Page 1 of 10 123 ... LastLast
Results 1 to 10 of 91

Thread: Ondemand governor dramatically slows down mesa perfomance

  1. #1
    Join Date
    Feb 2011
    Location
    Ukraine
    Posts
    116

    Default Ondemand governor dramatically slows down mesa perfomance

    With kernel >=3.8 and Mesa 9.2 and ondemand governor cpu frequency don't change in 3d.

    ondemand performance
    glxgears 3400 6200
    openarena 100 125
    lightsmark 240 340

    can somebody confirm this?

  2. #2
    Join Date
    Feb 2008
    Location
    Linuxland
    Posts
    4,991

    Default

    Kernel 3.7.10 and Mesa 9.1.1, but I see the same.

    glxgears pegs one core, but the frequency stays on low.


    Used this to check: watch -d -n 1 "grep MHz /proc/cpuinfo | uniq"

  3. #3
    Join Date
    Feb 2011
    Location
    Ukraine
    Posts
    116

    Default

    Quote Originally Posted by curaga View Post
    Kernel 3.7.10 and Mesa 9.1.1, but I see the same.

    glxgears pegs one core, but the frequency stays on low.


    Used this to check: watch -d -n 1 "grep MHz /proc/cpuinfo | uniq"
    cpupower frequency-set -g performance

    Increases the FPS?

  4. #4
    Join Date
    Feb 2008
    Location
    Linuxland
    Posts
    4,991

    Default

    Yes, the performance governor about doubles the fps.

  5. #5
    Join Date
    Jan 2009
    Location
    Vienna, Austria; Germany; hello world :)
    Posts
    622

    Default

    Quote Originally Posted by Pontostroy View Post
    With kernel >=3.8 and Mesa 9.2 and ondemand governor cpu frequency don't change in 3d.

    ondemand performance
    glxgears 3400 6200
    openarena 100 125
    lightsmark 240 340

    can somebody confirm this?

    works fine here

    just noticed that the conservative governor that I had used all the time didn't clock up all that much at all

    glxgears always stayed around 2500 - now with ondemand it goes up to 3800

    that's with a 5850

    what card is that on your side ?

    screen resolution is 1920x1080 here


    had to use R600_DEBUG=sb vblank_mode=0 glxgears to use its full potential

  6. #6
    Join Date
    Jan 2008
    Posts
    68

    Default

    Yes this is well known, and I always warn people for this on IRC
    when gaming always use performance governor it's +33-50% fps difference in many scenes
    (and get a fast cpu per core also 3-4+ GHz )

    ideally you would bind your game to 1 core (and mesa to another?) and only put that core into performance mode.

  7. #7
    Join Date
    Jan 2009
    Location
    Vienna, Austria; Germany; hello world :)
    Posts
    622

    Default

    Quote Originally Posted by pheldens View Post
    Yes this is well known, and I always warn people for this on IRC
    when gaming always use performance governor it's +33-50% fps difference in many scenes
    (and get a fast cpu per core also 3-4+ GHz )

    ideally you would bind your game to 1 core (and mesa to another?) and only put that core into performance mode.
    had the same issue occuring to me a few days ago

    so ondemand isn't working very reliably (it didn't clock the cpu up)


    you could also try conservative governor a try

    or try tweaking conservative & ondemand governor to clock up more aggressively

  8. #8

    Default

    I think it makes more sense to just use the performance governor all the time when running on a desktop, as you won't save that much power from using ondemand anyway. And on a laptop you can use laptop-mode-tools to set ondemand/conservative when on battery, and have it force to performance when AC power is supplied.

    Alternatively, you could setup up a script to force performance whenever you play a demanding game.

  9. #9
    Join Date
    Feb 2008
    Location
    Linuxland
    Posts
    4,991

    Default

    Well, it depends on how much is much. This desktop here has a 60W difference when running idle vs running with all cores full.

  10. #10

    Default

    Quote Originally Posted by curaga View Post
    Well, it depends on how much is much. This desktop here has a 60W difference when running idle vs running with all cores full.
    Yeah, that's pretty substantial. I remember testing it a long time back and it was more like 10W difference or something like that. But it probably varies by CPU and the scaling implementation; they have both probably gotten a lot better.

    I should test my current systems to check for myself.

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •