Bug 23083

Summary: [G43] GtkPerf performance measurements 15% slower with 2.8.0 than 2.6.x (regression)
Product: xorg Reporter: Bryce Harrington <bryce>
Component: Driver/intelAssignee: Carl Worth <cworth>
Status: RESOLVED WORKSFORME QA Contact: Xorg Project Team <xorg-team>
Severity: major    
Priority: high CC: Michael, xunx.fang
Version: 7.4 (2008.09)Keywords: NEEDINFO, regression
Hardware: All   
OS: Linux (All)   
i915 platform: i915 features:

Description Bryce Harrington 2009-08-01 15:01:07 UTC
Forwarding this performance testing issue as reported in Phoronix by Michael Larabel:

On the tester's G43 hardware running Ubuntu Karmic Alpha-3, he got GtkPerf results which were about 15% slower than against a stock Jaunty (with 2.6.x).

Hardware tested:
ASRock G43Twins-FullHD

Note:  I have not been hearing complaints about karmic performance in general from users (who are not shy about flagging performance problems), however Michael is the only one I know of who runs quantitative tests regularly against Ubuntu's intel stack.  I would guess it to be something peculiar to the hardware he tested, but it would be worth investigation.
Comment 1 Bryce Harrington 2009-08-01 15:12:48 UTC
Michael, when you subscribe here could you attach your Xorg.0.log with ModeDebug turned on, and the output of 'dmesg'?
Comment 2 Michael Larabel 2009-08-01 18:39:25 UTC
Yep, will do, Bryce. See my other bug comment.
Comment 3 Gordon Jin 2009-08-01 19:32:05 UTC
Let me share some info from my side:
I used to run Phoronix Test Suite (v1.8.1) on G35 (with 32-bit OS) to compare Intel Q2-rc2 release (xf86-video-intel master and mesa_7_5_branch git tip around July 6th) and Intel Q1 release (2.7.0). The result shows Q2 wins Q1 for most cases. As for the 3 cases of GtkPerf used by Michael, my result shows Q2 wins about 20% for all of them.
Padman, OpenArena and Tremulous all runs successfully on my side. (I didn't try Urban Terror)

That means the result differs much between Michael and my side, and we'd better investigate which element causes the big difference:
1) PTS version (Michael, is it fine for me to run 1.8.1? It might take me more time to upgrade to the latest version) 
2) the difference between Intel Q2-rc2 (I suppose it's close to Q2 release, i.e. 2.8.0) and Ubuntu 9.10 Alpha3 (Bryce, can you elaborate the kernel in Alpha3?)
3) 32-bit v.s. 64-bit (I didn't expect there's huge difference. But I'll try to confirm)
4) G35 v.s. G43 (I didn't expect difference)
5) 2.7.0 regressed much against 2.6.3 (for EXA)? But this can't explain the 3d broken observed by Michael.

Once 1) get confirmed, I'd be able to replicate the rest.
Comment 4 Michael Larabel 2009-08-01 20:02:49 UTC
Hi Gordon,

Using version 1.8 of the Phoronix Test Suite should be fine for the 2D tests. The results between the versions of PTS should not matter and should produce nearly identical results. Of course, the exception is where a newer version of PTS is using a newer upstream software package, but there have been no GtkPerf or QGears2 test profiles in recent times, so their test profiles still should be identical, but in the event they were not, PTS will automatically inform you of that when trying to view the results.

When I am unpacked I will see if I still have any G35 hardware around, if so I can then see how that compares. I know last year immediately after the G43 launched I ran into a few G4x-specific issues, but if my memory serves me they weren't performance related, Keith might recall.

Comment 5 Ivan Bulatovic 2009-08-02 00:52:01 UTC
Hi, my results with pts benchmarking differs also, I've tested it though with kms enabled on kernel and 2.7.1 xf86-video-intel against 2.6.31-rc4 and 2.8.0 intel video drivers. So my testing only regards to performance under UXA.

In every single test that I run, 2.6.31-rc4 and 2.8.0 wins, and somehow I don't think that it is got much to do with the xf86 driver as much as with intel drivers in the kernel tree. I've tested the new kernel with the old driver and that is not making as much difference in performance.

Let aside PTS I can actually feel that with 2.6.31 intel is much more responsive, I know it is subjective but urt gives more fps and is more responsive and compiz and everything else is nippy as ever.

As for stabillity issues, with 2.6.31 I haven't had a single freeze, prolems with native resolution while booting are fixed, fullscreen apps with resolution lower then a native one run maximized now, compiz doesn't leak anymore (related to mesa upgrade), now firefox and other apps dont freeze the entire system, only bug left is with brightness hotkeys but that is in my opinion not your fault, ati users have this issue also, must be ACPI related. 

UXA was terrible until 2.6.31 was released, now I'm pretty happy how things run. I'm running Arch linux.

I belive Michael when he says that there is performace regression between Jaunty and Karmic, but for me on Arch linux things are only geting better - both PTS and my user experience can vouch for that... So somehow I think that this is a problem that should be resolved by guys @Canonical...     
Comment 6 Gordon Jin 2009-08-05 02:13:42 UTC
I tried replicating Michael's test configurations (G43, with x86-64 build), but get the opposite result -- my result shows 9.10 Alpha3 is better than 9.04:
                     GtkComboBox    GtkCheckButton  GtkDrawingArea Pixbufs
Ubuntu 9.04:          44.87           9.98               14.83
Ubuntu 9.10 Alpha3:   34.63           5.10               13.87

(Note: when I install PTS by apt on Ubuntu, PTS 1.6.0 is installed on Ubuntu 9.04 and PTS 2.0 is installed on Ubuntu 9.10 Alpha3, and GtkPerf version shows v0.40 in both cases. I guess this is expected?) 
Comment 7 Carl Worth 2009-08-06 10:41:21 UTC
Gordon and Ivan,

Thanks for the careful testing.

Since we've been totally unable to replicate the bug, I'll go ahead and close it now.

If someone can point out something we're missing or can replicate it, please feel free to reopen it.


Use of freedesktop.org services, including Bugzilla, is subject to our Code of Conduct. How we collect and use information is described in our Privacy Policy.