After the benchmarks a few days back of Intel Sandy Bridge Acceleration On Non-SNB Hardware, Chris Wilson of Intel who has been responsible for much of the "Sandy Bridge New Acceleration" work requested more tests, but this time to see the effect that the compositing window manager has on this new acceleration architecture. As a result, here is some quick tests of Intel's Sandy Bridge graphics under the Unity, Unity 2D, and GNOME Shell desktops.
This was last minute testing before XDC2011 Chicago and Oktoberfest, so this article is a bit brief. With an Ubuntu 11.10 development snapshot as of 8 September, the Mesa and DDX were pulled from Git. The latest Linux 3.1 kernel was also employed. The test system was the HP EliteBook Sandy Bridge system with the Core i5 2520M CPU.
A variety of 2D/3D tests were carried out with the xf86-video-intel Git driver in its stock configuration under the Unity, Unity 2D, and GNOME Shell desktops. These tests were then carried out a second time when the DDX driver was built with the SNA acceleration architecture enabled.
Chris saw these results in advance. Here's his commentary about them.
Thanks, I had just been looking through those. Those results are just as perplexing! Eventually I'll find the missing bit of configuration that is causing such disagreement between me and the rest of the world. ;(
When running such tests locally, the impact of having to flush the render queue after handling each client and the overhead of having to perform damage tracking demonstrates the stark performance disadvantage of the X/Compositing Manager split.
Back to finding the source of the few regressions (which I think I have already) but more importantly why our results occasionally differ by an order of magnitude for similar hardware...
Into the xf86-video-intel Git repository since then, he has already made some more performance optimizations for Sandy Bridge New Acceleration thanks to the Phoronix testing.