Re: [RFC/RFT,v7] cpuidle: New timer events oriented governor for tickless systems

From: Quentin Perret
Date: Fri Dec 07 2018 - 04:38:03 EST


Hi Chen,

On Friday 07 Dec 2018 at 15:58:35 (+0800), Chen, Hu wrote:
> From: Chen Hu <hu1.chen@xxxxxxxxx>
>
> Hi Rafael,
>
> I run several popular Android performance benchmarks on teov7, using kernel
> 4.19.0 as my baseline because I happen to work on it. To backport teov7 to
> kernel 4.19.0, I also backport patch 5f26bdc: "cpuidle: menu: Fix wakeup
> statistics updates for polling state". The teov7 doesn't show regressions on
> such perf KPIs.
>
> Compare "4.19 + 5f26bdc" and "4.19 + 5f26bdc + teov7" on Android with Intel
> Apollo Lake SoC:
>
> Test Case Diff after appling teov7
> Antutu_6 0.51%
> GFX5_openGL_Car_Chase 0.43%
> GFX5_openGL_Car_Chase_offscreen 0.25%
> GFX5_openGL_Manhattan31 0.75%
> GFX5_openGL_Manhattan31_1080_offscreen -0.10%
> Geekbench3.3 -0.13%
> H264_1080P_60FPS 0.00%
> H264_2160P_60FPS 0.00%
> H265_2K_10bit 0.00%
> H265_2K_8bit 0.00%
> Resume_time 1.31%
> full_boot -0.39%

That's really nice that you guys are able to get good test results on
Android 4.19 that early after the release :-)

Do you happen to run interactive use-cases as well ? I think Jankbench
is one of the workload most folk care about for example. Also, did you
manage to measure energy numbers for the above tests ?

FWIW, I'd like to give a go to TEO on an Arm-based setup when I can.

Thanks,
Quentin