_______________________________________________
LKP mailing list
[email protected]
On Thu, Nov 6, 2014 at 10:00 PM, LKP <[email protected]> wrote:
> FYI, we noticed the below changes on
>
> git://git.kernel.org/pub/scm/linux/kernel/git/luto/linux.git x86/entry
> commit f04e05b81e4d9ae88bee379f12176f551a24394a ("x86_64,entry: Use sysret to return to userspace when possible")
>
>
> f2ee9bebb99dd4a4 f04e05b81e4d9ae88bee379f12 testbox/testcase/testparams
> ---------------- --------------------------
> %stddev %change %stddev
> \ | \
> 24.09 ± 4% -62.7% 8.99 ± 3% lkp-wsx02/aim9/performance-300s-creat-clo
> 24.09 -62.7% 8.99 GEO-MEAN time.user_time
>
I don't really know how to interpret this. Is this a good thing or a bad thing?
I actually expect this change to be a dramatic speedup for some
workloads. Any chance you can send me the .config, kernel command
line, and whether you have any tracing features enabled?
--Andy
On Thu, Nov 06, 2014 at 10:26:23PM -0800, Andy Lutomirski wrote:
> On Thu, Nov 6, 2014 at 10:00 PM, LKP <[email protected]> wrote:
> > FYI, we noticed the below changes on
> >
> > git://git.kernel.org/pub/scm/linux/kernel/git/luto/linux.git x86/entry
> > commit f04e05b81e4d9ae88bee379f12176f551a24394a ("x86_64,entry: Use sysret to return to userspace when possible")
> >
> >
> > f2ee9bebb99dd4a4 f04e05b81e4d9ae88bee379f12 testbox/testcase/testparams
> > ---------------- --------------------------
> > %stddev %change %stddev
> > \ | \
> > 24.09 ? 4% -62.7% 8.99 ? 3% lkp-wsx02/aim9/performance-300s-creat-clo
> > 24.09 -62.7% 8.99 GEO-MEAN time.user_time
> >
>
> I don't really know how to interpret this. Is this a good thing or a bad thing?
TBH, I don't, either. Here we just list the changes we captured, which
is then for your information to decide it's good or bad.
>
> I actually expect this change to be a dramatic speedup for some
> workloads.
Would you like to share your workload? If so, I can add them to our
testing system, and give feedback to you when datat is available.
> Any chance you can send me the .config, kernel command
> line, and whether you have any tracing features enabled?
.config and dmesg are attached.
--yliu
On Thu, Nov 6, 2014 at 10:40 PM, Yuanhan Liu
<[email protected]> wrote:
> On Thu, Nov 06, 2014 at 10:26:23PM -0800, Andy Lutomirski wrote:
>> On Thu, Nov 6, 2014 at 10:00 PM, LKP <[email protected]> wrote:
>> > FYI, we noticed the below changes on
>> >
>> > git://git.kernel.org/pub/scm/linux/kernel/git/luto/linux.git x86/entry
>> > commit f04e05b81e4d9ae88bee379f12176f551a24394a ("x86_64,entry: Use sysret to return to userspace when possible")
>> >
>> >
>> > f2ee9bebb99dd4a4 f04e05b81e4d9ae88bee379f12 testbox/testcase/testparams
>> > ---------------- --------------------------
>> > %stddev %change %stddev
>> > \ | \
>> > 24.09 ą 4% -62.7% 8.99 ą 3% lkp-wsx02/aim9/performance-300s-creat-clo
>> > 24.09 -62.7% 8.99 GEO-MEAN time.user_time
>> >
>>
>> I don't really know how to interpret this. Is this a good thing or a bad thing?
>
> TBH, I don't, either. Here we just list the changes we captured, which
> is then for your information to decide it's good or bad.
>
>>
>> I actually expect this change to be a dramatic speedup for some
>> workloads.
>
> Would you like to share your workload? If so, I can add them to our
> testing system, and give feedback to you when datat is available.
My workload is just syscall performance under different conditions.
There's no obvious interesting benchmark, although I may write one at
some point.
>
>> Any chance you can send me the .config, kernel command
>> line, and whether you have any tracing features enabled?
>
> .config and dmesg are attached.
Hmm. I'm not entirely sure what's going on. I think that you have
NO_HZ_FULL active, but I can't see what's actually activating it. My
initial attempt to replicate this failed, but I see a big performance
gain if CONFIG_CONTEXT_TRACKING_FORCE=y, so I suspect that you've just
managed to activate context tracking without forcing it.
--Andy