Testing For The Latest Linux Kernel Power Regression

[phoronix]

Phoronix: Testing For The Latest Linux Kernel Power Regression

As I wrote about yesterday, there appears to be a new Linux kernel power regression that's yet to be solved by the latest Linux 3.17 code. The issue was originally tracked down to being a regression introduced during the Linux 3.15 stable cycle that disabled frame-buffer compression support by default for the Intel DRM graphics driver, but the impact it's had on the system power draw is much greater than what was anticipated by the Intel developers. A separate Intel employee is also reporting increased power draw, so I decided to run some tests on a few local systems to see what I'm encountering in the power consumption primarily between Linux 3.15 and 3.16.

http://www.phoronix.com/vr.php?view=20821

[saski]

Finally! Thank you Michael for bringing up this issue! That confirms my experience with mobile HSW over the last 9 Months

[smitty3268]

This testing scheme of Michael's is pretty bizarre

How can you still not have the most important test of all? A sitting idle doing nothing test.

Trying to test what power does during other benchmark runs can be interesting, i guess, but it brings in so many other potential factors and typically won't find anything anyway because it swamps out any real regressions with noise.

What everyone cares about is idle power usage. Surely it'd be easy to add a test which does nothing...

[liam]

How can you still not have the most important test of all? A sitting idle doing nothing test.

Trying to test what power does during other benchmark runs can be interesting, i guess, but it brings in so many other potential factors and typically won't find anything anyway because it swamps out any real regressions with noise.

What everyone cares about is idle power usage. Surely it'd be easy to add a test which does nothing...

He included that on page four, iirc. He idled the system for three minutes.
You are right, though. Idling is the most important test. I wish he'd included the idle test on the high res zenbook. Also, he really needs to integrate those power curves to find the Joules used.

[saski]

He included that on page four, iirc. He idled the system for three minutes.
You are right, though. Idling is the most important test. I wish he'd included the idle test on the high res zenbook. Also, he really needs to integrate those power curves to find the Joules used.

I totally agree. Keeping the idle drain low is very important especially during business meetings with the windows laptops from my coleagues runing twice as long than my linux box with identical hardware.

Michael, you might want to take a closer look at the intel pstate driver. I suspect it as a major culprit (see changes between 3.14 and 3.15).

[BSDude]

I totally agree. Keeping the idle drain low is very important especially during business meetings with the windows laptops from my coleagues runing twice as long than my linux box with identical hardware.

Aww that just sucks. Now I wonder how much my laptop would hold while running windows. I wiped it as soon as I got it. Running Antergos I get 4-4.30 hours of life compared to 9 advertised by Lenovo.

[halo9en]

Michael, you might want to take a closer look at the intel pstate driver. I suspect it as a major culprit (see changes between 3.14 and 3.15).

I wonder if booting 3.17 with intel_pstate=disable makes a difference?

[saski]

I wonder if booting 3.17 with intel_pstate=disable makes a difference?

Yes and no. When idle the CPU core frequency stays at its lowest value constantly when using the on demand govenour. however pc states above pc2 are not entered. Power consumption stays with 14-16 Watts idle roughly the same between on demand and intel_pstate.

[Michael]

How can you still not have the most important test of all? A sitting idle doing nothing test.

Trying to test what power does during other benchmark runs can be interesting, i guess, but it brings in so many other potential factors and typically won't find anything anyway because it swamps out any real regressions with noise.

What everyone cares about is idle power usage. Surely it'd be easy to add a test which does nothing...

I did.... If you read all the pages of the article.

There is an 'idle' test profile - http://openbenchmarking.org/test/pts/idle

[smitty3268]

I did.... If you read all the pages of the article.

There is an 'idle' test profile - http://openbenchmarking.org/test/pts/idle

Whoops, sorry about that.

I actually did read through the whole thing, but somehow skipped over that one.

In my defense, that should really be the highlighted test, rather than one buried deep down.

But thanks for making sure you do have that test in there somewhere.