Performance regression between 2.6.35 and 3.0

[Michael Hope]

Using the current git head, CPU bound tests like CoreMark run at 78 % of the speed of a 2.6.35 kernel.

The .config, branches and refs of the 2.6.35 kernel are here:
 http://bazaar.launchpad.net/~michaelh1/+junk/hardware/files/head:/ursa/r2/

which is basically kernel-ubuntu from dev.omapzoom.org.

This is with linux-linaro-3.0 v3.0.3-751-gdbf7603. The .config of the 3.0 kernel is attached. It's from hwpack_linaro-panda_20110821-0 with the ethernet driver and EXT4 linked in.

This is on an A2 PandaBoard running Natty. The board is perfectly stable under 2.6.35.

[Michael Hope (michaelh1)]

/proc/cpuinfo reads ~1500 BogoMips vs ~2000 on the earlier kernel as well.

[Michael Hope (michaelh1)]

Adding 'nosmp' to the kernel command line has no effect on the CoreMark score.

[Rajendra Nayak (rnayak)]

The difference in BogoMips, I guess is due to the different bootlaoders used with the .35 kernel and the 3.0 one.
I see in the Bootlogs attached that SPL is used with the 3.0 kernel, and FWIK SPL configures the MPU on OMAP4 to
be at 792 Mhz (100% OPP) and the x-loader that would have been used with the .35 kernel was instead configuring the
MPU at overdrive (1Ghz) and so the performance drop is probably expected.