Google CPU Profiler使用指南及小工具
Google CPU Profiler使用指南及小工具
–五竹,2012/12/27
(一) gperftools-2.0 编译
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：
#includeint ProfilerStart(const char* fname); void ProfilerFlush(); void ProfilerStop();
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
#include#include #include #include #include using namespace std; int loopop() { char buffer1 [1024]; char buffer2 [1024]; int n = 0; for(int i = 0; i < 10000000; i++) { for(int j = 0; j < 1000; j++) { n |= i%100 + j/100; } memset(buffer1,0xa,1024); memcpy(buffer2,buffer1,1024); } return n; } int main(int argc,char** argv) { char program[1024]={0}; snprintf(program,1023,"%s_%d.prof",argv[0],getpid()); ProfilerStart(program); printf("result: %d\n", (loopop)() ); ProfilerStop(); return 0; }
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 [email protected] $^ -L./ -L$(GPerfTools)/lib -Wl,-Bdynamic -ltcmalloc_and_profiler
.cpp.o:
g++ $(CCFLAGS) -c -I./ -I$(GPerfTools)/include -fPIC -o [email protected] $<
clean:
rm -f $(ALL_BINS) *.o
3) 执行
[[email protected] cpu_profiler]$ ./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
//TestProfiler.h
extern "C"
{
int loopop();
}
//TestProfiler.cpp只定义了一个耗时计算函数，便于分析。
#include
#include "TestProfiler.h"
extern "C"
{
int loopop()
{
char buffer1 [1024];
char buffer2 [1024];
int n = 0;
for(int i = 0; i < 10000000; i++)
{
for(int j = 0; j < 1000; j++)
{
n |= i%100 + j/100;
}
memset(buffer1,0xa,1024);
memcpy(buffer2,buffer1,1024);
}
return n;
}
}
#include
#include
#include
#include
#include
#include "TestProfiler.h"
using namespace std;
int main(int argc,char** argv)
{
char program[1024]={0};
snprintf(program,1023,"%s_%d.prof",argv[0],getpid());
ProfilerStart(program);
printf("result: %d\n", (loopop)() );
ProfilerStop();
return 0;
}
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 [email protected] $^ -L./ -L$(GPerfTools)/lib -Wl,-Bdynamic -lTestProfiler -ltcmalloc_and_profiler
libTestProfiler.so:TestProfiler.o
g++ -shared $(CCFLAGS) -o [email protected] $^
.cpp.o:
g++ $(CCFLAGS) -c -I./ -I$(GPerfTools)/include -fPIC -o [email protected] $<
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)
#include#include #include #include #include #include #include typedef int (*op_t) (); int main(int argc,char** argv) { int not_close=0; if(argc >= 2) not_close=atoi(argv[1]); void* dl_handle=NULL; op_t loopop; char* error=NULL; char program[1024]={0}; snprintf(program,1023,"%s_%d.prof",argv[0],getpid()); ProfilerStart(program); dl_handle = dlopen("./libTestProfiler.so", RTLD_LAZY ); if (!dl_handle) { printf( "dlopen failed! %s\n", dlerror() ); return -1; } loopop = (op_t)dlsym( dl_handle, "loopop"); error = dlerror(); if (error != NULL) { printf( "dlsym failed! %s\n", error ); return -1; } printf("result: %d\n", (loopop)() ); if(not_close == 0) { printf("do dlclose()\n"); dlclose( dl_handle ); } else { printf("not dlclose()\n"); } ProfilerStop(); return 0; }
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 [email protected] $^ -L./ -L$(GPerfTools)/lib -Wl,-Bdynamic -ltcmalloc_and_profiler -ldl
libTestProfiler.so:TestProfiler.o
g++ -shared $(CCFLAGS) -o [email protected] $^
.cpp.o:
g++ $(CCFLAGS) -c -I./ -I$(GPerfTools)/include -fPIC -o [email protected] $<
clean:
rm -f $(ALL_BINS) *.o *.prof
3) 运行
[[email protected] cpu_profiler]$ ./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()
[[email protected] 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
#ifndef __GOOGLE_PROFILER__ #define __GOOGLE_PROFILER__ #include#include #include #ifdef CPU_PROFILER #include #endif class GoogleProfiler { public: enum { MAXBUFSIZE=1024 }; public: GoogleProfiler():_ready(false) { } ~GoogleProfiler() { ProfilerStop(); } void ProfilerStart() { #ifdef CPU_PROFILER char buf[MAXBUFSIZE]={0}; int count = readlink("/proc/self/exe",buf, MAXBUFSIZE ); if(count < 0 || count > MAXBUFSIZE) return; _programBin.assign(buf); printf("ProgramBin=%s\n",_programBin.c_str()); _pid=getpid(); size_t pos=_programBin.find_last_of('/'); if(pos != std::string::npos) { _programName=_programBin.substr(pos+1); } else { _programName=_programBin; } printf("ProfilerStart(%s)\n",buf); snprintf(buf,MAXBUFSIZE,"./%s_%d.prof",_programName.c_str(),_pid); ::ProfilerStart(buf); _ready=true; #endif } void ProfilerStop() { #ifdef CPU_PROFILER if(_ready) { ::ProfilerStop(); char cmd[2048]={0}; snprintf(cmd,2047,"pprof --callgrind %s %s_%d.prof >%s_%d.callgrind" ,_programBin.c_str(),_programName.c_str(),_pid,_programName.c_str(),_pid); printf("cmd=%s\n",cmd); system(cmd); _ready=false; } #endif } void ProfilerFlush() { ::ProfilerFlush(); } private: bool _ready; int _pid; std::string _programBin; std::string _programName; }; #endif
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
#include#include #include #include #include #include #include #include "GoogleProfiler.h" typedef int (*op_t) (); int main(int argc,char** argv) { int not_close=0; if(argc >= 2) not_close=atoi(argv[1]); void* dl_handle=NULL; op_t loopop; char* error=NULL; GoogleProfiler profiler; profiler.ProfilerStart(); dl_handle = dlopen("./libTestProfiler.so", RTLD_LAZY ); if (!dl_handle) { printf( "dlopen failed! %s\n", dlerror() ); return -1; } loopop = (op_t)dlsym( dl_handle, "loopop"); error = dlerror(); if (error != NULL) { printf( "dlsym failed! %s\n", error ); return -1; } printf("result: %d\n", (loopop)() ); if(not_close == 0) { printf("do dlclose()\n"); dlclose( dl_handle ); } else { printf("not dlclose()\n"); } profiler.ProfilerStop(); return 0; }
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 [email protected] $^ -L./ -L$(GPerfTools)/lib -Wl,-Bdynamic -ltcmalloc_and_profiler -ldl
libTestProfiler.so:TestProfiler.o
g++ -shared $(CCFLAGS) -o [email protected] $^
.cpp.o:
g++ $(CCFLAGS) -c -I./ -I$(GPerfTools)/include -fPIC -o [email protected] $<
clean:
rm -f $(ALL_BINS) *.o *.prof
执行
[[email protected] cpu_profiler]$ ./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.
[[email protected] 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