Re: [clocksource] 388450c708: netperf.Throughput_tps -65.1% regression

From: Feng Tang
Date: Sun May 16 2021 - 02:34:49 EST


On Fri, May 14, 2021 at 10:49:08AM -0700, Paul E. McKenney wrote:
> On Fri, May 14, 2021 at 03:43:14PM +0800, Feng Tang wrote:
> > Hi Paul,
> >
> > On Thu, May 13, 2021 at 10:07:07AM -0700, Paul E. McKenney wrote:
> > > On Thu, May 13, 2021 at 11:55:15PM +0800, kernel test robot wrote:
> > > >
> > > >
> > > > Greeting,
> > > >
> > > > FYI, we noticed a -65.1% regression of netperf.Throughput_tps due to commit:
> > > >
> > > >
> > > > commit: 388450c7081ded73432e2b7148c1bb9a0b039963 ("[PATCH v12 clocksource 4/5] clocksource: Reduce clocksource-skew threshold for TSC")
> > > > url: https://github.com/0day-ci/linux/commits/Paul-E-McKenney/Do-not-mark-clocks-unstable-due-to-delays-for-v5-13/20210501-083404
> > > > base: https://git.kernel.org/cgit/linux/kernel/git/tip/tip.git 2d036dfa5f10df9782f5278fc591d79d283c1fad
> > > >
> > > > in testcase: netperf
> > > > on test machine: 96 threads 2 sockets Ice Lake with 256G memory
> > > > with following parameters:
> > > >
> > > > ip: ipv4
> > > > runtime: 300s
> > > > nr_threads: 25%
> > > > cluster: cs-localhost
> > > > test: UDP_RR
> > > > cpufreq_governor: performance
> > > > ucode: 0xb000280
> > > >
> > > > test-description: Netperf is a benchmark that can be use to measure various aspect of networking performance.
> > > > test-url: http://www.netperf.org/netperf/
> > > >
> > > >
> > > >
> > > > If you fix the issue, kindly add following tag
> > > > Reported-by: kernel test robot <oliver.sang@xxxxxxxxx>
> > > >
> > > >
> > > > also as Feng Tang checked, this is a "unstable clocksource" case.
> > > > attached dmesg FYI.
> > >
> > > Agreed, given the clock-skew event and the resulting switch to HPET,
> > > performance regressions are expected behavior.
> > >
> > > That dmesg output does demonstrate the value of Feng Tang's patch!
> > >
> > > I don't see how to obtain the values of ->mult and ->shift that would
> > > allow me to compute the delta. So if you don't tell me otherwise, I
> > > will assume that the skew itself was expected on this hardware, perhaps
> > > somehow due to the tpm_tis_status warning immediately preceding the
> > > clock-skew event. If my assumption is incorrect, please let me know.
> >
> > I run the case with the debug patch applied, the info is:
> >
> > [ 13.796429] clocksource: timekeeping watchdog on CPU19: Marking clocksource 'tsc' as unstable because the skew is too large:
> > [ 13.797413] clocksource: 'hpet' wd_nesc: 505192062 wd_now: 10657158 wd_last: fac6f97 mask: ffffffff
> > [ 13.797413] clocksource: 'tsc' cs_nsec: 504008008 cs_now: 3445570292aa5 cs_last: 344551f0cad6f mask: ffffffffffffffff
> > [ 13.797413] clocksource: 'tsc' is current clocksource.
> > [ 13.797413] tsc: Marking TSC unstable due to clocksource watchdog
> > [ 13.844513] clocksource: Checking clocksource tsc synchronization from CPU 50 to CPUs 0-1,12,22,32-33,60,65.
> > [ 13.855080] clocksource: Switched to clocksource hpet
> >
> > So the delta is 1184 us (505192062 - 504008008), and I agree with
> > you that it should be related with the tpm_tis_status warning stuff.
> >
> > But this re-trigger my old concerns, that if the margins calculated
> > for tsc, hpet are too small?
>
> If the error really did disturb either tsc or hpet, then we really
> do not have a false positive, and nothing should change (aside from
> perhaps documenting that TPM issues can disturb the clocks, or better
> yet treating that perturbation as a separate bug that should be fixed).
> But if this is yet another way to get a confused measurement, then it
> would be better to work out a way to reject the confusion and keep the
> tighter margins. I cannot think right off of a way that this could
> cause measurement confusion, but you never know.

I have no doubt in the correctness of the measuring method, but was
just afraid some platforms which use to 'just work' will be caught :)

> So any thoughts on exactly how the tpm_tis_status warning might have
> resulted in the skew?

The tpm error message has been reported before, and from google there
were some similar errors, we'll do some further check.

> > With current math algorithm, the 'uncertainty_margin' is
> > calculated against the frequency, and those tsc/hpet/acpi_pm
> > timer is multiple of MHz or GHz, which gives them to have margin of
> > 100 us. It works with normal systems. But in the wild world, there
> > could be some sparkles due to some immature HW components, their
> > firmwares or drivers etc, just like this case.
>
> Isn't diagnosing issues from immature hardware, firmware, and drivers
> actually a benefit? It would after all be quite unfortunate if some issue
> that was visible only due to clock skew were to escape into production.

Yes, it surely will expose some mal-functional cases which haven't
been caught before.

Thanks,
Feng


> Thanx, Paul