Re: kernel performance: 3.16 vs. 4.9
Am 06.02.2017 um 00:12 schrieb Ben Caradoc-Davies:
> On 06/02/17 09:59, Matthias Bodenbinder wrote:
>> Am 05.02.2017 um 09:03 schrieb Matthias Bodenbinder:
>>> Hi,
>>> I have upgraded my PC to newest chipset and CPU: Kaby Lake, Z270 with i7-7700K.
>>> Out of curiosity I did a kernel benchmark. Comparing darktable performance with kernel 3.16.0-4-amd64 and 4.9.0-1-amd64. I use the following command to run darktable:
>>> darktable-cli test.CR2 test.jpg --core -d perf -d opencl
>>> The results are surprising for me. Kernel 4.9 very much outperformance kernel 3.16. Here are the results with and without opencl (using a Geforce GTX750TI):
>>> kernel 3.16 kernel 4.9
>>> with opencl 16 s 9 s
>>> without opencl 120 s 23 s
>>> Without opencl, that is with pure CPU performance, the difference is a factor of 5!
>>> Why is that? What am I missing? I can hardly believe that kernel 4.9 is so much faster.
>>> Matthias
>> Hi,
>> the issue is solved. It is related to the CPU frequency driver acpi-cpufreq vs. intel_pstate and the governor powersave vs.
>> performance.
>> With kernel 4.9 the default CPU frequency driver for my hardware is intel_pstate. And I had the governor set to "powersave" in /etc/default/cpufrequtils. I do not see big performance differences between "powersave" and "performance" settings with intel_pstate.
>> But with kernel 3.16 the default CPU frequency driver is acpi-cpufreq. With my settings in /etc/default/cpufrequtils it running in "powersave" mode which is giving this dramatic performance collapse. With governor "ondemand" or "performance" kernel 3.16 is as fast as kernel 4.9.
>> Sorry for the confusion.
>> Matthias
>
> Matthias that is very interesting. Your findings seem different to those reported on Phoronix; no doubt the p-state driver is undergoing some changes to improve Kaby Lake support:
> http://www.phoronix.com/scan.php?page=article&item=pstate-cpufreq-kbl&num=1
> http://www.phoronix.com/scan.php?page=article&item=pstate-cpufreq-kbl&num=2
> https://www.phoronix.com/scan.php?page=news_item&px=P-State-Kabylake-Patching
>
> Kind regards,
>
The phoronix article only compares the performance governors. It does not investigate the performance decrease with acpi-cpufreq+powersave.
Based on the articles I made another test with kernel 4.9 + acpi-cpufreq + performance and compared it to kernel 4.9 + intel_pstate + performance. The difference after several runs is minimal 23-24 s for intel_pstate vs. 22 s for acpi-cpufreq. Not sure if that is meaningful but it seems to support what phoronix is reporting.
Matthias
Reply to: