Announcement

Collapse
No announcement yet.

After A Bumpy Cycle, AMD Performance Will Shine Brighter On Linux 5.11

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

  • After A Bumpy Cycle, AMD Performance Will Shine Brighter On Linux 5.11

    Phoronix: After A Bumpy Cycle, AMD Performance Will Shine Brighter On Linux 5.11

    For those following the saga of the AMD frequency invariance regression on Linux 5.11 since the Christmas investigative benchmarking looking at the performance regressions, everything now looks like it will be buttoned up in time for the Linux 5.11 stable release. As noted yesterday, there was a curve ball this week in that the patch proposed by SUSE's Giovanni Gherdovich in January to address the frequency invariance regression was turned down by the Linux power management maintainer and instead he (Rafael Wysocki of Intel) proposed an alternative patch that instead modified the CPUFreq driver. Given it's getting late into the cycle, it's been a mad rush of re-conducting benchmarks on this new kernel patch and now it looks like that solution will be sent in the coming days for Linux 5.11.

    http://www.phoronix.com/vr.php?view=29909

  • #2
    Awesome!
    And thanks for your efforts to provide swift feedback!

    Comment


    • #3
      Would it make sense to include idle cycle counts for tests when changes to the scheduler are being made? ...maybe in a separate run to not obstruct the measurement too much... just out of curiosity.

      Comment


      • #4
        Given the improvements in some of the more bursty (and especially ST) tests, it looks like schedutil may have actually been outright PREVENTING the CPU from turboing at all, rather than it "only" failing to reach the right clock rates.

        While Michael deserves credit for identifying the problem and the bulk of the root cause here, I think the bigger takeaway is that schedutil is basically trying to tackle a problem that some of its devs don't even fully UNDERSTAND. Given that they've proven over and over again that they literally don't even bother testing the code in any meaningful way, ISTM that the only way this thing will ever actually usable is by chance more than anything else - which fits pretty well with the literal YEARS of utter failure that are all it's managed so far.

        I'm not saying it's an EASY problem and that the devs are just stupid - it's tricky stuff. But I AM saying that the fact that they don't even test it AT ALL is entirely on them, and if you have a problem that you don't even understand correctly, and then you don't even bother trying TO understand it better by doing that testing, it's no surprise that schedutil has been garbage for as long as it has.

        Still, this once, at least, it's managed to take a step in the right direction. Progress!

        Comment


        • #5
          Thanks Michael ! I send you a tip to get yourselve a well deserved beer!

          Comment


          • #6
            Originally posted by baka0815 View Post
            Thanks Michael ! I send you a tip to get yourselve a well deserved beer!
            more beer for michael is always a good investment.
            Phantom circuit Sequence Reducer Dyslexia

            Comment


            • #7
              So, should I switch to using schedutil on my AMD laptop, when 5.11 hits? I'm using ondemand right now.

              Comment

              Working...
              X