Re: [LKP] [lkp] [sched/fair] 53d3bc773e: hackbench.throughput -32.9% regression
From: Huang\, Ying
Date: Wed Jun 01 2016 - 04:53:45 EST
Peter Zijlstra <peterz@xxxxxxxxxxxxx> writes:
> On Wed, Jun 01, 2016 at 01:00:10PM +0800, Huang, Ying wrote:
>> Hi, Peter,
>> Peter Zijlstra <peterz@xxxxxxxxxxxxx> writes:
>> > On Tue, May 31, 2016 at 04:34:36PM +0800, Huang, Ying wrote:
>> >> Hi, Ingo,
>> >> Part of the regression has been recovered in v4.7-rc1 from -32.9% to
>> >> -9.8%. But there is still some regression. Is it possible for fully
>> >> restore it?
>> > after much searching on how you guys run hackbench... I figured
>> > something like:
>> > perf bench sched messaging -g 20 --thread -l 60000
>> There is a reproduce file attached in the original report email, its
>> contents is something like below:
>> 2016-05-15 08:57:02 echo performance > /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor
> <snip stupid large output>
>> 2016-05-15 09:06:24 /usr/bin/hackbench -g 24 --threads -l 60000
>> Hope that will help you for reproduce.
> It did not, because I didn't have the exact same machine and its not
> apparent how I should modify -- if at all -- the arguments to be
> representative when ran on my machine.
>> > on my IVB-EP (2*10*2) is similar to your IVT thing.
>> > And running something like:
>> > for i in /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor ; do echo performance > $i ; done
>> > perf stat --null --repeat 10 -- perf bench sched messaging -g 20 --thread -l 60000 | grep "seconds time elapsed"
>> > gets me:
>> > v4.6:
>> > 36.786914089 seconds time elapsed ( +- 0.49% )
>> > 37.054017355 seconds time elapsed ( +- 1.05% )
>> > origin/master (v4.7-rc1-ish):
>> > 34.757435264 seconds time elapsed ( +- 3.34% )
>> > 35.396252515 seconds time elapsed ( +- 3.38% )
>> > Which doesn't show a regression between v4.6 and HEAD; in fact it shows
>> > an improvement.
>> Yes. For hackbench test, linus/master (v4.7-rc1+) is better than v4.6,
>> but it is worse than v4.6-rc7. Details is as below.
> That kernel was broken.. what your point?
You mean the commit is a functionality fix? I found the
hackbench.throughput for the test is
So some commit in v4.6-rc1 introduce some issue but improve the score
for the test, which is fixed in v4.6, and some improvement merged by