Hi Jeremy,
I think the reason why we observe the performance degradation after
applying wastedcores patches is this issue:
With HT turned OFF we see that wastedcores kernel compares as following to
mainline kernel
- performance regression in order of 1-5% for SPECjbb2005 and for
Linpack and Stream
- SPECjvm shows some minor improvements for some subtests but also some
regressions for other subtests
I think we will need to test the four proposed "wastedcores" patches one by
one.
Could you please prepare 4 different kernels where always just one patch
from
will be applied?
Jirka
On Tue, May 3, 2016 at 11:43 PM, Jirka Hladky <jhladky(a)redhat.com> wrote:
Hi Jeremy,
please have a look at Linpack and Stream results
http://perf-desktop.brq.redhat.com/LinpackAndStream/Red-Hat-Enterprise-Li...
and
http://perf-desktop.brq.redhat.com/LinpackAndStream/Red-Hat-Enterprise-Li...
We have collected /proc/sched_debug every 10 seconds - there are
""*scheddebug" files for every run. We are also collecting mpstat
statistics every 2 seconds.
For example during a test when 64 stream jobs were running in parallel
there are corresponding /proc/sched_debug logs:
4.6.0-0.rc5.git0.1.el7.x86_64/throughput-performance/Results-2016-Apr-28_14h01m32s/RawData/stream.affinityRun.064streams.loop01.scheddebug
<
http://perf-desktop.brq.redhat.com/LinpackAndStream/Red-Hat-Enterprise-Li...
4.6.0-0.rc5.git0.1.wastedcores.el7.x86_64/throughput-performance/Results-2016-Apr-29_15h37m17s/RawData/stream.affinityRun.064streams.loop01.scheddebug
<
http://perf-desktop.brq.redhat.com/LinpackAndStream/Red-Hat-Enterprise-Li...
However, we are not sure how to process these log files. Is there any tool
to do so or could you share some hints what to look for?
Thanks a lot!
Jirka
On Tue, May 3, 2016 at 4:18 PM, Jeremy Eder <jeder(a)redhat.com> wrote:
> Yikes. 20% off...why? Do you have any system stats? Particularly
> /proc/sched_debug collected at some interval.
>
> On Mon, May 2, 2016 at 12:39 PM, Jirka Hladky <jhladky(a)redhat.com> wrote:
>
>> Hi all,
>>
>> we have compared scheduler performance
>> for kernel-4.6.0-0.rc5.git0.1.wastedcores
>> against kernel-4.6.0-0.rc5.git0.1.el7.
>>
>> *Summary: *against all expectations* kernel-4.6.0-0.rc5.git0.1.wastedcores
performs on
>> average 20% WORSE than the regular kernel. *
>>
>> Results for different servers:
>>
>> SPECjbb2005:
>>
>> - karkulka-02.lab.eng.brq.redhat.com-htON
>>
<
http://perf-desktop.brq.redhat.com/SPECjbb/.vs./karkulka-02.lab.eng.brq.r...
>> - kiff-01.cluster-qe.lab.eng.brq.redhat.com-htON
>>
<
http://perf-desktop.brq.redhat.com/SPECjbb/.vs./kiff-01.cluster-qe.lab.en...
>>
>> SPECjvm2008:
>>
>> - karkulka-02.lab.eng.brq.redhat.com-htON
>>
<
http://perf-desktop.brq.redhat.com/SPECjvm/.vs./karkulka-02.lab.eng.brq.r...
>> - kiff-01.cluster-qe.lab.eng.brq.redhat.com-htON
>>
<
http://perf-desktop.brq.redhat.com/SPECjvm/.vs./kiff-01.cluster-qe.lab.en...
>>
>> Linpack and Stream:
>>
>> - karkulka-02.lab.eng.brq.redhat.com-htON
>>
<
http://perf-desktop.brq.redhat.com/LinpackAndStream/.vs./karkulka-02.lab....
>> - kiff-01.cluster-qe.lab.eng.brq.redhat.com-htON
>>
<
http://perf-desktop.brq.redhat.com/LinpackAndStream/.vs./kiff-01.cluster-...
>>
>> Any comments, questions, suggestions?
>>
>> Jirka and Kamil
>>
>> On Thu, Apr 28, 2016 at 12:25 AM, Jirka Hladky <jhladky(a)redhat.com>
>> wrote:
>>
>>> Hi Jeremy,
>>>
>>> I have just read
>>>
>>>
http://www.i3s.unice.fr/~jplozi/wastedcores/files/extended_talk.pdf
>>>
http://www.ece.ubc.ca/~sasha/papers/eurosys16-final29.pdf
>>>
>>> We will definitely test it and share the results here.
>>>
>>> Jirka
>>>
>>>
>>>
>
>
> --
>
> -- Jeremy Eder
>