Fedora 12 vs. Ubuntu 9.10 Benchmarks

[Dragoran]

Originally Posted by ldesnogu

The question here was not to measure exactly the impact of compositing by using glxgears, but to see there is an impact.

All we need to know is whether Michael left compiz on when running the gaming benchmarks.

Still glxgears is the wrong way to do this, because just running a DRI2 driver has an negative impact on glxgears, while real apps tend to be faster/not impacted.

glxgears just renders a very simply frame as often as possible so _any_ additional work could have a great impact on it.

So if you want to test the impact just use any real app.
It would be close to zero if measurable at all.

[ldesnogu]

OK, you know better than I do about this

So would you have any guess about why Michael's results were so bad?

[BlackStar]

Originally Posted by natewiebe13

Hmm.. That's strange, but my guess is because it has to do with Wine. I just ran the standard (glxgears -fullscreen) on an old computer of mine. I'm running Ubuntu Karmic with an nVidia 7600GS and the nVidia 185 Drivers from the repositories. With effects set to standard (Compiz), I average at around 68 FPS, if I turn off Comiz by setting effects to none, I average 228 FPS. So there is definitely a chance that the reason Fedora benched higher was because Comiz was still enabled.

I have a gut feeling that Karmic disables the "unredirect fullscreen windows" option by default. I'm saying that because popup notifications no longer cause my fullscreen applications to flicker and flash like crazy, like they used in Jaunty.

While such a change would make for a better desktop experience, it would impact fullscreen 3d performance quite a bit, especially on lower-end / bandwidth-limited hardware (redirected windows cannot page-flip, so they have to copy memory from back- to front-buffers every frame - which translates into a few hundred MB/sec of "lost" texture bandwidth).

[damentz]

It's not compiz.

It would make sense for Fedora to backport the fixes in 2.6.32 of CFS due to the performance regressions exposed during benchmarks against BFS. They tend to grab a lot of new features and mix stable with new when there's no obvious stability penalty.