Announcement

Collapse
No announcement yet.

Ubuntu 32-bit, 32-bit PAE, 64-bit Benchmarks

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

  • yuhong
    replied
    My thread on this in LKML

    I have created a thread on this in LKML right there:

    Leave a comment:


  • Dragoran
    replied
    Originally posted by caramerdo View Post
    So is it a 64 bit kernel on 32 bit userland or a 64 bit kernel with 64 bit userland? That makes a world's difference.
    The later as I already said just switching the kernel can't and does NOT provide performance gains like this.

    Leave a comment:


  • caramerdo
    replied
    So is it a 64 bit kernel on 32 bit userland or a 64 bit kernel with 64 bit userland? That makes a world's difference.

    Leave a comment:


  • yuhong
    replied
    Originally posted by Kano View Post
    @xeros
    Your bios has to support remapping too.
    And the chipset too, and unfortunately Intel did not add support to their desktop chipsets until the i955X and i965 series, so the only option is for the BIOS to free more physical address space below the 4GB line. As you have only 3GB of RAM, this should be easy, try changing some BIOS options (maybe relating to the AGP aperture) or updating the BIOS.

    Leave a comment:


  • Kano
    replied
    @xeros

    Your bios has to support remapping too.

    Leave a comment:


  • yuhong
    replied
    BTW, the article says that the system used for testing has 4GB of RAM, was that with memory mapped over the 4GB line or not? If in doubt, look at the beginning of the dmesg for the e820 map, and look for memory mapped above 0x100000000. This is a detail particularly important for DMA to devices.

    Leave a comment:


  • xeros
    replied
    I've just tested PAE enabled kernel in Kubuntu 9.10 on P4 3GHz (32bit only) with HT and 3GB RAM physical (2,5-2.7GHz logical mapped by BIOS) and I didn't get more logical RAM with PAE - it's the same as without PAE but it looks like running from BIOS to... KDE 4.4b2 (two instances run by two users) takes at least few seconds more with PAE (I haven't measured that precisely).

    Leave a comment:


  • chithanh
    replied
    I think there is an error in the article, HIGHMEM_4G enables highmem but not PAE. Only HIGHMEM_64G will enable PAE. HIGHMEM_4G performance hit is measurable in some cases, but only significant if the work dataset is larger than lowmem.

    Leave a comment:


  • monraaf
    replied
    Originally posted by clavko View Post
    I know it all gets down to
    out-of-box experience, but AnandTech took the same route with battery
    tests and it didn't bode well for GNU/Linux. At all.

    Cheers!
    LOL, so at anandtech they found out what most of us already know:

    We were left with running the proprietary fglxr 8.600 driver, and while it worked fine in general we had problems with DVD playback. VLC repeatedly crashed during our benchmarks, sometimes after a few minutes, sometimes after 50 minutes. Eventually, we decided to uninstall the proprietary ATI driver and test out the open-source MESA driver. Surprisingly, the open-source driver actually provided a better experience.

    Leave a comment:


  • tettamanti
    replied
    Originally posted by linuxjacques View Post
    Can anyone explain the results of the 64-bit Postmark test?

    I think the explanation of the Apache test results is related.
    Apache result is indeed strange... the test uses the prefork mpm with 100 concurrent request so apache spawns about the same number of processes, but it's nowhere near to putting pressure to the VM (plus apache should use zero-copy sendfile for static content).
    My sheevaplug with a tuned apache easily outperforms that CD2 in 32bit mode...
    Another slight problem with this (local) ab test is that basically one core is lost to ab itself.

    Leave a comment:

Working...
X