You are not logged in.

#1 2009-04-11 18:29:33

Denacke
Member
Registered: 2008-04-09
Posts: 106

Kernel26 2.6.29.1-3 decreases intel performance

Hey everyone,

After upgrading my testing packages I noticed my performance dropped noticably (games would not run properly anymore). After some downgrading i found the culprit smile.

The upgrade from Kernel26 2.6.29.1-2 to Kernel26 2.6.29.1-3 made my cards performance drop:

My setup
-Intel card:

[nand@nandtop ~]$ lspci |grep VGA
00:02.0 VGA compatible controller: Intel Corporation Mobile 915GM/GMS/910GML Express Graphics Controller (rev 03)

- No xorg.conf
- Related packages:

xf86-video-intel 2.6.3-3
intel-dri 7.4-2
xorg-server 1.6.0-3
xorg-server-utils 7.4-4
xorg-utils 7.4-3

Performance
Kernel26 2.6.29.1-2 performance

Kernel26 2.6.29.1-2

[nand@nandtop ~]$ glxgears 
5942 frames in 5.0 seconds = 1188.265 FPS
5969 frames in 5.0 seconds = 1193.735 FPS
5959 frames in 5.0 seconds = 1191.691 FPS

glxgears cpu-usage > 90%

Kernel26 2.6.29.1-3 performance

Kernel26 2.6.29.1-3

[nand@nandtop ~]$ glxgears 
831 frames in 5.0 seconds = 166.040 FPS
815 frames in 5.0 seconds = 162.930 FPS
811 frames in 5.0 seconds = 161.825 FPS

glxgears cpu-usage < 15%

Has anyone else been expiriencing this?

Offline

#2 2009-04-11 19:28:57

brebs
Member
Registered: 2007-04-03
Posts: 3,742

Re: Kernel26 2.6.29.1-3 decreases intel performance

That's an old intel driver - see driver versions.

Offline

#3 2009-04-11 21:47:16

Bogart
Member
From: Madrid, Spain
Registered: 2005-06-22
Posts: 272

Re: Kernel26 2.6.29.1-3 decreases intel performance

It is _very_ strange that the kernel alone is the culprit. Looking at the diff between both versions it is just an irrelevant config change about AUFS.

The problem must be elsewhere. You should probably compare /var/log/Xorg.0.log from booting the two kernels to see what changed (maybe on one boot you're using EXA and in the other UXA? Are you using KMS?)

Offline

#4 2009-04-12 07:50:08

vincnd
Member
Registered: 2008-09-14
Posts: 8

Re: Kernel26 2.6.29.1-3 decreases intel performance

Denacke wrote:

Hey everyone,

After upgrading my testing packages I noticed my performance dropped noticably (games would not run properly anymore). After some downgrading i found the culprit smile.

The upgrade from Kernel26 2.6.29.1-2 to Kernel26 2.6.29.1-3 made my cards performance drop:

I have met the same problem.After i upgrade from  Kernel26 2.6.29.1-2 to Kernel26 2.6.29.1-3 , my glxgears decreased from 1100 to 800fps, i'm using "exa" not "uxa". Anyone can help us with the problem ?

ps: how can i get the old kernel26 pakage?

Offline

#5 2009-04-12 08:13:11

Denacke
Member
Registered: 2008-04-09
Posts: 106

Re: Kernel26 2.6.29.1-3 decreases intel performance

vincnd wrote:
Denacke wrote:

Hey everyone,

After upgrading my testing packages I noticed my performance dropped noticably (games would not run properly anymore). After some downgrading i found the culprit smile.

The upgrade from Kernel26 2.6.29.1-2 to Kernel26 2.6.29.1-3 made my cards performance drop:

I have met the same problem.After i upgrade from  Kernel26 2.6.29.1-2 to Kernel26 2.6.29.1-3 , my glxgears decreased from 1100 to 800fps, i'm using "exa" not "uxa". Anyone can help us with the problem ?

ps: how can i get the old kernel26 pakage?

If you haven't cleared your pacman cache yet you should find an older version in /var/cache/pacman/pkg/. You can 'downgrade' via this command then:

pacman -U /var/cache/pacman/pkg/kernel26-2.6.29.1-2-i686.pkg.tar.gz

If you already cleared your cache, or not keeping one, you could search some mirrors for an older version or i could upload it for you smile.



