[lttng-dev] large liblttng-ust startup overhead (ust_lock)
Mathieu Desnoyers
mathieu.desnoyers at efficios.com
Wed Sep 6 20:23:40 UTC 2017
----- On Sep 6, 2017, at 3:57 PM, Mathieu Desnoyers mathieu.desnoyers at efficios.com wrote:
> ----- On Sep 6, 2017, at 3:35 AM, Milian Wolff milian.wolff at kdab.com wrote:
>
>> On Dienstag, 5. September 2017 20:11:58 CEST Mathieu Desnoyers wrote:
>>> ----- On Sep 5, 2017, at 11:08 AM, Milian Wolff milian.wolff at kdab.com wrote:
>>> > On Tuesday, September 5, 2017 4:51:42 PM CEST Mathieu Desnoyers wrote:
>>> >> ----- On Sep 5, 2017, at 10:34 AM, Milian Wolff milian.wolff at kdab.com
>> wrote:
>>> >> > Hey all,
>>> >> >
>>> >> > I have noticed a very large overhead when linking against liblttng-ust:
>>> >> >
>>> >> > ~~~~~
>>> >> > ┌milian at milian-kdab2:/tmp
>>> >> > └$ cat lttng-test.c
>>> >> > int main()
>>> >> > {
>>> >> >
>>> >> > return 0;
>>> >> >
>>> >> > }
>>> >> > ┌milian at milian-kdab2:/tmp
>>> >> > └$ gcc -O2 -g -ldl lttng-test.c
>>> >> > ┌milian at milian-kdab2:/tmp
>>> >> > └$ perf stat -r 5 ./a.out
>>> >> >
>>> >> > Performance counter stats for './a.out' (5 runs):
>>> >> > 0.209587 task-clock (msec) # 0.596 CPUs
>>> >> > utilized
>>> >> >
>>> >> > ( +- 8.76% )
>>> >> >
>>> >> > 0 context-switches # 0.000 K/sec
>>> >> > 0 cpu-migrations # 0.000 K/sec
>>> >> >
>>> >> > 49 page-faults # 0.235 M/sec
>>> >> >
>>> >> > ( +- 1.19% )
>>> >> >
>>> >> > 706,854 cycles # 3.373 GHz
>>> >> >
>>> >> > ( +- 8.82% )
>>> >> >
>>> >> > 773,603 instructions # 1.09 insn per
>>> >> > cycle
>>> >> >
>>> >> > ( +- 0.75% )
>>> >> >
>>> >> > 147,128 branches # 701.987 M/sec
>>> >> >
>>> >> > ( +- 0.70% )
>>> >> >
>>> >> > 4,096 branch-misses # 2.78% of all
>>> >> > branches
>>> >> >
>>> >> > ( +- 5.27% )
>>> >> >
>>> >> > 0.000351422 seconds time elapsed
>>> >> >
>>> >> > ( +- 11.85% )
>>> >> >
>>> >> > ┌milian at milian-kdab2:/tmp
>>> >> > └$ gcc -O2 -g -ldl -llttng-ust lttng-test.c
>>> >> > ┌milian at milian-kdab2:/tmp
>>> >> > └$ perf stat -r 5 ./a.out
>>> >> >
>>> >> > Performance counter stats for './a.out' (5 runs):
>>> >> > 2.063040 task-clock (msec) # 0.009 CPUs
>>> >> > utilized
>>> >> >
>>> >> > ( +- 1.37% )
>>> >> >
>>> >> > 44 context-switches # 0.021 M/sec
>>> >> >
>>> >> > ( +- 1.95% )
>>> >> >
>>> >> > 2 cpu-migrations # 0.776 K/sec
>>> >> >
>>> >> > ( +- 25.00% )
>>> >> >
>>> >> > 209 page-faults # 0.101 M/sec
>>> >> >
>>> >> > ( +- 0.34% )
>>> >> >
>>> >> > 7,053,686 cycles # 3.419 GHz
>>> >> >
>>> >> > ( +- 2.03% )
>>> >> >
>>> >> > 6,893,783 instructions # 0.98 insn per
>>> >> > cycle
>>> >> >
>>> >> > ( +- 0.25% )
