Comment 46 for bug 339555

One thing to keep in mind here (at least in my case) is that on the live CD I put a different kind of load on the graphics card. On my installed version I do certain things differently, like for example the first thing I do at boot is to restore my never ending Firefox session which has got like 20 windows with around 10 TABs in each window (I'm speculating here but maybe that will fill up the graphics card memory or GART aperture or whatever these things are called). There is also other things that are different for example, I use a special theme (Dust) on the installed version. Again, this is complete speculation but what if one theme used a special kind of gradient or other graphics primitive that taxed the graphics card heavily (of course the default human theme also has gradients and glossyness etc but you get the idea, the scenarios might be hitting different executation paths in the driver). As far as I understand it's very possible that certain operations on the card are dead slow while others are fast.

Also about the kernel fix, I was under the impression that we were dealing with three separate intel perf regressions at once. These are A) tiling being disable due to some MCHBAR thing, B) tiling being disabled due to some A17 sizzling thing and then C) the thing I was having and possibly also the original bug reporter for this bug. Note that I never once so any error about tiling not working correctly, not in xorg.log and not in dmesg so basically I think the tiling issues is limited to older intel chipsets like 945 and whatnot (I have G45). Even just given the possilibity that we're dealing with different vaguely formulated perf bugs, probably means we should do some more measurements on this. Unfortunately, I'll be traveling for the next week so I can't help with G45 testing.

I think the best way to proceed is to try to isolate some kind of repeatable benchmark so that we can then run this scenario on both intrepid EXA and jaunty EXA and then say it's this and that much slower or faster. It might also be necessary to create/find a test workflow that uses lots of big bitmaps / textures (as in multiple fullscreen apps in compiz). I was thinking about Phoronix Test Suite as one alternative but while they have a lot of gaming benchmark I don't think they have anything that is similar to fullscreen operations in compiz (like maximize/minimize a window or rotate the cube etc). In the last few weeks while trying to figure this out, I used the gtkperf benchmark from phoronix suite which indeed has one pixmap test but I fear the pixmaps where too small to really give you obvious clear results. If you want to try this anyway install the phoronix-test-suite package and then run "phoronix-test-suite install gtkperf" followed by "phoronix-test-suite run gtkperf".

I was mostly confused by the numbers. Another option for a benchmark idea is to activate the compiz "benchmarking" plugin (install compizconfig-settings-manager package from universe, activate the plugin and then maximize say 10 firefox webpages and spin the cube with the compiz FPS benchmark plugin active).