[x86, vdso] cfda7bb9ecb: +14.7% will-it-scale.per_thread_ops

[x86, vdso] cfda7bb9ecb: +14.7% will-it-scale.per_thread_ops

[Jet Chen]

[-- Attachment #1: Type: text/plain, Size: 977 bytes --]

Hi Andy,

FYI, we noticed the below changes on

git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git x86/vdso
commit cfda7bb9ecbf9d96264bb5bade33a842966d1062 ("x86, vdso: Move syscall and sysenter setup into kernel/cpu/common.c")

test case: nhm4/will-it-scale/sched_yield

3d7ee969bffcc98  cfda7bb9ecbf9d96264bb5bad
---------------  -------------------------
     5497021 ~ 0%     +14.7%    6303424 ~ 0%  TOTAL will-it-scale.per_thread_ops
        0.54 ~ 0%      +5.6%       0.57 ~ 0%  TOTAL will-it-scale.scalability
     6209483 ~ 0%      +1.6%    6305917 ~ 0%  TOTAL will-it-scale.per_process_ops
        2455 ~ 5%     +16.9%       2870 ~ 5%  TOTAL cpuidle.C1-NHM.usage
        8829 ~ 7%     +15.2%      10169 ~10%  TOTAL slabinfo.kmalloc-64.active_objs
       24.13 ~12%     +48.9%      35.93 ~14%  TOTAL time.user_time
         393 ~ 0%      -3.0%        382 ~ 1%  TOTAL time.system_time


Legend:
	~XX%    - stddev percent
	[+-]XX% - change percent


Thanks,
Jet



[-- Attachment #2: reproduce --]
[-- Type: text/plain, Size: 622 bytes --]

echo performance > /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor
echo performance > /sys/devices/system/cpu/cpu1/cpufreq/scaling_governor
echo performance > /sys/devices/system/cpu/cpu2/cpufreq/scaling_governor
echo performance > /sys/devices/system/cpu/cpu3/cpufreq/scaling_governor
echo performance > /sys/devices/system/cpu/cpu4/cpufreq/scaling_governor
echo performance > /sys/devices/system/cpu/cpu5/cpufreq/scaling_governor
echo performance > /sys/devices/system/cpu/cpu6/cpufreq/scaling_governor
echo performance > /sys/devices/system/cpu/cpu7/cpufreq/scaling_governor
./runtest.py sched_yield 32 1 4 6 8

Re: [x86, vdso] cfda7bb9ecb: +14.7% will-it-scale.per_thread_ops

[Andy Lutomirski]

On Mon, May 19, 2014 at 10:59 PM, Jet Chen <jet.chen@intel.com> wrote:
> Hi Andy,
>
> FYI, we noticed the below changes on
>
> git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git x86/vdso
> commit cfda7bb9ecbf9d96264bb5bade33a842966d1062 ("x86, vdso: Move syscall
> and sysenter setup into kernel/cpu/common.c")

This is really strange.  That change shouldn't have any effect except
that some kernels might boot a cycle or two faster.

>
> test case: nhm4/will-it-scale/sched_yield
>
> 3d7ee969bffcc98  cfda7bb9ecbf9d96264bb5bad
> ---------------  -------------------------
>     5497021 ~ 0%     +14.7%    6303424 ~ 0%  TOTAL
> will-it-scale.per_thread_ops
>        0.54 ~ 0%      +5.6%       0.57 ~ 0%  TOTAL will-it-scale.scalability
>     6209483 ~ 0%      +1.6%    6305917 ~ 0%  TOTAL
> will-it-scale.per_process_ops
>        2455 ~ 5%     +16.9%       2870 ~ 5%  TOTAL cpuidle.C1-NHM.usage
>        8829 ~ 7%     +15.2%      10169 ~10%  TOTAL
> slabinfo.kmalloc-64.active_objs
>       24.13 ~12%     +48.9%      35.93 ~14%  TOTAL time.user_time
>         393 ~ 0%      -3.0%        382 ~ 1%  TOTAL time.system_time
>

Is this a speedup or a slowdown?

