Kernel 3.6.2 Power Regressions (Sandy Bridge)

PurpleAlert · 2013-08-07 13:35:22

Well, it look like the fix "drm/i915: fix long-standing SNB regression in power consumption after resume" made it into 3.10.5. Let's hope it works!

Druedain · 2013-08-15 09:48:23

In my case now it's much worse. After first suspend CPU runs at max and never comes back to normal state until restart.

arcon · 2013-08-15 10:05:58

Did anyone tried new microcode-20130808 from intel. If it is of some help.

Angizia · 2013-08-15 18:44:16

How do I do this? The wiki doesn't help
Where should I put the microcode.dat file?
THX

bobzxr · 2013-08-15 19:42:36

I did some cpu stress testing today under Arch and Ubuntu with different kernels. These values are after 2 minutes stressing 2 cpu cores. (on a thinkpad X220 w/core i-5)

Same conditions under every test. Whats interesting are the temperatures under the lts kernel.

ubuntu 3.8.0-19-generic
Physical id 0: +87.0°C

ubuntu 3.5.0-23-generic
Physical id 0: +90.0°C

arch 3.10.6-2-ARCH
Physical id 0: +89.0°C

arch 3.0.90-1-lts
Physical id 0: +56.0°C

As you can see, the results are more or less the same, except for the lts kernel which ran the machine ~30C cooler. Using the lts kernel is really not an option cause it is slower and I have graphical glitches with it.

Maybe I will try the 3.4 lts kernel from kernel.org and see if it helps, but upgrading it all the time would be also a pain.

EDIT: these are NOT after suspend, I never use it. My machine *just* runs on high temperatures.

Last edited by bobzxr (2013-08-15 19:47:38)

KaiSforza · 2013-08-15 20:05:06

bobzxr wrote:

I did some cpu stress testing today under Arch and Ubuntu with different kernels. These values are after 2 minutes stressing 2 cpu cores. (on a thinkpad X220 w/core i-5)
Same conditions under every test. Whats interesting are the temperatures under the lts kernel.
ubuntu 3.8.0-19-generic
Physical id 0: +87.0°C
ubuntu 3.5.0-23-generic
Physical id 0: +90.0°C
arch 3.10.6-2-ARCH
Physical id 0: +89.0°C
arch 3.0.90-1-lts
Physical id 0: +56.0°C
As you can see, the results are more or less the same, except for the lts kernel which ran the machine ~30C cooler. Using the lts kernel is really not an option cause it is slower and I have graphical glitches with it.
Maybe I will try the 3.4 lts kernel from kernel.org and see if it helps, but upgrading it all the time would be also a pain.

The regressions are not from stress tests, they are from the kernel not putting the GPU and CPU to sleep properly. The lts kernel probably ran the test a lot slower than the three other kernels, and didn't use all of the cores (for some reason, don't ask me). That's all I can think of to justify that.

EDIT: these are NOT after suspend, I never use it. My machine *just* runs on high temperatures.

What are these 'high temperatures'? Are your fan speeds set correctly?

bobzxr · 2013-08-15 20:41:22

KaiSforza wrote:

The lts kernel probably ran the test a lot slower than the three other kernels, and didn't use all of the cores (for some reason, don't ask me). That's all I can think of to justify that.

The stress test uses the number of cores I tell him to use. So the did not use all of the cores does not apply here.

From stressing only 1 cpu core I got:
75C from current kernel
53C from lts kernel

So 1 core was on 100%, the others are on ~1-3%

Lts kernel could run the test slower, but I doubt that because the core usage was 100% on the stressed cores.

What are these 'high temperatures'? Are your fan speeds set correctly?

For example I get ~85C, even 90C after like 10 minutes of playing with Hotline Miami (steam), which is not really a demanding game...
Fan spins up to 4500 RPM, which I think is the maximum level.

The regressions are not from stress tests, they are from the kernel not putting the GPU and CPU to sleep properly.

