Announcement

Collapse
No announcement yet.

Intel P-State Gets More Cleanups & Optimizations

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • #11
    Originally posted by aufkrawall View Post
    In my test with Talos Principle Vulkan, ondemand btw. failed badly, it was even quite slower than pstate powersave. So probably schedutil would also be a good choice for Ryzen CPUs.
    In my tests with Talos Vulkan, ondemand was slightly better than intel-pstate, on a 4690K and linux 4.9.

    Comment


    • #12
      I've been having hard lockups with kernels newer than 4.8 on my Skylake system as have others where the 2nd resume from suspend locks the system up. I have a similar issue with logrotate service at midnight causing a lockup, but unlike the first issue which can be resolved by disabling async power management(echo 0 > /sys/power/pm_async), it's only fixed by changing the IO scheduler from BFQ to something else like CFQ, this also fixes the suspend/resume lockup.

      Bit confused that the first issue can still use BFQ as long as pm_async is disabled in 4.9 and newer. Is it possible using p-state or a different governor/scheduler for CPU, or preventing certain power states would also fix the issue? It's been suggested to do a bisect and compile vanilla kernels(I'm using Manjaro) but I've no experience with that or I'd assume the time required to do that many times :|

      Comment


      • #13
        Originally posted by polarathene View Post
        t's been suggested to do a bisect and compile vanilla kernels(I'm using Manjaro) but I've no experience with that or I'd assume the time required to do that many times :|
        No one had that experience before doing that
        It's not very hard, though it can be a bit slow... I use ccache so it's not too bad.

        Comment


        • #14
          I tried both P-state and acpi-cpufreq recently on a 4.10 kernel, P-state performance was giving me choppiness/stutters in games and other desktop 3D tasks under full load, like compiling something and moving a 1080p video or browsing Google Maps in Chromium.

          Was also getting choppiness/stutters in certain games with nothing else running at all.

          Stuff that I didn't experience with schedutil or performance governors from acpi-cpufreq, or much less, I think they're a bit better in keeping desktop applications performing well under full load, and evenly distributing processor resources.

          Comment


          • #15
            Originally posted by aufkrawall View Post
            Gave SC2 a quick try in Wine staging, no sound cracking to report with schedutil.
            Game shows some stuttering though, but also with cpufreq performance. Probably the typical non-optimal performance caused by Wine.
            Strange

            Rafael gave me a patch against linux-next of linux-pm, I just tried it and schedutil seems to be fine now. So there is a lot of hope for 4.12

            Comment


            • #16
              Cool, thx for letting me know!

              Comment


              • #17
                Originally posted by polarathene View Post
                I've been having hard lockups with kernels newer than 4.8 on my Skylake system as have others where the 2nd resume from suspend locks the system up. I have a similar issue with logrotate service at midnight causing a lockup, but unlike the first issue which can be resolved by disabling async power management(echo 0 > /sys/power/pm_async), it's only fixed by changing the IO scheduler from BFQ to something else like CFQ, this also fixes the suspend/resume lockup.

                Bit confused that the first issue can still use BFQ as long as pm_async is disabled in 4.9 and newer. Is it possible using p-state or a different governor/scheduler for CPU, or preventing certain power states would also fix the issue? It's been suggested to do a bisect and compile vanilla kernels(I'm using Manjaro) but I've no experience with that or I'd assume the time required to do that many times :|
                BFQ is an io scheduler. not a cpu scheduler.

                Comment


                • #18
                  Originally posted by aufkrawall View Post
                  Cool, thx for letting me know!
                  Sure!

                  I also forgot to mention I use PulseAudio, with resample-method = speex-float-10, maybe that's where the difference is between our 2 setups? or maybe not...

                  What do you have in your setup?
                  Last edited by geearf; 30 March 2017, 11:52 PM.

                  Comment


                  • #19
                    Oh latest comment I got from the dev, if the patch that fixes my issue with schedutil is accepted upstream (it may not, as it may increase power consumption...), then the behavior of scheduitl and pstate powersave should be fairly similar. Then I'd probably use pstate passive and schedutil. No real reason though, it just seems like a good choice

                    Comment


                    • #20
                      Originally posted by geearf View Post
                      Oh latest comment I got from the dev, if the patch that fixes my issue with schedutil is accepted upstream (it may not, as it may increase power consumption...), then the behavior of scheduitl and pstate powersave should be fairly similar. Then I'd probably use pstate passive and schedutil. No real reason though, it just seems like a good choice
                      I'm on PulseAudio too, but with default settings.

                      Could you please elaborate what "pstate passive" means?

                      Edit: What does i7z tell you? Is there a noteable difference in clock/vcore/states with the patch compared to without?
                      i7z terminal seems to be the only tool that reads out correct values for me when the pstate module isn't loaded.
                      Last edited by aufkrawall; 31 March 2017, 03:04 PM.

                      Comment

                      Working...
                      X