Kernel 2.6.38 really impresses!

Message

Anon-E-moose · Post by **Anon-E-moose** » Tue Apr 19, 2011 5:35 pm

I was getting a few random seg-faults with 2.6.38.2 one in less (man) one in tcp (during system boot)
I haven't noticed any since going to 2.6.38.3.

pilla · Post by **pilla** » Tue Apr 19, 2011 6:50 pm

VinzC wrote:
pilla wrote:2.6.38-gentoo-r2 is still segfaulting Xorg for me.
Does it crash right away or after a certain, random period of time?

After some random time, usually not so long.

VinzC · Post by **VinzC** » Fri Apr 22, 2011 2:09 pm

pilla wrote:2.6.38-gentoo-r2 is still segfaulting Xorg for me.

VinzC wrote:Does it crash right away or after a certain, random period of time?

pilla wrote:After some random time, usually not so long.

I think we might experience the same bug.

EDIT: Apparently there's a pending fix for radeon video cards. I'll try it and report if the bug is fixed. I think if there's no issue in a week or so then the patch may appear to actually fix the issue.

robinmarlow · Post by **robinmarlow** » Tue Apr 26, 2011 9:34 pm

I too am having problems with 2.6.38. I compile it using my working config from 2.6.37 & answering the questions posed by make oldconfig.
The resulting kernel gives a blank screen with no hd activity or response to magic keys. There are no logs & no clues as to what's gone wrong!

I have tried compiling vanilla as well as gentoo sources with the same result. I'm now trying excluding sections of the config e.g. audio individually to see if they are causing the problem - but it's rather time consuming!
How should I troubleshoot it & does anyone know what the problem is?

Thanks,

Robin

jbouzan · Post by **jbouzan** » Wed Apr 27, 2011 3:42 am

Interesting. I'd say use make oldconfig and then manually check with make nconfig or make xconfig for your video, hard drive, and filesystem drivers to be compiled in.

wswartzendruber · Post by **wswartzendruber** » Wed Apr 27, 2011 3:47 am

Phoronix is reporting massive power consumption issues with this kernel. Anyone else notice this?

VinzC · Post by **VinzC** » Wed Apr 27, 2011 7:29 am

robinmarlow wrote:I too am having problems with 2.6.38. I compile it using my working config from 2.6.37 & answering the questions posed by make oldconfig.
The resulting kernel gives a blank screen with no hd activity or response to magic keys. There are no logs & no clues as to what's gone wrong!

I have tried compiling vanilla as well as gentoo sources with the same result. I'm now trying excluding sections of the config e.g. audio individually to see if they are causing the problem - but it's rather time consuming!
How should I troubleshoot it & does anyone know what the problem is?

You might want to try Pappy's Kernel Seeds or [topic=829476]this thread[/topic]. You «just» need to add your hardware-specific options like drivers, firmwares...

wswartzendruber wrote:Phoronix is reporting massive power consumption issues with this kernel. Anyone else notice this?

I'm watching Phoronix too. It even looks like there is a «growing increase» in consumption since 2.6.25.

pilla · Post by **pilla** » Mon May 02, 2011 4:12 pm

2.6.39-rc5 in my Thinkpad T60 consumes 26W instead of 20W of 2.6.35.

Jaglover · Post by **Jaglover** » Wed May 04, 2011 12:02 pm

This automatic group scheduling is not for me. Turned it off with -r4 and got back my responsive desktop. I've Core2 Duo, dual monitors, most of time MythTV is running on the second one. With group scheduling on every time I started something on first monitor MythTV picture stopped for a fraction of a second, very annoying.

red-wolf76 · Post by **red-wolf76** » Mon May 09, 2011 9:07 pm

I don't get it. Whatever happened to the choice factor? Even for mainline kernel development, we get to choose between CFQ and no-op, even if the latter is really not a viable alternative. There's ton's of stuff in the kernels marked "Experimental" with rolls of red tape around it.

What's wrong with including the new scheduler and then slapping a big warning on top of it. "This scheduler is reported to have better performance on systems with fewer than 16 CPUs. If you know what you're doing and have such a system, you may want to select this. If you have no idea or don't care, choose CFQ instead."

End of story. Make it available to people that are sensible enough to actually read the help for kernel settings - provided there is any...

Honestly... This politics crap sounds like even programmers aren't immune to start masturbating on kindergarten power games once they get bored... Why not let Joe User decide?