>>> >> >
>>> >> > 1,342,492 branches # 650.735 M/sec
>>> >> >
>>> >> > ( +- 0.20% )
>>> >> >
>>> >> > 29,390 branch-misses # 2.19% of all
>>> >> > branches
>>> >> >
>>> >> > ( +- 0.61% )
>>> >> >
>>> >> > 0.225597302 seconds time elapsed
>>> >> >
>>> >> > ( +- 6.68% )
>>> >> > ~~~~~
>>> >> >
>>> >> > This is without any LTTng session configured. If I enable LTTng kernel
>>> >> > and
>>> >> > userspace events, this becomes even worse:
>>> >> >
>>> >> > ~~~~~
>>> >> > ┌milian at milian-kdab2:/tmp
>>> >> > └$ cat $(which run_lttng_trace.sh)
>>> >> > #!/bin/sh
>>> >> >
>>> >> > if [ -z "$(pidof lttng-sessiond)" ]; then
>>> >> >
>>> >> > sudo lttng-sessiond --daemonize
>>> >> >
>>> >> > fi
>>> >> >
>>> >> > sudo lttng create -o ~/lttng-traces/$(date -Iseconds)
>>> >> > sudo lttng enable-channel kernel -k --subbuf-size 16M --num-subbuf 8
>>> >> > sudo lttng enable-event -c kernel -k -a
>>> >> > sudo lttng enable-channel ust -u --subbuf-size 16M --num-subbuf 8
>>> >> > sudo lttng enable-event -c ust -u lttng_ust_tracef:*
>>> >> > sudo lttng start
>>> >> > $@
>>> >> > sudo lttng stop
>>> >> >
>>> >> > sudo chmod a+rx -R ~/lttng-traces
>>> >> > ┌milian at milian-kdab2:/tmp
>>> >> > └$ run_lttng_trace.sh perf stat -r 5 ./a.out
>>> >> > Session auto-20170905-162818 created.
>>> >> > Traces will be written in
>>> >> > /home/milian/lttng-traces/2017-09-05T16:28:18+02:00 Kernel channel
>>> >> > kernel
>>> >> > enabled for session auto-20170905-162818
>>> >> > All Kernel events are enabled in channel kernel
>>> >> > UST channel ust enabled for session auto-20170905-162818
>>> >> > UST event lttng_ust_tracef:* created in channel ust
>>> >> > Tracing started for session auto-20170905-162818
>>> >> >
>>> >> > Performance counter stats for './a.out' (5 runs):
>>> >> > 4.937820 task-clock (msec) # 0.005 CPUs
>>> >> > utilized
>>> >> >
>>> >> > ( +- 1.28% )
>>> >> >
>>> >> > 117 context-switches # 0.024 M/sec
>>> >> >
>>> >> > ( +- 0.70% )
>>> >> >
>>> >> > 3 cpu-migrations # 0.608 K/sec
>>> >> >
>>> >> > ( +- 23.57% )
>>> >> >
>>> >> > 245 page-faults # 0.050 M/sec
>>> >> >
>>> >> > ( +- 0.20% )
>>> >> >
>>> >> > 16,541,355 cycles # 3.350 GHz
>>> >> >
>>> >> > ( +- 0.94% )
>>> >> >
>>> >> > 20,611,637 instructions # 1.25 insn per
>>> >> > cycle
>>> >> >
>>> >> > ( +- 0.23% )
>>> >> >
>>> >> > 3,580,525 branches # 725.123 M/sec
>>> >> >
>>> >> > ( +- 0.19% )
>>> >> >
>>> >> > 52,093 branch-misses # 1.45% of all
>>> >> > branches
>>> >> >
>>> >> > ( +- 1.73% )
>>> >> >
>>> >> > 0.965545882 seconds time elapsed
>>> >> >
>>> >> > ( +- 16.96% )
>>> >> >
>>> >> > Waiting for data availability...
