No announcement yet.

Mesa Looks At Switching To Jemalloc For Faster Performance

  • Filter
  • Time
  • Show
Clear All
new posts

  • #51
    Originally posted by bridgman View Post
    Isn't mesa being used on raspberry pi and similar systems as well (the ones Eric A is working on) ?
    rpi falls in the range of many gigabytes of storage(technically you could buy sdcard with small nubmer of gigabytes, but it makes zero sense)
    scaled to jemalloc, linus should be worried of increasing kernel size by hundreds or even dozens of bytes. he has to stop accepting patches to make bloatphobes happy


    • #52
      Originally posted by pal666 View Post
      it is
      I don't think you know what "irrational" means. But y'know what is irrational? Your arrogance.
      kernel is used on routers with few megabytes of storage. mesa is used on desktops with few terabytes or many gigabytes of storage
      Irrelevant point - do you really think the modern kernel is used as-is on routers? Because it isn't - it's heavily modified.


      • #53
        Originally posted by pal666 View Post
        i said storage, you replied with ram
        ok. then rpi3 has 0 bytes storage. provide your own sd card extra as big or small as you like. mesa is used on fairly small systems with small amounts of storage. it's generally the RAM that is the biggest limitation (routers generally excepted).

        Originally posted by pal666 View Post
        correct question is why the fuck whine about jemalloc dependency, considering

        $ rpm -q --qf '%{SIZE} %{NAME}\n' mesa-dri-drivers.x86_64 llvm-libs.x86_64 jemalloc
        38501303 mesa-dri-drivers
        43403008 llvm-libs
        604751 jemalloc
        well you don't need to build ALL drivers into mesa... it doesn't have to be that big. not when you only have a single gpu to support. yeah llvm is needed and that's also not small. it really doesn't much matter. if it speeds up shader compiling by 10% its almost definitely worth it.


        • #54
          Originally posted by schmidtbag View Post
          do you really think the modern kernel is used as-is on routers? Because it isn't - it's heavily modified.
          Where "heavily modified" is translated as :
          - uses a couple of binary proprietary only driver for a few key component (usually WLAN and/or DSL modem or DOCSIS cable modem)
          - uses a static platform definition on most router (they don't rely on a standard device discovery scheme, unlike x86 hardware), though these tend to get modularized over time and start using some form of proper device tree recognition. (e.g.: Most Raspberry Pi models are now supported by the latest mainstream Linux Kernel 4.9)

          The rest is just plain vanilla kernel.
          (Including file system used on internal flash (e.g.: jffs2) or on external USB flash (FAT32), including the networking and filtering stack, etc.)
          (but it's also configured with a lot of useless stuff disabled from the kernel)

          The only reason why you cannot currently compile kernel 4.9 on any random router, is that the kernel will fail to understand what is available beyond the CPU on which it is running (what other devices are present on the SoC or elsewhere in the PCB ?) and that the non-ethernet interface will probably lack any open-source drivers.
          At least Linaro is helping a lot for the first part.
          Some are also helping the later part (Atheros and Intel hardware tend to have opensource drivers. Some broadcom hardware is supported by "mac80211" driver) and given that lots of router actually use a small form factor card for WLAN (e.g.: a miniPCIe port holding the WiFi card) you might actually end-up being able to install the latest vanilla kernel on your router.