Hi all,
Yesterday during the performance call there was some talk around
measuring lock contention in wallclock, and it was mentioned that an
option would be to monitor the futex system call. I just uploaded to
Github a script I've used in the past that does exactly that, in case
there was interest in pursuing this approach
(https://github.com/mogeb/bt-futex). The Python script expects a CTF
trace (e.g. LTTng). The functionality is very minimal at this point, but
if there is interest I'd be happy to enhance it further. The main
drawback of this approach is that it only works for locks that use futex
(spinlocks, for instance, wouldn't be monitored). No instrumentation to
the application is required.
Here's an example output of the amount of time spent in the futex
syscall, per thread (system-wide). The LOCK_TIME column shows the amount
of time waiting on a lock, in nanoseconds:
$> python3 count_futex.py kernel/
PROCESS_NAME (TID): LOCK_TIME ns
git (17550): 2,122
git (17551): 8,124
example (17554): 80,690,708
example (17555): 80,020,536
example (17556): 367,010,253
example (17557): 459,009,153
example (17558): 199,676,209
example (17559): 108,401
example (17560): 475,562,084
example (17561): 80,549,102
example (17562): 468,002,252
example (17564): 92,108,616
example (17566): 450,655,745
example (17567): 49,904,388
libvirtd (3772): 2,389
libvirtd (3773): 2,839
libvirtd (3775): 2,417
libvirtd (3776): 3,577
libvirtd (3777): 2,492
Mohamad
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