>>> >> > Tracing stopped for session auto-20170905-162818
>>> >> > ~~~~~
>>> >> >
>>> >> > I see this ~1s delay during startup in nearly all test applications
>>> >> > where
>>> >> > I
>>> >> > wanted to use lttng-ust. In tracecompass, and according to perf sleep
>>> >> > time
>>> >> > profiling, this seems to be mostly due to extensive sleeping on mutex
>>> >> > locks
>>> >> > (from ust_lock in liblttng-ust). With perf trace, I also see membarrier
>>> >> > taking a lot of time:
>>> >> >
>>> >> > ~~~~~
>>> >> > # without a running lttng session:
>>> >> > ┌milian at milian-kdab2:/tmp
>>> >> > └$ perf trace --duration 1 ./a.out
>>> >> >
>>> >> > 6.492 (52.468 ms): a.out/23672 recvmsg(fd: 3<socket:[1178439]>,
>> msg:
>>> >> > 0x7fbe2fbb1070 ) = 1
>>> >> >
>>> >> > 5.077 (54.271 ms): a.out/23671 futex(uaddr: 0x7fbe30d508a0, op:
>>> >> > WAIT_BITSET|PRIV|CLKRT, utime: 0x7ffc474ff5a0, val3: 4294967295) = 0
>>> >> >
>>> >> > 59.598 (79.379 ms): a.out/23671 membarrier(cmd: 1
>>> >> >
>>> >> > ) = 0
>>> >> >
>>> >> > 138.984 (39.945 ms): a.out/23671 membarrier(cmd: 1
>>> >> >
>>> >> > ) = 0
>>> >> >
>>> >> > 59.621 (119.324 ms): a.out/23672 futex(uaddr: 0x7fbe303e54c0, op:
>>> >> > WAIT|
>>> >> >
>>> >> > PRIV, val: 2 ) = 0
>>> >> >
>>> >> > 179.045 (79.918 ms): a.out/23671 membarrier(cmd: 1
>>> >> >
>>> >> > ) = 0
>>> >> >
>>> >> > 258.971 (39.997 ms): a.out/23671 membarrier(cmd: 1
>>> >> >
>>> >> > ) = 0
>>> >> >
>>> >> > 299.052 (79.883 ms): a.out/23671 membarrier(cmd: 1
>>> >> >
>>> >> > ) = 0
>>> >> >
>>> >> > 378.942 (59.988 ms): a.out/23671 membarrier(cmd: 1
>>> >> >
>>> >> > ) = 0
>>> >> >
>>> >> > 439.022 (69.908 ms): a.out/23671 membarrier(cmd: 1
>>> >> >
>>> >> > ) = 0
>>> >> >
>>> >> > 508.937 (80.027 ms): a.out/23671 membarrier(cmd: 1
>>> >> >
>>> >> > ) = 0
>>> >> >
>>> >> > # with a running lttng session:
>>> >> > ┌milian at milian-kdab2:/tmp
>>> >> > └$ run_lttng_trace.sh perf trace --duration 1 ./a.out
>>> >> >
>>> >> > 6.666 (392.496 ms): a.out-ust/23790 recvmsg(fd:
>>> >> > 3<socket:[1185362]>,
>>> >
>>> > msg:
>>> >> > 0x7fcd81d8a0b0 ) = 612
>>> >> >
>>> >> > 400.238 (68.452 ms): a.out-ust/23790 membarrier(cmd: 1
>>> >> >
>>> >> > ) = 0
>>> >> >
>>> >> > 468.698 (39.983 ms): a.out-ust/23790 membarrier(cmd: 1
>>> >> >
>>> >> > ) = 0
>>> >> >
>>> >> > 510.567 (98.113 ms): a.out-ust/23790 recvmsg(fd: 3<socket:[1185362]>,
>>> >
>>> > msg:
>>> >> > 0x7fcd81d8a070 ) = 1
>>> >> >
>>> >> > 6.467 (603.500 ms): a.out/23789 futex(uaddr: 0x7fcd82f298a0, op:
