gprof, threads and forks
It's rare that I cross-post to this many lists, but
1. I blew a week trying to track this little monster down, and
2. It isn't at all clear to me where the appropriate fixes and documentation changes need to be applied.
So here goes:
If you have an executable compiled with *gcc* using the "-pg" option, it is supposed to generate a "gmon.out" file, which you can later process with the *binutils* utility "gprof" to produce a profile. All this magic works just dandy for a single threaded executable, but apparently child threads and child processes created with "fork" don't get their execution time counted. See the following links for the gory details:
http://sources.redhat.com/ml/bug-binutils/2001-q3/msg00090.html
http://uwsg.iu.edu/hypermail/linux/kernel/0101.3/1516.html
The second link implies that it's possible to fix this at the *kernel* level. So my question to all of you is, "what's the best way to get this fixed?" I need to profile a multi-threaded executable and personally don't care about the "fork" case, but I'm sure there are others who would care about forks and not threads.
--
M. Edward (Ed) Borasky, Chief Scientist, Borasky Research
http://www.borasky-research.net http://www.aracnet.com/~znmeb
mailto:[email protected] mailto:[email protected]
Stand-Up Comedy: Because Man Does Not Live By Dread Alone
Quoted 3 lines · M. Edward Borasky
fixed?" I need to profile a multi-threaded executable and personally don't care about the "fork" case, but I'm sure there are others who would care about forks and not threads.
Did you try testing against an -ac kernel? I believe there is a patch in there that allows profiling of multi-threaded apps. If it has made it into the main kernel, then you can just ignore this message...
Quoted 17 lines · M. Edward Borasky
It's rare that I cross-post to this many lists, but
1. I blew a week trying to track this little monster down, and
2. It isn't at all clear to me where the appropriate fixes and documentation changes need to be applied.So here goes:
If you have an executable compiled with *gcc* using the "-pg" option, it is supposed to generate a "gmon.out" file, which you can later process with the *binutils* utility "gprof" to produce a profile. All this magic works just dandy for a single threaded executable, but apparently child threads and child processes created with "fork" don't get their execution time counted. See the following links for the gory details:
http://sources.redhat.com/ml/bug-binutils/2001-q3/msg00090.html
Quoted 7 lines
http://uwsg.iu.edu/hypermail/linux/kernel/0101.3/1516.html
The second link implies that it's possible to fix this at the *kernel* level. So my question to all of you is, "what's the best way to get this fixed?" I need to profile a multi-threaded executable and personally don't care about the "fork" case, but I'm sure there are others who would care about forks and not threads.
Fork is easy enough to handle, at least on GNU/Linux. If you set the environment variable GMON_OUT_PREFIX, glibc will automatically append the process ID when writing the gmon.out file. If you then also do this in the child after calling fork:
extern void _start (void), etext (void); monstartup ((u_long) &_start, (u_long) &etext);
you will a set of ${GMON_OUT_PREFIX}.pid files, one for the parent and one for each child.
As far as I know, none of this is documented, though that might just mean that I missed it. I think glibc ought to document GMON_OUT_PREFIX and the right way to call monstartup.
In principle threading could be handled the same way, but glibc would need to cooperate when generating the profiling output file so that each thread doesn't overwrite the same file.
Ian