[Intel-gfx] Intel 945GM with >=2.6 and newer kernels: massive performance and or memory leak problems

[Eric Anholt]

On Sun, 2009-03-29 at 17:51 +0200, Thomas Bächler wrote:
> This is a summary of my experience with latest Intel developments: All 
> comments below apply to a KDE 4.2 Desktop with compositing enabled on 
> Arch Linux x86_64 and an Intel 945GM / Core 2 Duo CPU.
> 
> - 2.4 driver, Linux 2.6.28 and older, don't remember the Mesa version:
> Using EXA, performance was okay, but not very good. Leaked a few hundred 
> MB of memory within a week of active usage.
> 
> - 2.6 driver, Linux 2.6.28, Mesa 7.3:
> Using EXA, the performance is bad, you can count the FPS with the bare eye.
> Using UXA, it is amazingly fast (faster as ever before), but it leaks 
> hundreds of megabytes of memory within hours. After two or three 
> suspend/resume cycles and one day of active usage, the X server uses 
> over 1GB of memory! Even without suspend/resume cycles, it may use over 
> 500MB of memory within only 6 hours.
> This applies to both the 2.6.3 and 2.6.99.902 driver.
> 
> - 2.6 driver, Linux 2.6.29, Mesa 7.3
> Using EXA or UXA, the performance is very bad (as with EXA in the above 
> section), I have to disable compositing to be able to work at all. This 
> seems like a huge performance regression in 2.6.29! It is the same with 
> and without KMS enabled.
> I can't say anything about the memory problems, as I didn't use it with 
> compositing long enough to find out.
> 
> - 2.6 driver, linux-2.6.git tree, Mesa 7.3
> Basically, same as the last section, except the "weird colors" are fixed 
> when enabling KMS.
> 
> So my question is:
> 1) How do I get composite performance back with 2.6.29 and newer kernels?
> 2) How do I get rid of the massive memory leaks I experience with UXA?
> 
> Thanks for your replies. I hope me joining this mailing list can help 
> sorting these problems out, as sadly the intel driver is becoming less 
> useful for me over time.

If performance regressed for you with updating the kernel, could you
bisect to find when it happened?

Or, assuming your workload is CPU-bound, use sysprof to see where your
time is being spent.

-- 
Eric Anholt
eric at anholt.net                         eric.anholt at intel.com


-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 197 bytes
Desc: This is a digitally signed message part
URL: <http://lists.freedesktop.org/archives/intel-gfx/attachments/20090330/3899dc01/attachment.sig>