The Cost Of Running Compiz
Beginning with Warsow we see the Intel Clarkdale graphics were 3~4 FPS faster when Metacity was running rather than Compiz, which with the low-speeds of that hardware and ill-optimized Linux driver equates to being about a 15% difference. With our Warsow test and the classic Mesa R600/700 OpenGL driver, the frame-rates were not more than three frames per second so we do not have any valid numbers here. With the Catalyst driver on the Radeon HD 4830 graphics card we see Compiz played little role into the performance of this game with the frame-rates being about the same. Lastly, with NVIDIA's proprietary driver, Compiz definitely caused a performance hit. When Compiz was not running, the frame-rates were 20% higher with the NVIDIA GeForce 9800GT graphics card.
With the OpenArena game, disabling the Compiz desktop effects caused Intel's graphics performance to improve again by 15%. We also have our first numbers for the open-source Radeon driver. Using the modern Radeon KMS-DRI2 driver that is the default for Ubuntu 10.04 there too was a 15% penalty incurred when running Compiz. However, with the older UMS-DRI1 code-paths for the open-source Radeon driver the frame-rate was the same between Compiz and Metacity.
Like in our first test with Warsow, the Catalyst driver performed the same regardless of whether Compiz was in use. Likewise, the NVIDIA driver again took a performance hit on its binary driver when Compiz was enabled. Running with just Metacity had boosted the 9800GT frame-rate by about 26%.
With World of Padman the performance drop caused by Compiz is even worse. The frame-rate dropped from 42 to 20 FPS simply when using the Compiz desktop effects that are the default in Ubuntu. With the Radeon driver using KMS-DRI2 the frame-rate of Metacity over Compiz increased from 15% with OpenArena to 22% with World of Padman. However, with the older UMS-DRI1 code-paths, there again was no change in performance when enabling Compiz nor was there any change when the Catalyst driver was being used. The NVIDIA driver was 11% faster when there were no desktop effects.