Announcement

Collapse
No announcement yet.

Linux 5.4 Kernel To Bring Improved Load Balancing On AMD EPYC Servers

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

  • Linux 5.4 Kernel To Bring Improved Load Balancing On AMD EPYC Servers

    Phoronix: Linux 5.4 Kernel To Bring Improved Load Balancing On AMD EPYC Servers

    Adding to the growing list of features for Linux 5.4 with its cycle officially kicking off in mid-September is a kernel scheduler optimization designed to improve load balancing on AMD EPYC servers...

    http://www.phoronix.com/scan.php?pag...e-EPYC-Balance

  • #2
    I'm sure this will have the greatest impact on heavy multitasking, rather than highly parallel tasks.

    Comment


    • #3
      Originally posted by schmidtbag View Post
      I'm sure this will have the greatest impact on heavy multitasking, rather than highly parallel tasks.
      How so? If you are choking a NUMA node instead of balancing the load your performance suffers even if the task is parallelized.

      Comment


      • #4
        Originally posted by schmidtbag View Post
        I'm sure this will have the greatest impact on heavy multitasking, rather than highly parallel tasks.
        it looks like it'll be a win for anything that forks and dies quickly

        Comment


        • #5
          Originally posted by starshipeleven View Post
          How so? If you are choking a NUMA node instead of balancing the load your performance suffers even if the task is parallelized.
          I said highly parallel, implying the application is going to be using all NUMA nodes.

          Comment


          • #6
            Originally posted by schmidtbag View Post
            I said highly parallel, implying the application is going to be using all NUMA nodes.
            How do you know you are using all NUMA nodes?
            How you know that all NUMA nodes have an equal share of the work?

            If the scheduler keeps a bunch of threads stuck on some node because it thinks that moving it to another node would be bad you're still hurting performance even if you spam threads with a highly parallel load.

            That's the example given in the very commit message.

            Comment

            Working...
            X