On Wed, Jul 03, 2019 at 04:01:03PM -0700, Paul E. McKenney wrote:
> On Wed, Jul 03, 2019 at 06:24:06PM -0400, Joel Fernandes wrote:
> > On Wed, Jul 03, 2019 at 02:57:14PM -0700, Paul E. McKenney wrote:
> > > On Wed, Jul 03, 2019 at 05:24:26PM -0400, Joel Fernandes wrote:
> > > > On Wed, Jul 03, 2019 at 10:39:35AM -0700, Paul E. McKenney wrote:
> > > > > On Wed, Jul 03, 2019 at 12:41:34PM -0400, Joel Fernandes wrote:
> > > > > > On Wed, Jul 03, 2019 at 11:30:36AM -0400, Steven Rostedt wrote:
> > > > > > > On Wed, 3 Jul 2019 11:25:20 -0400
> > > > > > > Joel Fernandes <joel@xxxxxxxxxxxxxxxxx> wrote:
> > > > > > >
> > > > > > >
> > > > > > > > I am sorry if this is not a realistic real-life problem, but more a
> > > > > > > > "doctor it hurts if I do this" problem as Steven once said ;-)
> > > > > > > >
> > > > > > > > I'll keep poking ;-)
> > > > > > >
> > > > > > > Hi Joel,
> > > > > > >
> > > > > > > Can you also share the tests you are performing as well as any
> > > > > > > module/code changes you made so that we can duplicate the results?
> > > > > >
> > > > > > Sure thing. Below is the diff that I applied to Paul's /dev branch. But I
> > > > > > believe Linus's tree should have same results.
> > > > > >
> > > > > > After applying the diff below, I run it like this:
> > > > > > tools/testing/selftests/rcutorture/bin/kvm.sh --bootargs rcuperf.pd_test=1 rcuperf.pd_busy_wait=5000 rcuperf.holdout=5 rcuperf.pd_resched=0 --duration 1 --torture rcuperf
> > > > > >
> > > > > > Some new options I added:
> > > > > > pd_test=1 runs the preempt disable loop test
> > > > > > pd_busy_wait is the busy wait time each pass through the loop in microseconds
> > > > > > pd_resched is whether the loop should set the need-resched flag periodically.
> > > > > >
> > > > > > If your qemu is a bit old or from debian, then you may also need to pass: --qemu-args "-net nic,model=e1000"
> > > > > >
> > > > > > With pd_resched = 0, I get quite high average grace-period latencies. The
> > > > > > preempt-disable loop thread is running on its own CPU. Enabling the rcu:*
> > > > > > tracepoints, I see that for long periods of time, the FQS rcu loop can be
> > > > > > running while the scheduler tick learns from rcu_preempt_deferred_qs() that
> > > > > > there's nothing to worry about (at least this is what I remember tracing).
> > > > > >
> > > > > > With pd_resched = 0, the output of the command above:
> > > > > > Average grace-period duration: 195629 microseconds
> > > > > > Minimum grace-period duration: 30111.7
> > > > > > 50th percentile grace-period duration: 211000
> > > > > > 90th percentile grace-period duration: 218000
> > > > > > 99th percentile grace-period duration: 222999
> > > > > > Maximum grace-period duration: 236351
> > > > > >
> > > > > > With pd_resched = 1, you get more like twice (10ms) the busy-wait time (5ms).