Joseph K. · Post by **Joseph K.** » Tue May 17, 2011 7:21 am

VinzC, did you already post a link to this kernel bug in regard to the screen saver oops that you mentioned? (I had a feeling that you did, but then I couldn't find it.) I hit this screen saver crash only every couple of days, but it's enough to motivate me to want to debug it. Do you want to open a Gentoo bug or a new forum thread for it if there isn't one already?

VinzC · Post by **VinzC** » Tue May 17, 2011 8:54 am

Joseph K. wrote:VinzC, did you already post a link to this kernel bug in regard to the screen saver oops that you mentioned? (I had a feeling that you did, but then I couldn't find it.)

Yes, I did in [post=6661269]this post[/post] (a little above).

Joseph K. wrote:Do you want to open a Gentoo bug or a new forum thread for it if there isn't one already?

Well, as I saw there is already a bug opened in kernel.org, I didn't feel ike doing it on Gentoo's side. Fact is there doesn't seem to be much progress on that lately.

Bircoph · Post by **Bircoph** » Thu May 19, 2011 9:10 pm

No problems with 2.6.38 on both laptop and server so far (I use vanilla-sources).

Transparent hugepages is a really nice feature in this kernel, it measurably speeds up world upgrade.

Also I can't understand why many people are so interested in cgroups scheduler. I tried this feature and it really sucks, because it hurt both throughoutput and peak multimedia performance (think of 1080p h264 on Atom N270 without frame drops). All this tests with -jMANY and "my desktop is still responsible" looks like completely synthetic and unpractical for me: just learn how to use nice and ionice, this will save both your system performance and user experience. Of course, anyone is free to configure their kernels as they like — that is Gentoo way after all —, but I personally see absolutely no gain from cgroups and refuse to use them by now.

Gusar · Post by **Gusar** » Thu May 19, 2011 10:22 pm

Bircoph wrote:think of 1080p h264 on Atom N270 without frame drops

Err, the thing can barely decode 720p, are you really playing 1080p on it? Without hardware help (Nvidia ION or Broadcom CrystalHD) it's not possible.

Bircoph · Post by **Bircoph** » Thu May 19, 2011 10:37 pm

Gusar wrote: Err, the thing can barely decode 720p, are you really playing 1080p on it? Without hardware help (Nvidia ION or Broadcom CrystalHD) it's not possible.

You forgot to mention on unoptimized software, perhaps.

I use SHE + ffmpeg-mt + hardly optimized general system + skiploopfilter for lavdopts (this doesn't hurt quality in general).
Of course, video is downscaled during playback to fit 1024x600 resolution.

Gusar · Post by **Gusar** » Fri May 20, 2011 9:31 am

What is SHE? And if you're skipping loop filtering, that's something else, I don't know what happens then. I do know that it actually does degrade video quality, sometimes very visibly so. And sometimes it causes artifacts, because the video does motion compensation on filtered frames. Also, what's the bitrate of those 1080p videos?

Bircoph · Post by **Bircoph** » Fri May 20, 2011 6:21 pm

Gusar wrote:What is SHE?

Super Hybrid Engine is a hardware clock frequency control for eeepc, including light overclocking and moderate underclocking to save battery power: http://event.asus.com/notebook/bamboo/external5.htm

I do know that it actually does degrade video quality, sometimes very visibly so. And sometimes it causes artifacts, because the video does motion compensation on filtered frames.

Read mplayer's manual then:

Code: Select all

                 skiploopfilter=<skipvalue> (H.264 only)
                      Skips  the  loop filter (AKA deblocking) during H.264 decoding.  Since the fil‐
                      tered frame is supposed to be used as reference for decoding  dependent  frames
                      this  has  a  worse  effect on quality than not doing deblocking on e.g. MPEG-2
                      video.  But at least for high bitrate HDTV this provides a big speedup with  no
                      visible quality loss.

                      <skipvalue> can be either one of the following:
                         none: Never skip.
                         default: Skip useless processing steps (e.g. 0 size packets in AVI).
                         nonref: Skip frames that are not referenced (i.e. not used for decoding oth‐
                         er frames, the error cannot "build up").
                         bidir: Skip B-Frames.
                         nonkey: Skip all frames except keyframes.
                         all: Skip all frames.

This is just deblocking skip. You wan't need this for 1080p in most of the cases anyway. And you can always safely skip deblocking for at least B-frames.

Also, what's the bitrate of those 1080p videos?

I tested on different samples, most of them are 25 fps digital video, some are 24 fps.

