kcachegrind 上展示

 提供宏定义，可通过是否传 -DCPU_PROFILER 宏来打开或关闭google cpu profiler 性能采集.

下面是使用示例:

 Main 程序

 Makefile

GPerfTools=/home/wuzhu/tools/gperftools

CCFLAGS=-fno-omit-frame-pointer -g -DCPU_PROFILER

ALL_BINS=libTestProfiler.so \

main_dlopen_link2

all:$(ALL_BINS)

main_dlopen_link2 :main_dlopen_link2.o

g++ $(CCFLAGS) -o $@ $^ -L./ -L$(GPerfTools)/lib -Wl,-Bdynamic -ltcmalloc_and_profiler -ldl

libTestProfiler.so:TestProfiler.o

g++ -shared $(CCFLAGS) -o $@ $^

.cpp.o:

g++ $(CCFLAGS) -c -I./ -I$(GPerfTools)/include -fPIC -o $@ $<

clean:

rm -f $(ALL_BINS) *.o *.prof

 执行

[wuzhu@search041142.sqa.cm4 cpu_profiler]$ ./main_dlopen_link2 1

ProgramBin=/home/wuzhu/develop/mytest/cpu_profiler/main_dlopen_link2

ProfilerStart(/home/wuzhu/develop/mytest/cpu_profiler/main_dlopen_link2)

result: 127

not dlclose()

PROFILE: interrupts/evictions/bytes = 5962/3555/171088

cmd=pprof –callgrind /home/wuzhu/develop/mytest/cpu_profiler/main_dlopen_link2 main_dlopen_link2_24515.prof >main_dlopen_link2_24515.callgrind

Using local file /home/wuzhu/develop/mytest/cpu_profiler/main_dlopen_link2.

Using local file main_dlopen_link2_24515.prof.

Removing killpg from all stack traces.

Removing main from all stack traces.

Removing __libc_start_main from all stack traces.

[wuzhu@search041142.sqa.cm4 cpu_profiler]$

生成 main_dlopen_link2_24515.prof 和 main_dlopen_link2_24515.callgrind

 用 kcachegrind 打开查看





(六) 参考

 http://www.cnblogs.com/lenolix/archive/2010/12/13/1904868.html
 http://gperftools.googlecode.com/svn/trunk/doc/cpuprofile.html