> > > > > > I wonder why its twice, but that's still Ok. It is as follows:
> > > > > > Average grace-period duration: 12302.2 microseconds
> > > > > > Minimum grace-period duration: 5998.35
> > > > > > 50th percentile grace-period duration: 12000.4
> > > > > > 90th percentile grace-period duration: 15996.4
> > > > > > 99th percentile grace-period duration: 18000.6
> > > > > > Maximum grace-period duration: 20998.6
> > > > >
> > > > > Both of these results are within the design range for normal
> > > > > RCU grace-period durations on busy systems. See the code in
> > > > > adjust_jiffies_till_sched_qs(), which is setting one of the "panic
> > > > > durations" at which RCU starts taking more aggressive actions to end
> > > > > the current grace period. See especially:
> > > > >
> > > > > if (j < HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV)
> > > > > j = HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
> > > > > pr_info("RCU calculated value of scheduler-enlistment delay is %ld jiffies.\n", j);
> > > > > WRITE_ONCE(jiffies_to_sched_qs, j);
> > > > >
> > > > > This usually gets you about 100 milliseconds, and if you are starting
> > > > > grace periods in quick succession from a single thread while other threads
> > > > > are doing likewise, each grace-period wait gets to wait about two grace
> > > > > periods worth due to the end of the previous grace period having started
> > > > > a new grace period before the thread is awakened.
> > > > >
> > > > > Of course, if this is causing trouble for some use case, it would not
> > > > > be hard to create a tunable to override this panic duration. But that
> > > > > would of course require a real use case in real use, given that RCU isn't
> > > > > exactly short on tunables at the moment. Significantly shortening this
> > > > > panic duration caused 0day to complain about slowness last I tried it,
> > > > > just so you know.
> > > >
> > > > Thanks a lot for the explanation.
> > > > Indeed this code in the tick is doing a good job and I just had to drop
> > > > jiffies_till_first_fqs to bring down the latencies. With a
> > > > jiffies_till_first_fqs of 50 instead of the default of 100, the latencies
> > > > drop by 4 fold.
> > >
> > > You lost me on this one. The normal value of jiffies_till_first_fqs
> > > is but three, for systems with 255 or fewer CPUs and HZ=1000. So I
> > > have to ask... What did you do to get jiffies_till_first_fqs=100?
> > > The normal default automatic settings would need something like 8,000
> > > CPUs to get it up to that level.
> > >
> > > Or did you instead mean replacing the "HZ / 10" in the code snippet
> > > above with "HZ / 20" or similar?
> > >
> >
> > I meant jiffies_to_sched_qs.
>
> Whew!!! ;-)
>
> > Without any changes, it is 100 jiffies on my
> > system. Setting rcutree.jiffies_till_sched_qs sets the jiffies_to_sched_qs. I
> > had set it to 50 and observed dramatic improvements.
> >
> > /* If jiffies_till_sched_qs was specified, respect the request. */
> > if (jiffies_till_sched_qs != ULONG_MAX) {
> > WRITE_ONCE(jiffies_to_sched_qs, jiffies_till_sched_qs);
> > return;
> > }
> >
> > > Or did you mean jiffies_to_sched_qs instead of jiffies_till_first_fqs?
> > > -That- does default to 100, and you could set it using the
> > > rcutree.jiffies_to_sched_qs kernel boot parameter. But even then, I
> > > must admit that I would naively expect halving jiffies_till_first_fqs to
> > > halve the latencies. But I have not looked at it closely, and there are
> > > lots of moving parts in RCU's grace-period encouragement code, so maybe
> > > that is the effect.
> >
> > It could also be my sloppy testing. Now I tried again with 50 and it cuts the
> > latencies by around half as you said. However my histogram does have several
> > really nasty outliers..
>
> OK, that is more what I would expect.
>
> > (rcu-kvm is my wrapper where I pass the -net qemu args I need)
> >
> > rcu-kvm --rcuperf --boot-args "rcuperf.pd_test=1 rcuperf.pd_busy_wait=5000
> > rcuperf.holdout=5 rcuperf.pd_resched=0 rcutree.jiffies_till_sched_qs=50"
> > --kvm-args "--duration 1
> >
> > Log says:
> > 0.087440] rcu: Boot-time adjustment of scheduler-enlistment delay to 50 jiffies.
> >
> > Output:
> > Histogram bucket size: 1000
> > 57000 4
> > 58000 1
> > 59000 1
> > 60000 2
> > 103000 1
> > 104000 2
> > 105000 8
> > 106000 44
> > 107000 60
> > 108000 131
> > 109000 164
> > 110000 143 <---------- most of the time its ~100ms.
