Announcement

Collapse
No announcement yet.

Linux's Load Balancer Still Needs To Be Better Adapted For Intel Hybrid CPUs

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

  • Linux's Load Balancer Still Needs To Be Better Adapted For Intel Hybrid CPUs

    Phoronix: The Linux Kernel's Load Balancer Still Needs To Be Better Adapted For Intel Hybrid CPUs

    Over the past year since launching Intel Alder Lake processors, Intel engineers have made a number of improvements to the Linux kernel for better dealing with the hybrid processor approach mixing P and E cores. While Alder Lake is running great with recent kernels and the P vs. E core selection for tasks on Linux is better than it was at launch, there still are areas for improvement as raised by Intel engineers this week...

    https://www.phoronix.com/news/Linux-...er-Hybrid-Need

  • #2
    Don't care about that smelly, rotten carcass that are x86 CPUs, but I always understood that there are a lot of issues with balancing between little/big cores with Linux in general.

    Hope Intel can help out there, their software skills are great.

    Comment


    • #3
      Michael

      Grammar

      "Also at LPC 2022 this week, Intel engineers Rui Zhang and Yu Chen further raised the Intel hybrid Linux still remaining."

      Maybe this

      "Also at LPC 2022 this week, Intel engineers Rui Zhang and Yu Chen raised the Intel hybrid Linux work still remaining."

      Comment


      • #4
        I'm sure I'm missing something, but as I understand it, why isn't it a matter of establishing tiers of performance, queuing tasks to run on the most efficient cores first, and if the task monopolizes a core for too long, then move it up a tier? Akin to tasks being migrated when too many of their resources are on a different NUMA node.

        Comment


        • #5
          Originally posted by Snaipersky View Post
          I'm sure I'm missing something, but as I understand it, why isn't it a matter of establishing tiers of performance, queuing tasks to run on the most efficient cores first, and if the task monopolizes a core for too long, then move it up a tier? Akin to tasks being migrated when too many of their resources are on a different NUMA node.
          Unfortunately it's not that simple. The strategy you describe would work for non-interactive tasks like a 3D render or compressing a file, but not if it's a task like responding to a network packet, or creating a GPU draw call. For short, interactive, tasks you'd much rather run them right away on the biggest performance cores first, then "race to sleep".

          There have been many Linux conference talks (Linuxconf, LPC, LinuxEmbedded, etc.) going back over a decade dedicated to task scheduling in modern computing. Part of that is deciphering which tasks are interactive vs. not interactive. But steering the hardware also continues to get more difficult since "modern computing" involves multiple cores, which could be in any number of boost and/or sleep states, now with varying energy characteristics, and now potentially different instruction sets.

          Comment


          • #6
            Originally posted by discordian View Post
            Don't care about that smelly, rotten carcass that are x86 CPUs, but I always understood that there are a lot of issues with balancing between little/big cores with Linux in general.

            Hope Intel can help out there, their software skills are great.
            It's a little surprising to see the to do list, considering support for big.LITTLE is supposedly mature.

            Comment

            Working...
            X