Announcement

Collapse
No announcement yet.

The v2 Rotary Interactivity Favor Scheduler

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

  • sure:

    http://pastebin.com/gaUv52Ue (valid 1 month)

    Comment


    • Originally posted by kernelOfTruth View Post
      sure:

      http://pastebin.com/gaUv52Ue (valid 1 month)
      CONFIG_NO_HZ=y

      Could you turn it off and re-compile the kernel ?
      Thanks.

      As I say
      The algorithm of RIFS-ES can be affected by variable clock interrupt.

      Comment


      • Originally posted by 3766691 View Post
        CONFIG_NO_HZ=y

        Could you turn it off and re-compile the kernel ?
        Thanks.

        As I say
        The algorithm of RIFS-ES can be affected by variable clock interrupt.

        sure, will do

        I read that there isn't that much of more power consumption:

        http://www.phoronix.com/scan.php?pag...item=651&num=1


        where others claim 10-20% more consumption:

        http://ubuntuforums.org/archive/index.php/t-524694.html


        now raw performance and efficiency counts (in terms of workflow) so I'll revert to tickless or more energy-friendly settings in 1-2 weeks ^^

        Comment


        • well, that massively helped

          pulseaudio now continually seems to stay at 1% cpu load instead of pegging around 100-300%


          it could have been partly an issue caused by configuration but I believe it's partly also due to tickless:

          Originally posted by diff config-3.4-geek_rifs_v2 config-3.4-geek_RIFS_ES-no-tickless_less-debug
          3c3
          < # Linux/x86_64 3.4.2-geek Kernel Configuration
          ---
          > # Linux/x86_64 3.4.2-geek_btrfs+_libata_RIFS-ES-v1-low-spec Kernel Configuration
          113d112
          < CONFIG_RCU_FAST_NO_HZ=y
          319c318
          < CONFIG_NO_HZ=y
          ---
          > # CONFIG_NO_HZ is not set
          326a326
          > CONFIG_SCHED_OMIT_FRAME_POINTER=y
          553d552
          < CONFIG_CPU_IDLE_GOV_MENU=y
          2113a2113
          > # CONFIG_NTP_PPS is not set
          3754,3755c3754
          < CONFIG_UNWIND_INFO=y
          < CONFIG_STACK_UNWIND=y
          ---
          > # CONFIG_UNWIND_INFO is not set
          3759c3758
          < CONFIG_RCU_CPU_STALL_VERBOSE=y
          ---
          > # CONFIG_RCU_CPU_STALL_VERBOSE is not set

          it's still for 1-2 seconds stalling when it's heavily writing even though the dirty values are low as "1" - but that obviously might be caused by either luks/device-mapper or the filesystems and the other subsystems (VFS, writeback, etc.) involved


          last kernel that - in my opinion - had really great writeback performance was 2.6.34, dunno


          will stay on this kernel now ...


          thanks !
          Last edited by kernelOfTruth; 06-23-2012, 05:38 AM.

          Comment


          • Originally posted by kernelOfTruth View Post
            sure, will do

            I read that there isn't that much of more power consumption:

            http://www.phoronix.com/scan.php?pag...item=651&num=1


            where others claim 10-20% more consumption:

            http://ubuntuforums.org/archive/index.php/t-524694.html


            now raw performance and efficiency counts (in terms of workflow) so I'll revert to tickless or more energy-friendly settings in 1-2 weeks ^^
            It seems that Tickless just help very little on power saving. On the other hand using tickless system can cause significantly high latency

            Comment


            • Originally posted by kernelOfTruth View Post
              well, that massively helped

              pulseaudio now continually seems to stay at 1% cpu load instead of pegging around 100-300%


              it could have been partly an issue caused by configuration but I believe it's partly also due to tickless:




              it's still for 1-2 seconds stalling when it's heavily writing even though the dirty values are low as "1" - but that obviously might be caused by either luks/device-mapper or the filesystems and the other subsystems (VFS, writeback, etc.) involved


              last kernel that - in my opinion - had really great writeback performance was 2.6.34, dunno
              How about the subjective feeling between RIFS and RIFS-ES-Low-Spec ?

              Comment


              • Originally posted by 3766691 View Post
                How about the subjective feeling between RIFS and RIFS-ES-Low-Spec ?
                like I wrote - I need that box for production - so I won't compile another RIFS with NOHZ right now (next week or after that at earliest)


                otherwise it would be RIFS + tickless vs. RIFS-ES-low-spec + NOHZ

                and that wouldn't be fair


                in my opinion RIFS was more consistent but less fluid whereas RIFS-ES-low-spec was more smooth but at times also less smooth

                this could be a whole different story with tickless disabled

                Comment


                • Originally posted by kernelOfTruth View Post
                  like I wrote - I need that box for production - so I won't compile another RIFS with NOHZ right now (next week or after that at earliest)


                  otherwise it would be RIFS + tickless vs. RIFS-ES-low-spec + NOHZ

                  and that wouldn't be fair


                  in my opinion RIFS was more consistent but less fluid whereas RIFS-ES-low-spec was more smooth but at times also less smooth

                  this could be a whole different story with tickless disabled
                  NOHZ and tickless are the same things.

                  Actually I can't feel difference between RIFS and -ES but with benchmark data -ES would gain more improvement

                  Comment


                  • Originally posted by 3766691 View Post
                    NOHZ and tickless are the same things.

                    Actually I can't feel difference between RIFS and -ES but with benchmark data -ES would gain more improvement
                    ah yes - I actually meant # NOHZ is not set


                    yeah, the differences are very subtle but under load RIFS-ES probably can really shine and show its strength



                    RIFS/-ES needs more testers and experiences to further improve and become even better


                    anyone ?



                    edit:

                    it's actually ridiculous how almost every app seemed to hog the cpu with NOHZ=y
                    (between 50-100% per app)

                    it might have been issues with the kernel over longer runtime - I'll see at the end of the time if it shows the same behavior or if it stays that calm with 0-1%
                    Last edited by kernelOfTruth; 06-23-2012, 05:56 AM.

                    Comment


                    • Originally posted by kernelOfTruth View Post
                      ah yes - I actually meant # NOHZ is not set


                      yeah, the differences are very subtle but under load RIFS-ES probably can really shine and show its strength



                      RIFS/-ES needs more testers and experiences to further improve and become even better
                      Ah, yes
                      Now the last things is we have to make it into modular.(Very last things) After this RIFS/-ES will support CGroups.
                      RIFS/-ES have beaten BFS but now we have to promote it more. If Gentoo developer can help me to promote I will say a big THANKS.

                      it might have been issues with the kernel over longer runtime - I'll see at the end of the time if it shows the same behavior or if it stays that calm with 0-1%
                      I seldom run my computer for long time but now I will try. Do you play LOL?

                      Comment

                      Working...
                      X