For example if I scan with clamscan, my fan starts to spin soon after start. It uses 1 core up to 100%. And the temperature starts to rise. Stressing only 1, or some cores (but not all) in my opinion is a good way to test power regression, since it is essentially the same. (because the other cores are not working)

Last edited by bobzxr (2013-08-15 20:45:01)

KaiSforza · 2013-08-15 20:52:42

bobzxr wrote:

KaiSforza wrote:
The lts kernel probably ran the test a lot slower than the three other kernels, and didn't use all of the cores (for some reason, don't ask me). That's all I can think of to justify that.
The stress test uses the number of cores I tell him to use. So the did not use all of the cores does not apply here.
From stressing only 1 cpu core I got:
75C from current kernel
53C from lts kernel
So 1 core was on 100%, the others are on ~1-3%
Lts kernel could run the test slower, but I doubt that because the core usage was 100% on the stressed cores.
What are these 'high temperatures'? Are your fan speeds set correctly?
For example I get ~85C, even 90C after like 10 minutes of playing with Hotline Miami (steam), which is not really a demanding game...
Fan spins up to 4500 RPM, which I think is the maximum level.
The regressions are not from stress tests, they are from the kernel not putting the GPU and CPU to sleep properly.
For example if I scan with clamscan, my fan starts to spin soon after start. It uses 1 core up to 100%. And the temperature starts to rise. Stressing only 1, or some cores (but not all) in my opinion is a good way to test power regression, since it is essentially the same. (because the other cores are not working)

Yes, but you are stressing cores. The main bug here is that no cores are being stressed at all, cpu usage is 0-2%, load a 0.01 or lower, yet clock speeds for the GPU and CPU are at max. Your issue does not seem like what this thread is for.

bobzxr · 2013-08-16 13:05:44

KaiSforza wrote:

Yes, but you are stressing cores. The main bug here is that no cores are being stressed at all, cpu usage is 0-2%, load a 0.01 or lower, yet clock speeds for the GPU and CPU are at max. Your issue does not seem like what this thread is for.

Sorry for the off then, I realised now I have a problem with turbo boost frequency scaling which I thought to be = power regression.

Loser777 · 2013-09-06 09:02:17

Have we abandoned all hope yet? (T420, i5 2520M)

I seem to recall some specific patterns:

pre 3.x.x: Everything was normal.
3.2-3.7.x: Irritatingly high power consumption most of the time.
3.8.x: Normal power consumption most of the time, 1/X suspend-resumes lock CPU @ max freq, power consumption becomes insane.
3.9.x-3.10.x Essentially the same behavior as 3.8.x.

It's gotten to the point where I fear putting my laptop to sleep. It's a bit of a nasty situation: I don't want to suspend my laptop because there's a chance it'll end up running unreasonably hot when I resume. On the other hand, I don't want to lose my workspace/the state of my machine when I'm done for the day, so I just leave my laptop on all night.

shadyabhi · 2013-09-06 09:09:49

Loser777 wrote:

It's gotten to the point where I fear putting my laptop to sleep. It's a bit of a nasty situation: I don't want to suspend my laptop because there's a chance it'll end up running unreasonably hot when I resume. On the other hand, I don't want to lose my workspace/the state of my machine when I'm done for the day, so I just leave my laptop on all night.

You don't need to keep your laptop on whole night. Everytime this issue happens, random 3 or 4 suspend resumes do the thing for me. Whenever I have this problem, I try to do a few more suspend/resume cycles and check with powertop if the issue is solved.

cris9288 · 2013-09-06 21:42:41

I can't speak for everyone else here, but the overheating from suspend issue appears to be solve - at least on my machine. I do notice that my laptop tends to run hotter than usual and that my fans aren't as quiet.

mar04 · 2013-09-06 22:15:34

Yup, fix landed in 3.10.6 if I remember correctly. If you still have that issue, you better let the developers know about it, because https://bugs.freedesktop.org/show_bug.cgi?id=54089 is closed as FIXED.

Chais · 2013-09-11 11:04:00

Right after boot my machine goes up to nearly 60°C, which is too much for my taste. I'd expect it to run at about 35°C when idle in Gnome.

trsk · 2013-10-10 04:31:45

I also have the impression that the power consumption and temperature of my X220 are higher than with 3.5.6-1 (which is the last version that worked I know). Can anyone confirm this?

WonderWoofy · 2013-10-10 05:02:09

trsk wrote:

I also have the impression that the power consumption and temperature of my X220 are higher than with 3.5.6-1 (which is the last version that worked I know). Can anyone confirm this?

Maybe you should run some quick tests to see if your suspicion is right? It would seem that complaints have really dies down in regard to this issue, so although I don't know that things are necessarily fully fixed, it would seem that they are likely at least better. Originally, the posters in this thread were experiencing 20-25W power consumption under normal light use. If you check powertop, I can pretty much gaurantee that you will not see this kind of insanity, unless you are running something like a desktop with some ridiculous discrete video card.

