[ovirt-users] VM resource allocation and IO Threads

Gianluca Cecchi gianluca.cecchi at gmail.com
Thu Nov 2 11:59:04 UTC 2017


On Mon, Oct 30, 2017 at 6:16 PM, Martin Polednik <mpoledni at redhat.com>
wrote:

> As stated previously, 1 IO thread is somewhat sane choice. Multiple IO
> threads make sense if you have multiple storage devices across
> different NUMA nodes.
>
> mpolednik
>
> On Mon, Oct 30, 2017 at 4:29 PM, Darrell Budic <budic at onholyground.com>
> wrote:
> > Best explanation I’ve found is
> > https://wiki.mikejung.biz/KVM_/_Xen#virtio-blk_iothreads_.28
> x-data-plane.29
>
>
[snip]


> >
> > I currently enable it on all VMs and assign 1 thread per drive on my
> > systems.
>
>
Thanks for sharing your experience, Darrell.
I'm testing some VMs with Oracle RDBMS where I configured 3 disks with
virtio-scsi and now I have assigned 3 IO threads. Let's see the numbers.
In the mean time I verified you have to deactivate/activate the disks in
the guest, otherwise you do get 3 SCSI controllers, but all the 3 disks are
bound to the first one (scsi0-0-0-0).
Now instead on hypervisor I correctly get this with dumpxml:

# virsh -r dumpxml my_guest | grep scsi
      <target dev='sda' bus='scsi'/>
      <alias name='scsi0-0-0-0'/>
      <target dev='sdb' bus='scsi'/>
      <alias name='scsi1-0-0-0'/>
      <target dev='sdc' bus='scsi'/>
      <alias name='scsi2-0-0-0'/>
    <controller type='scsi' index='0' model='virtio-scsi'>
      <alias name='scsi0'/>
    <controller type='scsi' index='1' model='virtio-scsi'>
      <alias name='scsi1'/>
    <controller type='scsi' index='2' model='virtio-scsi'>
      <alias name='scsi2'/>
#

Martin, I have not understood when you write about "if you have multiple
storage devices across different NUMA nodes."
Can you elaborate a bit or point to any documentation link?

Thanks,
Gianluca
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.ovirt.org/pipermail/users/attachments/20171102/edaf61d6/attachment.html>


More information about the Users mailing list