On Oct 4, 2014, at 6:08 PM, Milian Wolff <[email protected]> wrote:
>
> this sounds very interesting. How does it compare speed-wise to local
> unwinding?
>
Local unwinding is special. For single-threaded programs we can access process'
memory and registers without overhead. In multi-threaded programs we need a way
to stop threads and get their registers. It implies using IPC mechanisms, so the
speed will degrade again.
I conducted simple tests tracing one thread of multithreaded program. Stack
contained 14 frames in 3 shared objects + 1 executable.
In the first test I took a code piece from libunwind man page and measured time
with RDTSC.
void show_backtrace (void)
{
unw_cursor_t cursor; unw_context_t uc;
unw_word_t ip, sp;
unw_getcontext(&uc);
unw_init_local(&cursor, &uc);
while (unw_step(&cursor) > 0) {
unw_get_reg(&cursor, UNW_REG_IP, &ip);
unw_get_reg(&cursor, UNW_REG_SP, &sp);
}
}
In the second test I put never-ending select() in that place. In third test
there was infinite loop. tbstack copied data and registers for only one thread.
Tracing the process several times I got numbers in following range:
1) 37-56 microseconds for local unwinding (183 on cold start);
2) 129-166 microseconds with thread sleeping in select();
3) 124-141 microseconds with spinning thread.
_______________________________________________
Libunwind-devel mailing list
[email protected]
https://lists.nongnu.org/mailman/listinfo/libunwind-devel