Joseph K. · Post by **Joseph K.** » Sat May 21, 2011 6:15 am

VinzC wrote:Well, as I saw there is already a bug opened in kernel.org, I didn't feel ike doing it on Gentoo's side. Fact is there doesn't seem to be much progress on that lately.

Yeah, but I think it's a good way to draw the Gentoo kernel maintainers' attention to the bugs that particularly affect us, which can draw interest and spur progress on them. Now that 2.6.39 is out, though, I may be inclined to try using that in the hope that this bug was fixed. If I can get a screen shot of the oops, I'll open a Gentoo bug for it.

Gusar · Post by **Gusar** » Sat May 21, 2011 12:22 pm

Bircoph wrote:Super Hybrid Engine

Ah, marketing buzzword for on-the-fly CPU governor switching.

Bircoph wrote:This is just deblocking skip. You wan't need this for 1080p in most of the cases anyway.

The in-loop deblocker is an integral part of h264 encoding/decoding. If it was done during encoding, then turning it off for decoding degrades the picture and can lead to artifacts. That you aren't bothered by the difference is something else. I can imagine it's probably not noticeable in the netbook display, but on anything bigger than 15'' ...

Bircoph wrote:I tested on different samples, most of them are 25 fps digital video, some are 24 fps.

That's fps, I asked for bitrate.

Bircoph · Post by **Bircoph** » Sun May 22, 2011 6:53 pm

Gusar wrote:
Bircoph wrote:Super Hybrid Engine
Ah, marketing buzzword for on-the-fly CPU governor switching.

You failed to read documentation again. SHE is completely independent from CPU governors and affects not only CPU.

The in-loop deblocker is an integral part of h264 encoding/decoding. If it was done during encoding, then turning it off for decoding degrades the picture and can lead to artifacts. That you aren't bothered by the difference is something else. I can imagine it's probably not noticeable in the netbook display, but on anything bigger than 15''

I often use it on my 19" desktop, artifacts are rare and I have never seen them on high-res video. And deblocking is not always done during encoding. Anyway, the goal is to be able to see a 1080p movie on netbook with an acceptable quality. This is not my goal to obtain mathematically equivalent image after decoding.

That's fps, I asked for bitrate.

Bitrate is variable, obviously, and may vary greatly during playback. An average value for the samples, estimated by duration and file size, and by taking audio tracks into account, though variable as well, is 8 kbps.

Cyker · Post by **Cyker** » Sun May 22, 2011 8:48 pm

Has anyone had a problem with high iowait in 2.6.38 after several hours/days use?

I ran into this with -r1 - It seems that a kernel task called khugepaged goes into the D state for some reason, and then one of my cores gets hit with a 100% IOWAIT.

It seems to be related to Transparent hugepages; Have disabled it completely and while there is a noticeable performance hit (Not big, just noticeable) I so far haven't had the problem reoccur... but might be too early to judge...!

Gusar · Post by **Gusar** » Mon May 23, 2011 8:42 am

Bircoph wrote:You failed to read documentation again. SHE is completely independent from CPU governors and affects not only CPU.

Ok, I've read a bit about it. It scales the fsb freq instead of the cpu freq. The highest modes also do a bit of overclocking, and then there are modes that do dynamic scaling. So while I oversimplified, it's still just a buzzword frequency scaling and governor switching.

Bircoph wrote:And deblocking is not always done during encoding.

Deblocking is very, very rarely not done. I'll bet you you won't find videos out there that don't have it, unless it's conformance tests suites or the person encoding is doing some sort of test.

soya · Post by **soya** » Tue Jun 21, 2011 4:35 pm

Hi, to make use of the feature i just need to enable SCHED_AUTOGROUP or there is something more i should do?

VinzC · Post by **VinzC** » Tue Jun 21, 2011 10:14 pm

soya wrote:Hi, to make use of the feature i just need to enable SCHED_AUTOGROUP or there is something more i should do?

In his first post, VinzC wrote:Be sure to check CONFIG_CGROUPS and CONFIG_CGROUP_SCHED in your kernel config for it's not enabled by default.

depontius · Post by **depontius** » Tue Jun 21, 2011 10:16 pm

I believe SCHED_AUTOGROUP either turns on the necessary features, or isn't available until they are on, or some mix of the two. I didn't do a heck of a lot - I may have enabled CGROUPS, but I just turned on SCHED_AUTOGROUP and then several other options popped in, too.