Interactivity degrades with CONFIG_[FAIR]_GROUP_SCHED set in 2.6.24 and later
First of all, I am very excited to see the CFS in the kernel. Interactivity is really great; much better than before.
When going from kernel 2.6.23 to 2.6.24, however, I noticed that interactivity (namely the smoothness of mouse motion) was noticeably less smooth under load. The most obvious case was while pulling down a Subversion repository; the mouse would freeze for quite long periods while being moved (i.e. a large fraction of a second), which never happens under 2.6.23. I was using ZFS-FUSE on the partition I was writing to, and it uses a lot of CPU at times, but 2.6.23 seems to handle it better.
After reading many posts on similar sounding issues, I decided to try getting later versions of the kernel: 2.6.25-rc5 and the latest 2.6.25 from the git archive (as of yesterday, 3/16) in case recent patches addressed this. They both still exibit the issue, however.
I then tried compiling without CONFIG_GROUP_SCHED (which also turned off CONFIG_FAIR_GROUP_SCHED and a number of other options), and the behavior seems to be back to the (better) 2.6.23 performance, at least has far as what I can tell from my tests.
After playing around with the later kernels with CONFIG_GROUP_SCHED enabled, I noticed a number of random times the mouse was not smooth and seemed to not be getting an "interactive" amount of CPU time, but I have a couple of tasks that repeat the problem reliably, so if there is any debug information I can provide, please let me know, and I'll be glad to help out in any way I can.
Thanks, Joe
Quoted 8 lines · Joe Peterson
When going from kernel 2.6.23 to 2.6.24, however, I noticed that interactivity (namely the smoothness of mouse motion) was noticeably less smooth under load. The most obvious case was while pulling down a Subversion repository; the mouse would freeze for quite long periods while being moved (i.e. a large fraction of a second), which never happens under 2.6.23. I was using ZFS-FUSE on the partition I was writing to, and it uses a lot of CPU at times, but 2.6.23 seems to handle it better.
I just made a couple of debug output files using the default CONFIG_GROUP_SCHED setting in 2.6.25-rc5-git (from 3/16):
This one was done during a diff between two kernel source dirs:
http://www.skyrush.com/downloads/linux-cfs-debug/cfs-debug-info-2008.03.18-06.37.44
This one was done during an svn checkout:
http://www.skyrush.com/downloads/linux-cfs-debug/cfs-debug-info-2008.03.18-06.38.23
The latter one was done during a period exhibiting what felt like longer and more profound pauses in mouse movement, however, the first debug file shows several se.wait_max > 40. Both cases showed very choppy/jerky mouse movement with quite obvious freezes when moving the mouse in circles during the tests.
Here is a snippet of the se.wait_max lines from the first file:
se.wait_max : 0.000000 se.wait_max : 0.000000 se.wait_max : 7.375852 se.wait_max : 4.875690 se.wait_max : 7.211323 se.wait_max : 3.779489 se.wait_max : 2.770382 se.wait_max : 12.959876 se.wait_max : 3.524036 se.wait_max : 16.625484 se.wait_max : 73.991289 se.wait_max : 18.937293 se.wait_max : 6.955806 se.wait_max : 21.919110 se.wait_max : 6.922932 se.wait_max : 10.432544 se.wait_max : 76.222556 se.wait_max : 7.216472 se.wait_max : 90.987582 se.wait_max : 7.071110 se.wait_max : 16.332227 se.wait_max : 2.261235 se.wait_max : 67.892956 se.wait_max : 25.337064 se.wait_max : 44.180597 se.wait_max : 74.290628 se.wait_max : 12.631292 se.wait_max : 1.389533 se.wait_max : 3.653521 se.wait_max : 2.841697 se.wait_max : 16.234723 se.wait_max : 107.011747 se.wait_max : 110.638560 se.wait_max : 9.677538 se.wait_max : 15.852109 se.wait_max : 23.316106 se.wait_max : 18.428330 se.wait_max : 16.674552 se.wait_max : 110.630411 se.wait_max : 47.906824 se.wait_max : 74.191984 se.wait_max : 12.224021 se.wait_max : 2.184557 se.wait_max : 17.799343 se.wait_max : 17.670697 se.wait_max : 0.000000 se.wait_max : 0.000000
I hope these help; please let me know if I can help further or run more tests.
-Thanks, Joe