Re: [LKP] [lkp] [mm/vmstat] 6cdb18ad98: -8.5% will-it-scale.per_thread_ops

[Heiko Carstens]

On Fri, Jan 08, 2016 at 01:24:30PM +0800, Huang, Ying wrote:
> Heiko Carstens <heiko.carstens@xxxxxxxxxx> writes:
> 
> > On Wed, Jan 06, 2016 at 11:20:55AM +0800, kernel test robot wrote:
> >> FYI, we noticed the below changes on
> >> 
> >> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git master
> >> commit 6cdb18ad98a49f7e9b95d538a0614cde827404b8 ("mm/vmstat: fix overflow in mod_zone_page_state()")
> >> 
> >> 
> >> =========================================================================================
> >> compiler/cpufreq_governor/kconfig/rootfs/tbox_group/test/testcase:
> >>   gcc-4.9/performance/x86_64-rhel/debian-x86_64-2015-02-07.cgz/ivb42/pread1/will-it-scale
> >> 
> >> commit: 
> >>   cc28d6d80f6ab494b10f0e2ec949eacd610f66e3
> >>   6cdb18ad98a49f7e9b95d538a0614cde827404b8
> >> 
> >> cc28d6d80f6ab494 6cdb18ad98a49f7e9b95d538a0 
> >> ---------------- -------------------------- 
> >>          %stddev     %change         %stddev
> >>              \          |                \  
> >>    2733943   0%      -8.5%    2502129   0%  will-it-scale.per_thread_ops
> >>       3410   0%      -2.0%       3343   0%  will-it-scale.time.system_time
> >>     340.08   0%     +19.7%     406.99   0%  will-it-scale.time.user_time
> >>   69882822   2%     -24.3%   52926191   5%  cpuidle.C1-IVT.time
> >>     340.08   0%     +19.7%     406.99   0%  time.user_time
> >>     491.25   6%     -17.7%     404.25   7%  numa-vmstat.node0.nr_alloc_batch
> >>       2799  20%     -36.6%       1776   0%  numa-vmstat.node0.nr_mapped
> >>     630.00 140%    +244.4%       2169   1%  numa-vmstat.node1.nr_inactive_anon
> >
> > Hmm... this is odd. I did review all callers of mod_zone_page_state() and
> > couldn't find anything obvious that would go wrong after the int -> long
> > change.
> >
> > I also tried the "pread1_threads" test case from
> > https://github.com/antonblanchard/will-it-scale.git
> >
> > However the results seem to vary a lot after a reboot(!), at least on s390.
> >
> > So I'm not sure if this is really a regression.
> 
> The test is quite stable for my side.  We run the test case 7 times for
> your commit and its parent.  The standard variation is very low.
> 
> you commit:
> 
> [2493136, 2510964, 2508784, 2495632, 2506735, 2503016, 2510121]
> 
> parent commit:
> 
> [2735669, 2719566, 2739052, 2741485, 2735152, 2739356, 2739125]
> 
> The test result is stable for bisection too.  The below figure show the
> results of good commits and bad commits.  The distance between is quite
> big.  And the variation is quite small.

Ok, so it seems to be quite stable on your machine across reboots.

I have to admit I still cannot make much sense of this. Is the "pread1"
testcase the only one that performs worse, or are there more?

Also could you please provide the output of /proc/zoneinfo and the output
of "perf top" of the good/bad cases? Maybe that might help to figure out
what is happening.

> FYI, I test your patch on x86 platform.  I have no s390 system.

Sure, I wouldn't expect that.