--Andy

Re: [x86, vdso] cfda7bb9ecb: +14.7% will-it-scale.per_thread_ops

[Fengguang Wu]

Hi Andy,

On Wed, May 21, 2014 at 03:46:58PM -0700, Andy Lutomirski wrote:
> On Mon, May 19, 2014 at 10:59 PM, Jet Chen <jet.chen@intel.com> wrote:
> > Hi Andy,
> >
> > FYI, we noticed the below changes on
> >
> > git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git x86/vdso
> > commit cfda7bb9ecbf9d96264bb5bade33a842966d1062 ("x86, vdso: Move syscall
> > and sysenter setup into kernel/cpu/common.c")
> 
> This is really strange.  That change shouldn't have any effect except
> that some kernels might boot a cycle or two faster.
> 
> >
> > test case: nhm4/will-it-scale/sched_yield
> >
> > 3d7ee969bffcc98  cfda7bb9ecbf9d96264bb5bad
> > ---------------  -------------------------
> >     5497021 ~ 0%     +14.7%    6303424 ~ 0%  TOTAL
> > will-it-scale.per_thread_ops
> >        0.54 ~ 0%      +5.6%       0.57 ~ 0%  TOTAL will-it-scale.scalability
> >     6209483 ~ 0%      +1.6%    6305917 ~ 0%  TOTAL
> > will-it-scale.per_process_ops
> >        2455 ~ 5%     +16.9%       2870 ~ 5%  TOTAL cpuidle.C1-NHM.usage
> >        8829 ~ 7%     +15.2%      10169 ~10%  TOTAL
> > slabinfo.kmalloc-64.active_objs
> >       24.13 ~12%     +48.9%      35.93 ~14%  TOTAL time.user_time
> >         393 ~ 0%      -3.0%        382 ~ 1%  TOTAL time.system_time
> >
> 
> Is this a speedup or a slowdown?

It's a speedup. The will-it-scale/sched_yield test case's throughput
increased by +14.7% (multi-thread case) and +1.6% (multi-process case).

However the CPU %user time increased more, by +48.9%.

Thanks,
Fengguang

Re: [x86, vdso] cfda7bb9ecb: +14.7% will-it-scale.per_thread_ops

[H. Peter Anvin]

On 05/21/2014 06:54 PM, Fengguang Wu wrote:
>>>
>>> test case: nhm4/will-it-scale/sched_yield
>>>
>>> 3d7ee969bffcc98  cfda7bb9ecbf9d96264bb5bad
>>> ---------------  -------------------------
>>>     5497021 ~ 0%     +14.7%    6303424 ~ 0%  TOTAL
>>> will-it-scale.per_thread_ops
>>>        0.54 ~ 0%      +5.6%       0.57 ~ 0%  TOTAL will-it-scale.scalability
>>>     6209483 ~ 0%      +1.6%    6305917 ~ 0%  TOTAL
>>> will-it-scale.per_process_ops
>>>        2455 ~ 5%     +16.9%       2870 ~ 5%  TOTAL cpuidle.C1-NHM.usage
>>>        8829 ~ 7%     +15.2%      10169 ~10%  TOTAL
>>> slabinfo.kmalloc-64.active_objs
>>>       24.13 ~12%     +48.9%      35.93 ~14%  TOTAL time.user_time
>>>         393 ~ 0%      -3.0%        382 ~ 1%  TOTAL time.system_time
>>>
>>
>> Is this a speedup or a slowdown?
> 
> It's a speedup. The will-it-scale/sched_yield test case's throughput
> increased by +14.7% (multi-thread case) and +1.6% (multi-process case).
> 
> However the CPU %user time increased more, by +48.9%.
> 

That would be consistent with spending less time in the kernel, no?

But I agree... that is completely bizarre.  That checkin should have
absolutely zero effect on performance.

	-hpa