Re: Commit 554c8aa8ecad causing severe performance degression with pcc-cpufreq

[Rafael J. Wysocki]

On Tue, Jul 17, 2018 at 10:50 AM, Andreas Herrmann <aherrmann@xxxxxxxx> wrote:
> On Tue, Jul 17, 2018 at 10:03:41AM +0200, Rafael J. Wysocki wrote:
>> On Tue, Jul 17, 2018 at 9:33 AM, Rafael J. Wysocki <rafael@xxxxxxxxxx> wrote:
>> > Hi,
>> >
>> > Thanks for your report!
>> >
>> > On Tue, Jul 17, 2018 at 8:50 AM, Andreas Herrmann <aherrmann@xxxxxxxx> wrote:
>> >> Hello,
>> >>
>> >> I've recently noticed that commit 554c8aa8ecad ("sched: idle: Select
>> >> idle state before stopping the tick") causes severe performance drop
>> >> for systems using pcc-cpufreq driver. Depending on the number of CPUs
>> >> the system might be almost unusable. The OS jitter for 4.17.y and
>> >> 4.18.-rcx kernels is off the charts, you can even spot it with top
>> >> command (issued when the system is supposedly idle), e.g.
>> >>
>> >>  top - 14:44:24 up 2 min,  1 user,  load average: 90.11, 38.20, 14.38
>> >>  Tasks: 1199 total, 109 running, 541 sleeping,   0 stopped,   0 zombie
>> >>  %Cpu(s):  1.2 us, 58.7 sy,  0.0 ni, 39.3 id,  0.6 wa,  0.0 hi,  0.3 si,  0.0 st
>> >>  KiB Mem:  13137064+total,  1192168 used, 13017848+free,     2340 buffers
>> >>  KiB Swap:  2104316 total,        0 used,  2104316 free.   522296 cached Mem
>> >>
>> >>    PID USER      PR  NI    VIRT    RES    SHR S    %CPU  %MEM     TIME+ COMMAND
>> >>   3373 root      20   0  982024  49916  36120 R  96.691 0.038   0:19.54 kubelet
>> >>     67 root      20   0       0      0      0 R  78.676 0.000   0:49.36 kworker/9:0
>> >>     25 root      20   0       0      0      0 R  78.125 0.000   0:49.67 kworker/2:0
>> >>    182 root      20   0       0      0      0 R  75.735 0.000   1:18.17 kworker/28:0
>> >>     43 root      20   0       0      0      0 R  75.000 0.000   0:11.56 kworker/5:0
>> >>    103 root      20   0       0      0      0 R  74.449 0.000   0:46.83 kworker/15:0
>> >>    334 root      20   0       0      0      0 R  72.978 0.000   1:06.88 kworker/53:0
>> >>    789 root      20   0       0      0      0 R  69.853 0.000   1:29.50 kworker/38:1
>> >>    418 root      20   0       0      0      0 R  69.301 0.000   0:41.33 kworker/67:0
>> >>    779 root      20   0       0      0      0 R  68.934 0.000   1:33.60 kworker/27:1
>> >>    773 root      20   0       0      0      0 R  68.566 0.000   1:37.91 kworker/22:1
>> >>    762 root      20   0       0      0      0 R  68.015 0.000   1:41.01 kworker/11:1
>> >>    769 root      20   0       0      0      0 R  67.647 0.000   1:37.65 kworker/18:1
>> >>    805 root      20   0       0      0      0 R  67.096 0.000   1:30.96 kworker/54:1
>> >>    840 root      20   0       0      0      0 R  66.912 0.000   1:23.82 kworker/89:1
>> >>    812 root      20   0       0      0      0 R  66.728 0.000   1:31.89 kworker/59:1
>> >>    847 root      20   0       0      0      0 R  66.360 0.000   1:28.40 kworker/96:1
>> >>    763 root      20   0       0      0      0 R  66.176 0.000   1:42.57 kworker/12:1
>> >>    772 root      20   0       0      0      0 R  66.176 0.000   1:12.58 kworker/21:1
>> >>    821 root      20   0       0      0      0 R  66.176 0.000   1:29.62 kworker/69:1
>> >>    923 root      20   0       0      0      0 R  65.809 0.000   1:44.32 kworker/3:18
>> >>   1284 root      20   0       0      0      0 R  65.809 0.000   1:23.50 kworker/101:2
>> >>     61 root      20   0       0      0      0 R  65.625 0.000   1:29.37 kworker/8:0
>> >>   3531 root      20   0   24384   3768   2356 R  65.625 0.003   0:08.91 top
>> >>    771 root      20   0       0      0      0 R  65.074 0.000   1:37.90 kworker/20:1
>> >>    767 root      20   0       0      0      0 R  64.706 0.000   1:38.01 kworker/16:1
>> >>    764 root      20   0       0      0      0 R  64.522 0.000   1:40.28 kworker/13:1
>> >>    765 root      20   0       0      0      0 R  64.154 0.000   1:40.13 kworker/14:1
>> >>
>> >> When I apply below patch (trying to revert essential parts of commit
>> >> 554c8aa8ecad) behaviour seems back to normal.
>> >
>> > Well, that basically defeats the purpose of the change in commit
>> > 554c8aa8ecad, so it's not what I'd like to do to fix this problem.
>> >
>> > Also it would be good to understand what actually happens.
>> >
>> >> I know that pcc-cpufreq driver is not "state-of-the-art" when it comes
>> >> to cpufreq drivers and you better not use it.
>> >
>> > That's exactly right.
>> >
>> >> But I wonder whether commit 554c8aa8ecad ("sched: idle: Select idle state before
>> >> stopping the tick") introduced bad behaviour for other cases as well.
>> >
>> > It has been tested quite extensively in that respect, although
>> > admittedly not with the pcc-cpufreq driver.
>> >
>> > Nothing bad related to it has been has been reported so far, FWIW.
>> >
>> >> I'll send some performance results to illustrate the issue asap. I've
>> >> also tried to modify pcc-cpufreq to reduce the amount of frequency
>> >> changes triggered by this driver but this does not help for kernels
>> >> where commit 554c8aa8ecad is applied.
>> >
>> > Can you replace pcc-cpufreq with a different cpufreq driver on the
>> > affected systems?  If so, do performance numbers look bad after that
>> > too?
>>
>> Also, what cpufreq governor do you use with pcc-cpufreq?
>
> Ondemand governor. Which triggers a lot of PCC related platform calls.
> And as Peter noticed already the driver has a severe bottleneck (lock
> protecting shared memory used for all CPUs to pass data to/from
> platform for PCC calls).
>
>> Does changing it to something like "performance" improve things?
>
> With performance governor above mentioned bottleneck is no issue.

OK

> On balance before this commit users could use pcc-cpufreq but had
> already suboptimal performance (compared to say intel_pstate driver
> which can be used changing BIOS options). Starting with this commit
> systems using pcc-cpufreq are unusable with high number of CPUs (top
> output above is for system with 120 CPUs).

I see. :-)

> So should the driver be removed (sooner or later), or this behaviour
> be documented somewhere, or just leave it as is.

At least it should be documented in the driver that it is not scalable
and not for use on many-CPU systems (with "many" meaning anything
greater than 4 probably).

Or we could just make the driver not load if the number of CPUs in the
system is greater than 4 or similar.