I'll do some testing and report back with my xorg logs. I don't have KMS enabled (used to have it enabled, but disabled it to see if it was causing the framedrop)


Edit:

Alright, i'm back with the logs! Seems to be some weird things happening, but i don't have enough knowledge to pinpoint where the problem lies.
Here are the logs:
Kernel26 2.6.29.1-2 : http://denacke.net/archlinux/kernel3/kernel2.Xorg.0.log
Kernel26 2.6.29.1-3 : http://denacke.net/archlinux/kernel3/kernel3.Xorg.0.log

Some of the interesting things happening:
in Kernel26 2.6.29.1-2 it seems to load the i915 module, it doesn't have that line with Kernel26 2.6.29.1-3
Kernel26 2.6.29.1-2

II) [drm] loaded kernel module for "i915" driver.

There also seems to be some problem when setting tiling with the latest kernel version.
Kernel26 2.6.29.1-3

> (EE) intel(0): Failed to set tiling on front buffer: rejected by kernel
> (EE) intel(0): Failed to set tiling on back buffer: rejected by kernel
> (EE) intel(0): Failed to set tiling on depth buffer: rejected by kernel

Edit edit:

brebs wrote:

That's an old intel driver - see driver versions.

I'm using the latest package available in archlinux repos. I'll see if it make any difference when using the latest drivers when I have some time for compiling them smile

Last edited by Denacke (2009-04-12 08:38:20)

Offline

#6 2009-04-12 09:20:48

Aktau
Member
From: Antwerp, Belgium / BCN, Spain
Registered: 2009-04-04
Posts: 19
Website

Re: Kernel26 2.6.29.1-3 decreases intel performance

Hi! I'm very interested in this thread because I have the same card and have recently noticed a very similar performance regression.

For example, until recently my glxgears output was more like ~220, and now it's ~22 (a tenfold decrease). So your piece immediately reminded me of my own problem. I didn't however
notice immediately when the performance dropped so quite a lot of package upgrades have happened in the meanwhile, so I'll be virtually helpless in trying to determine a cause (I also cleared my cache for the very first time ever yestarday).

I'm also continuing my search into why the performance was so goddamn bad to begin with, I used to have ~1100 on ubuntu, ~300 on arch with EXA, ~220 on arch with KMS/DRI2/UXA and now  ~22 on arch with KMS/DRI2/UXA. Oh, and I'm using the testing repository and no external repositories, if it interests anyone.

Offline

#7 2009-04-12 09:27:08

Allan
Pacman
From: Brisbane, AU
Registered: 2007-06-09
Posts: 11,357
Website

Re: Kernel26 2.6.29.1-3 decreases intel performance

This is really weird....   the diff between -2 and -3 is only to do with unionfs and aufs2: http://repos.archlinux.org/viewvc.cgi/k … 8&r2=34241

Online

#8 2009-04-12 09:28:02

Denacke
Member
Registered: 2008-04-09
Posts: 106

Re: Kernel26 2.6.29.1-3 decreases intel performance

I'm uploading the Kernel26 2.6.29.1-2 package to my webspace, will edit post when it's done (ETA 15 mins). Will be interesting to see if it changes your performance!

Edit:

Here it is:  http://denacke.net/archlinux/kernel3/ke … pkg.tar.gz


Aktau wrote:

Hi! I'm very interested in this thread because I have the same card and have recently noticed a very similar performance regression.

It's weird how your performance seems to have dropped so badly. I used to have bad performance too a few months ago. But xorg 1.6 made the difference. I also stopped using a xorg.conf, not sure if it could help you.


Allan wrote:

This is really weird....   the diff between -2 and -3 is only to do with unionfs and aufs2: http://repos.archlinux.org/viewvc.cgi/k … 8&r2=34241

I also have no idea why these changes in the package make the difference for me, but performance really is completely different when using package -2 instead of -3 (and that's the only thing i'm downgrading). Perhaps it does something more which could explain this?

Last edited by Denacke (2009-04-12 09:44:44)

Offline

#9 2009-04-12 10:26:47

moljac024
Member
From: Serbia
Registered: 2008-01-29
Posts: 2,676

Re: Kernel26 2.6.29.1-3 decreases intel performance

Allan wrote:

This is really weird....   the diff between -2 and -3 is only to do with unionfs and aufs2: http://repos.archlinux.org/viewvc.cgi/k … 8&r2=34241

The OP didn't say he was running x86_64...perhaps there are more changes in the i686 package ?


The day Microsoft makes a product that doesn't suck, is the day they make a vacuum cleaner.
--------------------------------------------------------------------------------------------------------------
But if they tell you that I've lost my mind, maybe it's not gone just a little hard to find...