>>> >> > WAIT_BITSET|PRIV|CLKRT, utime: 0x7fff4df6c970, val3: 4294967295) = 0
>>> >> >
>>> >> > 610.020 ( 6.307 ms): a.out/23789 futex(uaddr: 0x7fcd82f29960, op:
>>> >> > WAIT|
>>> >> >
>>> >> > PRIV, val: 2 ) = 0
>>> >> >
>>> >> > 616.446 (62.265 ms): a.out/23789 membarrier(cmd: 1
>>> >> >
>>> >> > ) = 0
>>> >> >
>>> >> > 678.721 (59.916 ms): a.out/23789 membarrier(cmd: 1
>>> >> >
>>> >> > ) = 0
>>> >> >
>>> >> > 616.332 (122.335 ms): a.out-ust/23790 futex(uaddr: 0x7fcd82f29960,
>> op:
>>> >> > WAIT|PRIV, val: 2 ) = 0
>>> >> >
>>> >> > 738.976 (49.658 ms): a.out/23789 membarrier(cmd: 1
>>> >> >
>>> >> > ) = 0
>>> >> >
>>> >> > 788.642 (49.987 ms): a.out/23789 membarrier(cmd: 1
>>> >> >
>>> >> > ) = 0
>>> >> >
>>> >> > 739.014 (99.644 ms): a.out-ust/23790 futex(uaddr: 0x7fcd825be4c0, op:
>>> >> > WAIT|
>>> >> >
>>> >> > PRIV, val: 2 ) = 0
>>> >> >
>>> >> > 838.651 (49.977 ms): a.out/23789 membarrier(cmd: 1
>>> >> >
>>> >> > ) = 0
>>> >> >
>>> >> > 888.636 (69.983 ms): a.out/23789 membarrier(cmd: 1
>>> >> >
>>> >> > ) = 0
>>> >> >
>>> >> > 838.662 (119.978 ms): a.out-ust/23790 futex(uaddr: 0x7fcd825be4c0,
>> op:
>>> >> > WAIT|PRIV, val: 2 ) = 0
>>> >> >
>>> >> > 958.739 (59.873 ms): a.out/23789 membarrier(cmd: 1
>>> >> >
>>> >> > ) = 0
>>> >> >
>>> >> > 1018.621 (80.058 ms): a.out/23789 membarrier(cmd: 1
>>> >> >
>>> >> > ) = 0
>>> >> >
>>> >> > 1098.782 (49.844 ms): a.out/23789 membarrier(cmd: 1
>>> >> >
>>> >> > ) = 0
>>> >> >
>>> >> > 1148.634 (59.998 ms): a.out/23789 membarrier(cmd: 1
>>> >> >
>>> >> > ) = 0
>>> >> >
>>> >> > 1208.731 (69.940 ms): a.out/23789 membarrier(cmd: 1
>>> >> >
>>> >> > ) = 0
>>> >> >
>>> >> > 1278.680 (79.947 ms): a.out/23789 membarrier(cmd: 1
>>> >> >
>>> >> > ) = 0
>>> >> >
>>> >> > 1358.727 (39.881 ms): a.out/23789 membarrier(cmd: 1
>>> >> >
>>> >> > ) = 0
>>> >> >
>>> >> > 1398.616 (39.980 ms): a.out/23789 membarrier(cmd: 1
>>> >> >
>>> >> > ) = 0
>>> >> >
>>> >> > 6.640 (1432.157 ms): a.out-ust/23791 futex(uaddr: 0x7fcd82f90000
>>> >> >
>>> >> > ) = -1 (null) INTERNAL ERROR: strerror_r(512, 0x7ffdbc5467a0, 128)=22
>>> >> > ~~~~~
>>> >> >
>>> >> > This makes lttng-ust pretty much unusable for any short-lived
>>> >> > application.
>>> >> > Is this a known limitation?
>>> >>
>>> >> This kind of delay is not completely unexpected, but a bit higher than I
>>> >> would expect. I'm interested in figuring out if most of this delay
>>> >> happens
>>> >> to be caused by sys_membarrier, either directly, or indirectly (through a
>>> >> lock).
