gperftools cpu profiler does not support multi process?
according to the document, http://gperftools.googlecode.com/svn/trunk/doc/cpuprofile.html, the cpu profiles does support multi process and will generate independent output file:
If your program forks, the children will also be profiled (since they inherit the same CPUPROFILE setting). Each process is profiled separately; to distinguish the child profiles from the parent profile and from each other, all children will have their process-id appended to the CPUPROFILE name.
but when I try as follow:
// main_cmd_argv.cpp
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <string.h>
#include <sys/types.h>
#include <sys/wait.h>
#include <gperftools/profiler.h>
int loop(int n) {
int sum = 0;
for (int i = 0; i < n; i++) {
for (int j = 0; j < n; j++) {
sum = i + j;
if (sum %3 == 0) {
sum /= 3;
}
}
}
return 0;
}
int main(int argc, char* argv[]) {
printf("%sn%sn", getenv("CPUPROFILE"), getenv("CPUPROFILESIGNAL"));
if (argc > 1 && strcmp(argv[1], "-s")==0) {
// single process
loop(100000);
printf("stopedn");
} else if (argc > 1 && strcmp(argv[1], "-m")==0) {
// multi process
pid_t pid = fork();
if (pid < 0) {
printf("fork errorn");
return -1;
}
if (pid == 0) {
loop(100000);
printf("child stopedn");
} else if (pid > 0) {
loop(10000);
printf("father stopedn");
wait(NULL);
}
}
return 0;
}
// makefile
GPerfTools=/home/adenzhang/tools/gperftools
CCFLAGS=-fno-omit-frame-pointer -g -Wall
ALL_BINS=main_cmd_argv
all:$(ALL_BINS)
main_cmd_argv:main_cmd_argv.o
g++ $(CCFLAGS) -o $@ $^ -L./ -L$(GPerfTools)/lib -Wl,-Bdynamic -lprofiler -lunwind
.cpp.o:
g++ $(CCFLAGS) -c -I./ -I$(GPerfTools)/include -fPIC -o $@ $<
clean:
rm -f $(ALL_BINS) *.o *.prof
// shell command
$ make
g++ -fno-omit-frame-pointer -g -Wall -c -I./ -I/home/adenzhang/tools/gperftools/include -fPIC -o main_cmd_argv.o main_cmd_argv.cpp
g++ -fno-omit-frame-pointer -g -Wall -o main_cmd_argv main_cmd_argv.o -L./ -L/home/adenzhang/tools/gperftools/lib -Wl,-Bdynamic -lprofiler -lunwind
$ env CPUPROFILE=main_cmd_argv.prof ./main_cmd_argv -s
젩n_cmd_argv.prof
(null)
stoped
PROFILE: interrupts/evictions/bytes = 6686/3564/228416
$ /home/adenzhang/tools/gperftools/bin/pprof --text ./main_cmd_argv ./main_cmd_argv.prof
Using local file ./main_cmd_argv.
Using local file ./main_cmd_argv.prof.
Removing killpg from all stack traces.
Total: 6686 samples
6686 100.0% 100.0% 6686 100.0% loop
0 0.0% 100.0% 6686 100.0% __libc_start_main
0 0.0% 100.0% 6686 100.0% _start
0 0.0% 100.0% 6686 100.0% main
$ rm main_cmd_argv.prof
$ env CPUPROFILE=main_cmd_argv.prof ./main_cmd_argv -m
젩n_cmd_argv.prof
(null)
father stoped
child stoped
PROFILE: interrupts/evictions/bytes = 0/0/64
PROFILE: interrupts/evictions/bytes = 68/36/2624
$ ls
main_cmd_argv main_cmd_argv.cpp main_cmd_argv.o main_cmd_argv.prof Makefile
$ /home/adenzhang/tools/gperftools/bin/pprof --text ./main_cmd_argv ./main_cmd_argv.prof
Using local file ./main_cmd_argv.
Using local file ./main_cmd_argv.prof.
$
It semms that gperf does not support multi process, could anyone please explain? thanks!
Quite old, don't know if you found an answer or not, but...
Seems like every thread/fork should register itself using ProfilerRegisterThread(); You can find more information in those two issues: Here and Here.
Also here is an example code, similar to your test case where the forks can be registered.
I'm currently using gperftools to profile a mpi program and come across this problem. After googling I find that ProfilerStart(_YOUR_PROF_FILE_NAME_) and ProfilerStop() ought be called during every sub-process is executed, and _YOUR_PRO_FILE_NAME_ must be different among different process. Then you could analysis performance of every process.
link(also asked by ZRJ):
https://groups.google.com/forum/#!topic/google-perftools/bmysZILR4ik