Offline

#10 2009-04-12 10:36:31

moljac024
Member
From: Serbia
Registered: 2008-01-29
Posts: 2,676

Re: Kernel26 2.6.29.1-3 decreases intel performance

Aktau wrote:

...I'm also continuing my search into why the performance was so goddamn bad to begin with, I used to have ~1100 on ubuntu, ~300 on arch with EXA, ~220 on arch with KMS/DRI2/UXA and now  ~22 on arch with KMS/DRI2/UXA. Oh, and I'm using the testing repository and no external repositories, if it interests anyone.

That's one thing I can't seem to figure out either, I thought I was alone in that one, but it seems that performace is just much worse here on Arch... perhaps an examination of ours and some of the mainstream distros packages for xorg should be in order ?

Granted, there probably aren't a lot of distros which have the latest xorg but it was just as bad with 1.5, right ?
Let's wait and see how performance will be like in the upcoming versions of Ubuntu, Fedora and the likes, which should include the newer Xorg. There's definitely something fishy here...


The day Microsoft makes a product that doesn't suck, is the day they make a vacuum cleaner.
--------------------------------------------------------------------------------------------------------------
But if they tell you that I've lost my mind, maybe it's not gone just a little hard to find...

Offline

#11 2009-04-12 11:34:01

Bogart
Member
From: Madrid, Spain
Registered: 2005-06-22
Posts: 272

Re: Kernel26 2.6.29.1-3 decreases intel performance

Well, I still have no idea why the kernel upgrade caused that change. But it seems to be related with tiling not working correctly.

Does everyone experiencing problems have those error lines in /var/log/Xorg.0.log about tiling?

Does something like adding these lines to /etc/X11/xorg.conf workaround the problem?

Section "Device"
        Identifier  "card0"
        Driver      "intel"
        Option     "Tiling" "False"
EndSection

If so, you might have to file a bug report so that those changes related to AUFS and UNIONFS are reverted until it's known why they might cause this strange interaction.

(By the way, I'm using 2.6.29.1-3 in two computers with Intel graphics and I'm not seeing the problem. Strange)

Offline

#12 2009-04-12 16:36:31

Aktau
Member
From: Antwerp, Belgium / BCN, Spain
Registered: 2009-04-04
Posts: 19
Website

Re: Kernel26 2.6.29.1-3 decreases intel performance

moljac024 wrote:

Granted, there probably aren't a lot of distros which have the latest xorg but it was just as bad with 1.5, right ?
Let's wait and see how performance will be like in the upcoming versions of Ubuntu, Fedora and the likes, which should include the newer Xorg. There's definitely something fishy here...

(emphasis mine)

Exactly, it was equally bad with Xorg 1.5 / intel 2.4.x as with Xorg 1.6 / intel 2.6.x

I actually switched to testing and eventually even a fullblown KMS/DRI2/UXA config to get away from the performance problems, but it just keeps getting worse. I'm afraid I'm still learning when it comes to linux, especially when it comes to all kinds of package management, but I sure would like to lend a hand.

Denacke wrote:

It's weird how your performance seems to have dropped so badly. I used to have bad performance too a few months ago. But xorg 1.6 made the difference. I also stopped using a xorg.conf, not sure if it could help you.

With or without xorg.conf makes little difference with regards to performance, because with my current KMS setup it defaults to DRI2/UXA rendering anyway. I have an xorg.conf just to define my fonts folder in home and to turn on the dontzap property so I can use ctrl-alt-bksp.

Last edited by Aktau (2009-04-12 16:39:54)

Offline

#13 2009-04-25 00:09:34

karabaja4
Member
From: Croatia
Registered: 2008-09-14
Posts: 997
Website

Re: Kernel26 2.6.29.1-3 decreases intel performance

I confirm this problem. The 2.6.29.1-2 kernel works much better in 3D than the 2.6.29.1-3 or 2.6.29.1-4 (Intel i915, EXA, xorg 1.6.1, xf86-video-intel 2.7.0).

It indeed seems to be a tiling issue.