>>> >
>>> > See my other email which adds some harder numbers from a perf sleep
>>> > record.
>>> >
>>> >> > I'm using lttng-ust on Arch from the AUR packages, i.e. LTTng 2.9.4.
>>> >> > I've
>>> >> > seen this problem also on Ubuntu 17.04 though I didn't do any of the
>>> >> > above measurements there. So far, I suspected the ust statedump to take
>>> >> > too much time, but I disabled that one for the measurements above.
>>> >> >
>>> >> > Any help would be appreciated
>>> >>
>>> >> Can you reproduce with a Linux kernel that has CONFIG_MEMBARRIER=n and
>>> >> share the performance results ?
>>> >
>>> > I'd have to recompile the kernel, which is something I don't like to do.
>>> > Why is this required, or put differently - why would the kernel option
>>> > influence the userspace LTTng library?
>>>
>>> liburcu 0.9.x and 0.10.x (a dependency of lttng-ust) uses the membarrier
>>> system call to speed up tracing, at the expense of slower execution of
>>> tracing state update.
>>>
>>> I'm currently proposing a new sys_membarrier command for PRIVATE_EXPEDITED
>>> (single-process) use cases which should execute much faster than the
>>> non-expedited SHARED command currently implemented.
>>>
>>> liburcu 0.9 and 0.10 detects that this system call is available, and uses
>>> it when it can.
>>
>> May I ask whether you have such a configuration available? The test case I
>> have submitted should be easy to reproduce for everyone. Compiling my own
>> kernel version takes a huge amount of time which I'm quite frankly not willing
>> to invest at this point.
>
> We have reproduced the situation, and it's pointing towards our use of the
> sys_membarrier system call. We'll try narrowing it down between its use by
> lttng-ust (liburcu-bp.so) or by lttng-sessiond (liburcu.so).
>
> You can try to patch liburcu so it won't try using sys_membarrier at all
> (see attached patch)
>
>>
>> I simply wonder whether anyone else is actually using lttng-ust, or if they
>> are, whether they are only doing that for long-lived applications where this
>> overhead is not noticeable?
>
> It appears that this application startup slowdown mainly appears with
> LTTng linked against liburcu 0.9 and 0.10 running with rather recent kernels.
> This would explain why you are the first one to report this across our user
> base.
We have confirmed that it's the use of sys_membarrier SHARED by liburcu-bp,
used by lttng-ust, that causes those delays.
lttng-ust usage pattern is to call synchronize_rcu() directly. Having
membarrier system calls each time adds up to significant delays, because
membarrier SHARED blocks until a scheduler switch happens on each core.
We are invoking synchronize_rcu() directly rather than using the call_rcu
batching facilities because we want to minimize the number of threads we
have to insert into the application.
I'm currently proposing a PRIVATE_EXPEDITED command for the membarrier
system call in Linux. It will actively send IPIs rather than blocking.
I'm going to push a "Fix" in liburcu master, 0.10, 0.9 that will make it
stop using sys_membarrier in liburcu-bp. Then, after the new PRIVATE_EXPEDITED
command makes it way into the upstream Linux kernel, a future liburcu
(0.11 or +) will introduce usage of that new command, given that it has
a much, much smaller impact compared to the SHARED command.
The reason why we want sys_membarrier there is to speed up tracing. When
can trade memory barriers for simple compiler barriers on the tracing
fast-path by using sys_membarrier on the update side.
Thanks for reporting your finding!
Mathieu
>
> Thanks,
>
> Mathieu
>
>
>>
>> Thanks
>>
>> --
>> Milian Wolff | milian.wolff at kdab.com | Senior Software Engineer
>> KDAB (Deutschland) GmbH&Co KG, a KDAB Group company
>> Tel: +49-30-521325470
>> KDAB - The Qt Experts
>
> --
> Mathieu Desnoyers
> EfficiOS Inc.
> http://www.efficios.com
--
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com
More information about the lttng-dev
mailing list