<html>
  <head>
    <meta http-equiv="Content-Type" content="text/html; charset=UTF-8">
  </head>
  <body>
    <p>I would suggest to:</p>
    <p>- make sure that "none" I/O scheduler is used inside VM (e.g.
      /sys/block/sda/queue/scheduler). I assume quite recent kernel,
      otherwise "noop".<br>
    </p>
    <p>- make sure that host has CPU C-States above C1 disabled (check
      values of all /sys/devices/system/cpu/*/cpuidle/state*/disable for
      while [..]/name is different than "POLL", C1, C1E) or use some
      tool that disables that.<br>
    </p>
    <p>- Use raw images instead of qcow2: in [libvirt] section of
      nova.conf set force_raw_images=True and images_type=flat and
      recreate the instance<br>
    </p>
    <p>Is the difference so big also when you lower I/O depth (for
      example to 1) or increase block size (for example to 64k) ?<br>
    </p>
    <p><br>
    </p>
    <div class="moz-cite-prefix">On 09/08/2023 10:02, Jan Wasilewski
      wrote:<br>
    </div>
    <blockquote type="cite"
cite="mid:CAN4DDNgkk7qVUN3t8bqr4_pBudJ3MOJxJauc_bcWLV0C1qj5yw@mail.gmail.com">
      <meta http-equiv="content-type" content="text/html; charset=UTF-8">
      <div dir="ltr">
        <div>Hi,</div>
        <p>I am reaching out to inquire about the performance of our
          local storage setup. Currently, I am conducting tests using
          NVMe disks; however, the results appear to be underwhelming.</p>
        <p>In terms of my setup, I have recently incorporated two NVMe
          disks into my compute node. These disks have been configured
          as RAID1 under md127 and subsequently mounted at
          /var/lib/nova/instances [1]. During benchmarking using the fio
          tool within this directory, I am achieving approximately
          160,000 IOPS [2]. This figure serves as a satisfactory
          baseline and reference point for upcoming VM tests.</p>
        <p>As the next phase, I have established a flavor that employs a
          root disk for my virtual machine [3]. Regrettably, the
          resulting performance yields around 18,000 IOPS, which is
          nearly ten times poorer than the compute node results [4].
          While I expected some degradation, a tenfold decrease seems
          excessive. Realistically, I anticipated no more than a twofold
          reduction compared to the compute node's performance. Hence, I
          am led to ask: what should be configured to enhance
          performance?</p>
        <p>I have already experimented with the settings recommended on
          the Ceph page for image properties [5]; however, these changes
          did not yield the desired improvements. In addition, I
          attempted to modify the CPU architecture within the nova.conf
          file, switching to Cascade Lake architecture, yet this
          endeavor also proved ineffective. For your convenience, I have
          included a link to my current dumpxml results [6].</p>
        <p>Your insights and guidance would be greatly appreciated. I am
          confident that there is a solution to this performance
          disparity that I may have overlooked. Thank you in advance for
          your help.</p>
        <div>/Jan Wasilewski<br>
        </div>
        <div><br>
        </div>
        <div><i>References:</i></div>
        <div><i>[1] nvme allocation and raid configuration: <a
              href="https://paste.openstack.org/show/bMMgGqu5I6LWuoQWV7TV/"
              moz-do-not-send="true" class="moz-txt-link-freetext">https://paste.openstack.org/show/bMMgGqu5I6LWuoQWV7TV/</a></i></div>
        <div><i>[2] fio performance inside compute node: <a
              href="https://paste.openstack.org/show/bcMi4zG7QZwuJZX8nyct/"
              moz-do-not-send="true" class="moz-txt-link-freetext">https://paste.openstack.org/show/bcMi4zG7QZwuJZX8nyct/</a></i></div>
        <div><i>[3] Flavor configuration: <a
              href="https://paste.openstack.org/show/b7o9hCKilmJI3qyXsP5u/"
              moz-do-not-send="true" class="moz-txt-link-freetext">https://paste.openstack.org/show/b7o9hCKilmJI3qyXsP5u/</a></i></div>
        <div><i>[4] fio performance inside VM: <a
              href="https://paste.openstack.org/show/bUjqxfU4nEtSFqTlU8oH/"
              moz-do-not-send="true" class="moz-txt-link-freetext">https://paste.openstack.org/show/bUjqxfU4nEtSFqTlU8oH/</a></i></div>
        <div><i>[5] image properties: <a
href="https://docs.ceph.com/en/pacific/rbd/rbd-openstack/#image-properties"
              moz-do-not-send="true" class="moz-txt-link-freetext">https://docs.ceph.com/en/pacific/rbd/rbd-openstack/#image-properties</a></i></div>
        <div><i>[6] dumpxml of vm: <a
              href="https://paste.openstack.org/show/bRECcaSMqa8TlrPp0xrT/"
              moz-do-not-send="true" class="moz-txt-link-freetext">https://paste.openstack.org/show/bRECcaSMqa8TlrPp0xrT/</a></i></div>
      </div>
    </blockquote>
    <pre class="moz-signature" cols="72">-- 
Damian Pietras</pre>
  </body>
</html>