libvirt users - Nov 2020 - Re: Libvirt driver iothread property for virtio-scsi disks

The docs[1] say:

- The optional iothread attribute assigns the disk to an IOThread as defined by
  the range for the domain iothreads value. Multiple disks may be assigned to
  the same IOThread and are numbered from 1 to the domain iothreads value.
  Available for a disk device target configured to use "virtio" bus
and "pci"
  or "ccw" address types. Since 1.2.8 (QEMU 2.1)

Does it mean that virtio-scsi disks do not use iothreads?

I'm experiencing a horrible performance using nested vms (up to 2 levels of
nesting) when accessing NFS storage running on one of the VMs. The NFS
server is using scsi disk.

My theory is:
- Writing to NFS server is very slow (too much nesting, slow disk)
- Not using iothreads (because we don't use virtio?)
- Guest CPU is blocked by slow I/O

Does this make sense?

[1] https://libvirt.org/formatdomain.html#hard-drives-floppy-disks-cdroms

Nir

On Wed, Nov 04, 2020 at 05:48:40PM +0200, Nir Soffer
wrote:
> The docs[1] say:
> 
> - The optional iothread attribute assigns the disk to an IOThread as
defined by
>   the range for the domain iothreads value. Multiple disks may be assigned
to
>   the same IOThread and are numbered from 1 to the domain iothreads value.
>   Available for a disk device target configured to use "virtio"
bus and "pci"
>   or "ccw" address types. Since 1.2.8 (QEMU 2.1)
> 
> Does it mean that virtio-scsi disks do not use iothreads?
virtio-scsi disks can use iothreads, but they are configured in the
scsi controller, not in the disk itself. All disks attached to the
same controller will share the same iothread, but you can also attach
multiple controllers.
> I'm experiencing a horrible performance using nested vms (up to 2
levels of
> nesting) when accessing NFS storage running on one of the VMs. The NFS
> server is using scsi disk.
> 
> My theory is:
> - Writing to NFS server is very slow (too much nesting, slow disk)
> - Not using iothreads (because we don't use virtio?)
> - Guest CPU is blocked by slow I/O
I would discard the lack of iothreads as the culprit. They do improve
the performance, but without them the performance should be quite
decent anyway. Probably something else is causing the trouble.

I would do a step by step analysis, testing the NFS performance from
outside the VM first, and then elaborating upwards from that.

Sergio.
> Does this make sense?
> 
> [1] https://libvirt.org/formatdomain.html#hard-drives-floppy-disks-cdroms
> 
> Nir
>

On Wed, Nov 04, 2020 at 05:48:40PM +0200, Nir Soffer
wrote:
> The docs[1] say:
> 
> - The optional iothread attribute assigns the disk to an IOThread as
defined by
>   the range for the domain iothreads value. Multiple disks may be assigned
to
>   the same IOThread and are numbered from 1 to the domain iothreads value.
>   Available for a disk device target configured to use "virtio"
bus and "pci"
>   or "ccw" address types. Since 1.2.8 (QEMU 2.1)
> 
> Does it mean that virtio-scsi disks do not use iothreads?
> 
> I'm experiencing a horrible performance using nested vms (up to 2
levels of
> nesting) when accessing NFS storage running on one of the VMs. The NFS
> server is using scsi disk.
When you say  2 levels of nesting do you definitely have KVM enabled at
all levels, or are you ending up using TCG emulation, because the latter
would certainly explain terrible performance.
> 
> My theory is:
> - Writing to NFS server is very slow (too much nesting, slow disk)
> - Not using iothreads (because we don't use virtio?)
> - Guest CPU is blocked by slow I/O
Regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|

On Wed, Nov 4, 2020 at 6:42 PM Sergio Lopez <slp@redhat.com>
wrote:
>
> On Wed, Nov 04, 2020 at 05:48:40PM +0200, Nir Soffer wrote:
> > The docs[1] say:
> >
> > - The optional iothread attribute assigns the disk to an IOThread as
defined by
> >   the range for the domain iothreads value. Multiple disks may be
assigned to
> >   the same IOThread and are numbered from 1 to the domain iothreads
value.
> >   Available for a disk device target configured to use
"virtio" bus and "pci"
> >   or "ccw" address types. Since 1.2.8 (QEMU 2.1)
> >
> > Does it mean that virtio-scsi disks do not use iothreads?
>
> virtio-scsi disks can use iothreads, but they are configured in the
> scsi controller, not in the disk itself. All disks attached to the
> same controller will share the same iothread, but you can also attach
> multiple controllers.
Thanks, I found that we do use this in ovirt:

    <controller type='scsi' index='0'
model='virtio-scsi'>
      <driver iothread='1'/>
      <alias name='ua-6f070142-1dbe-4be3-90c6-1a2274a2f8a0'/>
      <address type='pci' domain='0x0000' bus='0x03'
slot='0x00'
function='0x0'/>
    </controller>

However the VMs in this setup are not created by oVirt, but manually using
libvirt. I'll make sure we configure the controller in the same way.
> > I'm experiencing a horrible performance using nested vms (up to 2
levels of
> > nesting) when accessing NFS storage running on one of the VMs. The NFS
> > server is using scsi disk.
> >
> > My theory is:
> > - Writing to NFS server is very slow (too much nesting, slow disk)
> > - Not using iothreads (because we don't use virtio?)
> > - Guest CPU is blocked by slow I/O
>
> I would discard the lack of iothreads as the culprit. They do improve
> the performance, but without them the performance should be quite
> decent anyway. Probably something else is causing the trouble.
>
> I would do a step by step analysis, testing the NFS performance from
> outside the VM first, and then elaborating upwards from that.
Makes sense, thanks.

On Wed, Nov 4, 2020 at 6:54 PM Daniel P. Berrangé <berrange@redhat.com>
wrote:
>
> On Wed, Nov 04, 2020 at 05:48:40PM +0200, Nir Soffer wrote:
> > The docs[1] say:
> >
> > - The optional iothread attribute assigns the disk to an IOThread as
defined by
> >   the range for the domain iothreads value. Multiple disks may be
assigned to
> >   the same IOThread and are numbered from 1 to the domain iothreads
value.
> >   Available for a disk device target configured to use
"virtio" bus and "pci"
> >   or "ccw" address types. Since 1.2.8 (QEMU 2.1)
> >
> > Does it mean that virtio-scsi disks do not use iothreads?
> >
> > I'm experiencing a horrible performance using nested vms (up to 2
levels of
> > nesting) when accessing NFS storage running on one of the VMs. The NFS
> > server is using scsi disk.
>
> When you say  2 levels of nesting do you definitely have KVM enabled at
> all levels, or are you ending up using TCG emulation, because the latter
> would certainly explain terrible performance.
Good point, I'll check that out, thanks.
> > My theory is:
> > - Writing to NFS server is very slow (too much nesting, slow disk)
> > - Not using iothreads (because we don't use virtio?)
> > - Guest CPU is blocked by slow I/O
>
> Regards,
> Daniel
> --
> |: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange
:|
> |: https://libvirt.org         -o-            https://fstop138.berrange.com
:|
> |: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange
:|
>