XCP-ng
    • Categories
    • Recent
    • Tags
    • Popular
    • Users
    • Groups
    • Register
    • Login

    Linux HVM big performance drop on Xeon E7450@HP DL 580g5, somewhere between 6.5 and 7.6 - still visible on 8.1

    Scheduled Pinned Locked Moved Compute
    7 Posts 4 Posters 443 Views 2 Watching
    Loading More Posts
    • Oldest to Newest
    • Newest to Oldest
    • Most Votes
    Reply
    • Reply as topic
    Log in to reply
    This topic has been deleted. Only users with topic management privileges can see it.
    • D Offline
      Dark199
      last edited by

      Hello everyone,

      could someone point me to a possible cause - and hopefully possible workarounds - in order to be able to use Xeon E7450 cores - even if only for sandbox/testing purposes.

      The VM's were running perfectly on 6.5, but after upgrade to 7.6 performance dropped significantly - to the point of unusability. The same VMs - when using PV - still run perfectly. Moreover, windows machines (HVM) show no change of performance. Dom0 - also seems unaffected.
      Cpu usage on those Linux HVM guests easily gets very high (100% all cores) - when given simple tasks, but I am unable to tell why.

      I've tried messing with guest I/O schedulers, and disabling ucode loader in Xen, but nothing helped. I've also tried different network hardware.

      I am aware that 'dunnington' cores are now unsupported (as of 8.1), but the problem exist also on 7.6.

      I've tested on both (up to date) Centos 6 and Centos 7.

      Thanks in advance,

      1 Reply Last reply Reply Quote 0
      • olivierlambertO Offline
        olivierlambert Vates πŸͺ Co-Founder CEO
        last edited by

        You could try disabling all spectre-like security mitigations.

        Could also be your 12 years old CPUs and HVM that isn't greatly working on those old things…

        1 Reply Last reply Reply Quote 0
        • D Offline
          Dark199
          last edited by

          Thanks for quick response,

          I've tried 'pti=off spectre_v2=off l1tf=off nospec_store_bypass_disable no_stf_barrier' (Xen/Dom0/guest) but no improvements.

          I'm still curious enough to investigate further πŸ™‚

          1 Reply Last reply Reply Quote 0
          • olivierlambertO Offline
            olivierlambert Vates πŸͺ Co-Founder CEO
            last edited by

            Any idea @r1 ?

            1 Reply Last reply Reply Quote 0
            • R Offline
              r1 XCP-ng Team
              last edited by

              Are you using a common benchmark tool between 6.5 and 7.6, e.g. sysbench?

              Can you share # dstat output from the VMs while the activity? Netdata would also give you good insights.

              Another thought it about Xen scheduler credit vs credit2.

              Lastly it may have to do with Xen clock source for Linux guests.

              H 1 Reply Last reply Reply Quote 0
              • H Offline
                Holger @r1
                last edited by

                @r1
                I have the same issue.
                Any news about this?

                1 Reply Last reply Reply Quote 0
                • olivierlambertO Offline
                  olivierlambert Vates πŸͺ Co-Founder CEO
                  last edited by

                  Please read the post before yours and try to provide more data if you want further help πŸ™‚

                  1 Reply Last reply Reply Quote 0
                  • First post
                    Last post