Here's the Xorg.0.log from 2.6.29.1-4 kernel:

...
(EE) intel(0): Failed to set tiling on front buffer: rejected by kernel
(EE) intel(0): Failed to set tiling on back buffer: rejected by kernel
(EE) intel(0): Failed to set tiling on depth buffer: rejected by kernel
...
(II) intel(0): 0x00800000-0x00bfffff: depth buffer (4096 kB)
(II) intel(0): 0x00c00000-0x00ffffff: back buffer (4096 kB)
(II) intel(0): 0x01000000-0x013fffff: front buffer (4096 kB)
...

and here's from 2.6.29.1-2 kernel:

...
(II) intel(0): 0x00800000-0x00bfffff: depth buffer (4096 kB) X tiled
(II) intel(0): 0x00c00000-0x00ffffff: back buffer (4096 kB) X tiled
(II) intel(0): 0x01000000-0x013fffff: front buffer (4096 kB) X tiled
...

in glxgears,

- 2.6.29.1-2 gives 1200fps and etracer runs smoothly (if I disable tiling with Option "Tiling" "false" I get very low performance in etracer and 170fps in glxgears).
- 2.6.29.1-3 or 2.6.29.1-4 gives 170fps and etracer runs at 2-3 fps.

Its the same situation with UXA, and with or without KMS.

Last edited by karabaja4 (2009-04-25 13:40:37)

Offline

#14 2009-04-26 17:36:37

Denacke
Member
Registered: 2008-04-09
Posts: 106

Re: Kernel26 2.6.29.1-3 decreases intel performance

Good to see i'm not alone with this problem! So it seems this is only happening on i915 cards.

Should we be filing a bug report or wait for a next kernel update?

Offline

#15 2009-04-26 18:19:45

karabaja4
Member
From: Croatia
Registered: 2008-09-14
Posts: 997
Website

Re: Kernel26 2.6.29.1-3 decreases intel performance

I already posted here since it seems like a similar issue: http://bugs.archlinux.org/task/14409?pr … &pagenum=2 - Maybe we should open a new bug report since this is related to 915 and not 945.

Last edited by karabaja4 (2009-04-29 23:25:58)

Offline

#16 2009-04-29 17:13:05

karabaja4
Member
From: Croatia
Registered: 2008-09-14
Posts: 997
Website

Re: Kernel26 2.6.29.1-3 decreases intel performance

...

Last edited by karabaja4 (2009-05-01 16:13:17)

Offline

#17 2009-04-29 23:25:36

karabaja4
Member
From: Croatia
Registered: 2008-09-14
Posts: 997
Website

Re: Kernel26 2.6.29.1-3 decreases intel performance

I found the cause of this. It's got nothing to do with aufs or unionfs - the real cause are the kernel patches, or to be more specific, this one: patch-2.6.29.1-3-ARCH (link to kernel patches)

There is massive diff in 2.6.29.1-3 patch compared to 2.6.29.1-2 (note that these are only small parts of the code from a huge patch file):

patch-2.6.29.1-2-ARCH
patch-2.6.29.1-3-ARCH

code that causes the slowdown on newest 2.6.29.2-1 kernel - <file deleted sorry> (this code is little different than 2.6.29.1-3, but causes the same slowdown).

Why is this huge code here, I don't know. Perhaps to make GEM work for some cards other than i915. I really hope this mumbo-jumbo will be fixed in 2.6.30 (the 2.6.29.2 has the same problem, 2.6.30-rc3 works fine - but then again, 2.6.30 is purely from git) sad

EDIT: I compiled a patched 2.6.29.2-1 kernel (newest at the moment) with good i915 code -> download (works for me, but use at your own risk).

P.S. Please use firefox to download the kernel package. Also, hosting isn't very stable so if you experience download troubles try again in a few minutes.
P.P.S Looks like this kernel I compiled fixed every single problem I had with i915. EXA and XAA work great (XAA is better in 2D like scrolling, they both work good in 3D), UXA works fine but not as good as XAA (this is expected since UXA is still in development). I use xf86-video-intel 2.7.0-2.

EDIT 1.6.2009:
I'll update the kernels as they come along and upload them here: http://archlinux.pondi.hr/  Latest version is 2.6.29.4-1.

Last edited by karabaja4 (2009-06-02 17:28:26)

Offline

#18 2009-05-01 19:54:11

Denacke
Member
Registered: 2008-04-09
Posts: 106

Re: Kernel26 2.6.29.1-3 decreases intel performance

Hey, nice find!

Upgraded my kernel to your package and glxgears does go at full speed now. smile (I have to admit, i used opera to download it though :$)

Offline

#19 2009-05-01 20:21:17

Bogart
Member
From: Madrid, Spain
Registered: 2005-06-22
Posts: 272

Re: Kernel26 2.6.29.1-3 decreases intel performance

I wonder where those patches come from and why there is no changelog about them.

