Announcement

Collapse
No announcement yet.

Intel Posts Reworked Linux Patches To Improve Hybrid CPU + HT/SMT Kernel Behavior

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

  • Intel Posts Reworked Linux Patches To Improve Hybrid CPU + HT/SMT Kernel Behavior

    Phoronix: Intel Posts Reworked Linux Patches To Improve Hybrid CPU + HT/SMT Kernel Behavior

    Back in August an Intel engineer posted a patch series to help Linux on Intel hybrid CPUs with Hypr Threading. That work benefiting newer Alder Lake and Raptor Lake processors is around avoiding unnecessary migrations within SMT domains. Finally an updated version of that patch series has now been posted as this Intel hybrid improvement works its way toward the kernel...

    https://www.phoronix.com/news/Intel-...-Avoid-Migrate

  • #2
    It seems like with SMT, you want to accomplish two things. One task would be to treat sibling threads equally, but the other task is to try and saturate the siblings with work so that SMT is being used. Otherwise, there's no reason to migrate work.

    If I understand this patch, we want to also consider less busy E and P cores when a specific P-Core is filled to the brim with a single thread of work, even when the P-Core has an empty sibling because it's more likely that SMT will hurt performance if we try too hard to saturate both siblings instead of use vacant cores.

    Comment


    • #3
      The E/P cores situation is already fixed, or it is still work in progress??

      Comment


      • #4
        Originally posted by evasb View Post
        The E/P cores situation is already fixed, or it is still work in progress??
        It's WIP.

        Comment


        • #5
          Michael

          Typo/grammar

          In the link "CPUs with Hypr Threading" should probably be "CPUs with Hyper-Threading"

          Comment


          • #6
            Originally posted by Mitch View Post
            It seems like with SMT, you want to accomplish two things. One task would be to treat sibling threads equally, but the other task is to try and saturate the siblings with work so that SMT is being used. Otherwise, there's no reason to migrate work.
            Speaking as one who's never seen the insides of an OS thread scheduler, here's my take on the challenges posed by SMT and hybrid CPUs:
            1. Latency-sensitive or fp-intensive threads should be prioritized for exclusive execution on a P-core. If there are fewer jobs in the run queue than hardware threads available, that should influence who gets "dibs" on the P-cores.
            2. The impact of sharing the core with a sibling thread should be accounted for, by the thread scheduler, so that QoS can be maintained.
            3. Similarly, the execution time on an E-core should be discounted relative to exclusive P-core utilization, to ensure fairness.

            Comment

            Working...
            X