> > 111000 136
> > 112000 51
> > 113000 45
> > 114000 11
> > 115000 4
> > 12464000 1
> > 12466000 2 <--- But what are these :(
> > 12467000 2
> > 12468000 1
> > 12470000 1
>
> Well, those are a bit over one second, which is when .b.need_qs is set
> in CONFIG_PREEMPT=y systems. Are you possibly seeing vCPU preeemption?
> (Not that .b.need_qs does anything about vCPU preemption just yet.)
Actually, I just realized, there is an extra 0 on those outliers. So it is
really 12 seconds, for example: 12,466,000 microseconds is ~12.4 seconds. So
these outliers are really odd since I don't see any RCU stalls, possible theories:
Looking closer into the logs, it is always the last iteration of a writer:
Third last iteration..
[ 62.157951] rcu-perf: 0 writer-duration: 99 109999933
[ 62.258131] rcu-perf: 1 writer-duration: 99 110999466
[ 62.353607] rcu-perf: 2 writer-duration: 99 112000830
[ 62.433249] rcu-perf: 3 writer-duration: 99 114000321
[ 62.510405] rcu-perf: 4 writer-duration: 99 109000240
[ 62.603643] rcu-perf: 5 writer-duration: 99 108999164
[ 62.702109] rcu-perf: 6 writer-duration: 99 111000721
[ 62.799970] rcu-perf: 7 writer-duration: 99 109000536
Second last iteration..
[ 62.158920] rcu-perf: 0 writer-duration: 100 109998415
[ 62.259061] rcu-perf: 1 writer-duration: 100 104024906
[ 62.354395] rcu-perf: 2 writer-duration: 100 105019175
[ 62.434072] rcu-perf: 3 writer-duration: 100 108019926
[ 62.511154] rcu-perf: 4 writer-duration: 100 109998839
[ 62.604572] rcu-perf: 5 writer-duration: 100 109000969
[ 62.703005] rcu-perf: 6 writer-duration: 100 107015416
[ 62.800894] rcu-perf: 7 writer-duration: 100 111018680
And last one which is over blown over..
[ 62.161123] rcu-perf: 0 writer-duration: 101 42665751175
[ 62.261115] rcu-perf: 1 writer-duration: 101 42693148470
[ 62.357093] rcu-perf: 2 writer-duration: 101 42692066685
[ 62.436059] rcu-perf: 3 writer-duration: 101 42692831737
[ 62.513063] rcu-perf: 4 writer-duration: 101 42693195036
[ 62.705105] rcu-perf: 6 writer-duration: 101 42692114079
[ 62.803104] rcu-perf: 7 writer-duration: 101 42693153435
I am wondering if this because the torture_stop() is having a hard time
stopping my preempt disable thread, which is odd because I am checking for
torture_must_stop() to break out of the loop as are other threads.
Anyway this is clearly more of a test issue than an RCU one ;-)
> > Average grace-period duration: 215642 microseconds <-- avg ended up skewed
> > Minimum grace-period duration: 57979.7
> > 50th percentile grace-period duration: 110000
> > 90th percentile grace-period duration: 112999 <-- but this reduced.
> > 99th percentile grace-period duration: 115000
>
> Indeed, medians and percentiles are often more stable than are averages.
> (But yes, you can cook up distributions that work the other way around.)
>
> > (And just did another test and the outliers are pretty consistent).
> >
> > Also, I do occassionally see that preempt disable test not getting its Need
> > resched bit set yet. Even though we have multiple ticks. I am trying to trace
> > this out and understand it as well.
>
> That might also be a sign that the RCU grace-period kthread is being
> starved for short time periods. One way to check this is to use the
> rcutree.kthread_prio kernel boot parameter to set RCU's kthreads to
> some real-time priority. This is not something you usually want to
> do because it can greatly increase the context-switch rate, but as an
> experiment it might provide useful clues.
Tried and did not change the result of outliers but that's just because it is
likely not an RCU issue.
thanks,
- Joel