I think there was a bug report for this issue, so maybe you should post there (if you haven't done already) and maybe the kernel maintainer can explain why that patch was used and if it can be reverted once it's known to cause this issue.

Offline

#20 2009-05-02 06:35:17

jacob_
Member
Registered: 2009-04-15
Posts: 12

Re: Kernel26 2.6.29.1-3 decreases intel performance

karabaja4 wrote:

EDIT: I compiled a patched 2.6.29.2-1 kernel (newest at the moment) with good i915 code -> download (works for me, but use at your own risk).

Good work!

What do you mean by good dri code? Do you mean you remove the arch patches and did you apply the tiling fix from Jesse Barnes?

I'm considering 2.26.30-rc4 now that the Jesse's dri fix is in.

The latest intel 2.8 release seems to have fixed some other recent issue that was causing my windows to redraw slowly.

Offline

#21 2009-05-02 12:44:44

karabaja4
Member
From: Croatia
Registered: 2008-09-14
Posts: 997
Website

Re: Kernel26 2.6.29.1-3 decreases intel performance

Denacke wrote:

Hey, nice find!

Upgraded my kernel to your package and glxgears does go at full speed now. smile (I have to admit, i used opera to download it though :$)

My host had some troubles accepting requests from other browsers, glad to see it works.

BTW, I have a question for you (since you are the one of the few people with i915). Does your glxgears show more FPS when you kill xorg and start it again (more than a first time)? For example, I start my computer, my xorg and glxgears shows 1000fps. I kill my xorg with ctrl-alt-backspace, and start it again, now glxgears shows 1200fps.

jacob_ wrote:

Good work!

What do you mean by good dri code? Do you mean you remove the arch patches and did you apply the tiling fix from Jesse Barnes?

I'm considering 2.26.30-rc4 now that the Jesse's dri fix is in.

The latest intel 2.8 release seems to have fixed some other recent issue that was causing my windows to redraw slowly.

I removed the whole diff related to the intel files from the patch, I only left the intel_bios.h diff (check the links in my previous post). I did that because 2.6.29.1-2 had the same configuration and was showing much better performance.

I didn't apply any other fixes, sorry. I tried 2.6.30-rc4 too, it works good, same as this patched kernel smile

Offline

#22 2009-05-03 15:32:08

afiestas
Member
Registered: 2008-12-28
Posts: 20

Re: Kernel26 2.6.29.1-3 decreases intel performance

From 150fps to 1100fps, thanks for the package!

Offline

#23 2009-05-03 15:47:36

Denacke
Member
Registered: 2008-04-09
Posts: 106

Re: Kernel26 2.6.29.1-3 decreases intel performance

karabaja4 wrote:

BTW, I have a question for you (since you are the one of the few people with i915). Does your glxgears show more FPS when you kill xorg and start it again (more than a first time)? For example, I start my computer, my xorg and glxgears shows 1000fps. I kill my xorg with ctrl-alt-backspace, and start it again, now glxgears shows 1200fps.

The difference here is barely noticable.

When running glxgears for the first time I got around 1150-1160, after restarting Xorg it showed around 1190 fps

Offline

#24 2009-05-03 16:40:30

Ashren
Member
From: Denmark
Registered: 2007-06-13
Posts: 1,229
Website

Re: Kernel26 2.6.29.1-3 decreases intel performance

Looking at the FPS's you guys get I guess 375 FPS in the default glxgears window for a "00:02.0 VGA compatible controller: Intel Corporation Mobile GM965/GL960 Integrated Graphics Controller (rev 0c)" is really bad.

I'm using the kernel supplied in this thread, do I need to build mesa-git, intel-dri-git etc. from AUR to get better performance?

Actually etracer seems to be running alright.

I know glxgears is no benchmark, but the difference between 375 FPS and above 1000 FPS is quite big.

Could it be because I'm using "AutoAddDevices" "False"?

Offline

#25 2009-05-03 16:53:54

karabaja4
Member
From: Croatia
Registered: 2008-09-14
Posts: 997
Website

Re: Kernel26 2.6.29.1-3 decreases intel performance

Ashren wrote:

Looking at the FPS's you guys get I guess 375 FPS in the default glxgears window for a "00:02.0 VGA compatible controller: Intel Corporation Mobile GM965/GL960 Integrated Graphics Controller (rev 0c)" is really bad.

I think this is i915 cards issue, sorry. You might get better kernel performance with kernel 2.6.30-rc4 which solves many other bugs: http://aur.archlinux.org/packages.php?ID=25304

Offline

Board footer

Powered by FluxBB