The Compiler · 2013-10-10 05:10:25

trsk wrote:

I also have the impression that the power consumption and temperature of my X220 are higher than with 3.5.6-1 (which is the last version that worked I know). Can anyone confirm this?

I'm running a vanilla 3.11.4-1 with a lightweight tiling desktop on a X220 Tablet, and I'm at 47°C / 13.2W right now, using WiFi and display on full brightness. Never has been better than this as far as I remember

Flo

Last edited by The Compiler (2013-10-10 05:10:44)

trsk · 2013-10-10 21:52:13

OK, I've now run some quantitative tests with powertop and it looks fine, thanks!

The Compiler · 2013-10-16 06:47:38

I just had the same issue (GPU does not downclock) again on 3.11.5...

trsk · 2013-10-16 07:19:03

And I have just resumed from suspend on battery power and the CPU frequency was at 100%...

3.11.5-1-ARCH

Last edited by trsk (2013-10-16 07:19:23)

eazy · 2013-11-12 17:27:39

Ah-a! Sorry for bumping this, but I just discovered something weird:

With chromium open, my i915 doesn't go into power saving, powertop indicates it running at 100% and fans start spinning.

Closing it, and using Firefox doesn't exhibit this behaviour. i915 goes into power saving mode correctly. Fans slow down to an acceptable speed.

Kernel: 3.12.0 from testing
chromium: 30.0.1599.114-1
/proc/cmdline: quiet pcie_aspm=force i915.i915_enable_rc6=1 acpi_backlight=vendor

Last edited by eazy (2013-11-12 17:28:15)

trsk · 2013-11-12 17:32:51

A follow-up from me: Since the one time I mentioned a month ago, the current stable Arch seems to be working fine on my X220 (SB). My current linux version is 3.11.6-1-ARCH.

WonderWoofy · 2013-11-12 18:04:30

eazy wrote:

/proc/cmdline: quiet pcie_aspm=force i915.i915_enable_rc6=1 acpi_backlight=vendor

I'm not sure about why your system would be suddenly using more energy, but do you really need these kernel command line arguments?

The first bold one, pcie_aspm=force, worked around an issue quite some time ago. It has the potential to cause instability, and is not needed most of the time. You might want to check to make sure that it is indeed something that you need to include.

The second bold one, i915.i915_enable_rc6=1 is actually allowing your integrated GPU to consume more power than if you didn't have this at all. Check the output of modinfo -p i915 and you'll see the various settings for that kernel module parameter. There are three levels of sleep states available to these newer integrated GPUs, rc6, rc6p, and rc6pp. You are essentially telling it that it can only go into the first sleep state and never into the deeper, more energy friendly, sleep states.

eazy · 2013-11-12 18:16:05

@WonderWoofy: that's a good question.
I'm now running without those options, and I confirm that the i915 goes to RC6 mode automatically, but doesn't go into lower power saving modes. Whatever.

But I can also confirm that using chromium the video card goes to 100% constant power usage. I'm not running any extension by the way.

Is it only me?

wlmeng11 · 2013-11-19 06:57:06

I'm using Firefox (not Chromium), and powertop reports the GPU to be powered on 100% of the time.
My temperatures are really hot (65 to 70 Celsius), causing the fan speed to be 3500 or 4000 rpm.

CPU usage is pretty low (less than 20%, load average 0.06).
My (external) display resolution is 2560x1440, but this problem has only started happening recently, so these temperatures are definitely abnormal.

3.12.0.1-ARCH
Thinkpad X220, i5

Arch Linux

#576 2013-08-07 13:35:22

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#577 2013-08-15 09:48:23

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#578 2013-08-15 10:05:58

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#579 2013-08-15 18:44:16

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#580 2013-08-15 19:42:36

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#581 2013-08-15 20:05:06

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#582 2013-08-15 20:41:22

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#583 2013-08-15 20:52:42

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#584 2013-08-16 13:05:44

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#585 2013-09-06 09:02:17

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#586 2013-09-06 09:09:49

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#587 2013-09-06 21:42:41

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#588 2013-09-06 22:15:34

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#589 2013-09-11 11:04:00

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#590 2013-10-10 04:31:45

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#591 2013-10-10 05:02:09

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#592 2013-10-10 05:10:25

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#593 2013-10-10 21:52:13

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#594 2013-10-16 06:47:38

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#595 2013-10-16 07:19:03

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#596 2013-11-12 17:27:39

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#597 2013-11-12 17:32:51

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#598 2013-11-12 18:04:30

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#599 2013-11-12 18:16:05

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#600 2013-11-19 06:57:06

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

Board footer