Announcement

Collapse
No announcement yet.

ACPI CPUfreq vs. Intel P-State Scaling With Linux 3.15

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

  • ACPI CPUfreq vs. Intel P-State Scaling With Linux 3.15

    Phoronix: ACPI CPUfreq vs. Intel P-State Scaling With Linux 3.15

    For this weekend's Linux benchmarks we are looking at the performance of the Intel P-State and ACPI cpufreq drivers and comparing their scaling governor options when testing from an Intel Core i7 Extreme Edition system running with the Linux 3.15 development kernel.

    http://www.phoronix.com/vr.php?view=20399

  • #2
    This makes no sense, Michael.

    The ACPI performance governor cannot be slower than ondemand one. It just cannot - it keeps the CPU at its highest frequency all the time.

    Comment


    • #3
      Originally posted by birdie View Post
      This makes no sense, Michael.

      The ACPI performance governor cannot be slower than ondemand one. It just cannot - it keeps the CPU at its highest frequency all the time.
      But does Turbo Boost works when performance governor holds the CPU at highest frequency?

      Comment


      • #4
        Yes, it surely does.

        Comment


        • #5
          When timing the compilation of the Linux kernel with these different CPU scaling settings, there was a huge difference. The Intel P-State Performance configuration led to the best results that were timed the same as ACPI CPUfreq with the on-demand governor while the other configurations were about twice as slow!
          Something is really wrong with scaling governors on linux... :/ How can "On-demand" be faster and much more power consuming than "Performance"? What kind of bullshit is this?

          Comment


          • #6
            Originally posted by Nuc!eoN View Post
            Something is really wrong with scaling governors on linux... :/ How can "On-demand" be faster and much more power consuming than "Performance"? What kind of bullshit is this?
            It is.
            On my atom netbook, even while idling, acpi powersave governor drains more power than the performance one, and is slower.

            Comment


            • #7
              Originally posted by kokoko3k View Post
              On my atom netbook, even while idling, acpi powersave governor drains more power than the performance one, and is slower.
              Which makes full sense. Power saving on modern CPUs is about the race-to-idle - finish tasks as fast as possible, so that the CPU can get to idle as soon as possible. The powersave governor, contrary to its name, does not save power because it makes tasks take longer to finish, which works against the race-to-idle.

              Because of clock and power gating, C-states are more important than P-states. Doesn't matter much that the performance governor keeps the CPU in the highest P-state all the time, because in deep C states the clock is gated and various parts of the CPU are power gated. Making tasks finish faster allows the CPU to spend more time in these gated C states.

              Comment


              • #8
                Originally posted by birdie View Post
                This makes no sense, Michael.

                The ACPI performance governor cannot be slower than ondemand one. It just cannot - it keeps the CPU at its highest frequency all the time.
                because it keeps the cpu cooler so "turbo boost" can be turned on more often (cpu's with boost check their temperature before "using" it)

                Comment


                • #9
                  Originally posted by kokoko3k View Post
                  It is.
                  On my atom netbook, even while idling, acpi powersave governor drains more power than the performance one, and is slower.
                  using the powersave governor is not enough- you need to use something like tlp to automatically force a low max cpu frequency to maximize it's effects.

                  Btw, every time i try to make a custom kernel with pstates, it crashes(as in, kernel panic so deep that i have to pull the battery) as soon as i pull the power cable- it does not like to be told what to do,apparently, so i stick with acpi.

                  Comment


                  • #10
                    intel p state

                    p-state benefits mostly when thermal daemon https://github.com/01org/thermal_daemon is also available and installed, and one side-note can someone do this on a notebook/laptop?

                    Comment


                    • #11
                      for best results use https://github.com/01org/thermal_daemon thermal daemon along with p-state

                      Comment


                      • #12
                        Originally posted by sireangelus View Post
                        using the powersave governor is not enough- you need to use something like tlp to automatically force a low max cpu frequency to maximize it's effects.
                        Using powersave AT ALL isn't enough, if you're actually trying to save power. Tasks will take longer to complete. The ONE use I've found for powersave is for thermalcapping when my laptop's CPU fan wasn't working. Using default P-State would make the laptop overheat under load, but using powersave let it run and run albeit slower.

                        Originally posted by sireangelus View Post
                        Btw, every time i try to make a custom kernel with pstates, it crashes(as in, kernel panic so deep that i have to pull the battery) as soon as i pull the power cable- it does not like to be told what to do,apparently, so i stick with acpi.
                        Read the panic and see if it tells you if something is missing. P-States might depend on another config option. All I can tell ya is they work perfectly fine on Fedora and Arch for me. (You could have also found a CPU-specific bug, but its a little less likely haha)

                        Comment


                        • #13
                          Yea, the results make no sense. Why is pstate performance sometimes slower and using less power than pstate powersave?

                          Comment


                          • #14
                            Originally posted by birdie View Post
                            This makes no sense, Michael.

                            The ACPI performance governor cannot be slower than ondemand one. It just cannot - it keeps the CPU at its highest frequency all the time.
                            Are you sure? My understanding of the performance governor was that it only had two states: max freq(turbo?) and idle (not sure which idle states... S0i1?). Perhaps the problem is that it is to aggressive going to idle and the resulting transition times, for this hardware, are causing performance hits?

                            Comment


                            • #15
                              Great article Michael! Best I've seen in awhile.
                              Would do mind attaching that wattmeter to the Nvidia board so we can get some official power numbers?

                              Comment

                              Working...
                              X