[ovirt-users] feedback and questions about SR-IOV feature

[Martin Polednik]

On 06/12/16 15:23 +0100, Nathanaël Blanchet wrote:
>
>
>Le 06/12/2016 à 13:19, Martin Polednik a écrit :
>>On 06/12/16 12:14 +0100, Nathanaël Blanchet wrote:
>>>Hello,
>>>
>>>My new 10G NICS support now SR-IOV, and I've played with this new 
>>>feature as passthrough device, so as to reduce my host CPU 
>>>consumption.
>>>
>>>At the origin, I set up a bond on both 10G PF nics.
>>>
>>>After many configurations, the only way I manage to use a VF into 
>>>a VM, is to get out of the bond one nic.
>>>
>>>So does it mean that it is impossible to run a VM with VF with PF 
>>>attached to a bond?
>>
>>As far as I know, it's not possible to do that. The reason is that the
>>bond normally creates new (logical) interface, what you are doing is
>>assigning "part" of the bond directly to a VM and the driver within VM
>>isn't aware of the bond.
>This is what I supposed, UI should prevent us to create VFfrom when 
>nic is attached to a bond. Pencil icon should'nt appear in this case.

Agreed.

>>
>>>Moreover, something strange happens : during the boot of the VM, 
>>>the passthrough device gets an dhcp IP on the native vlan of the 
>>>bond, and once finally up, the real vlan used by this device is on 
>>>the different predifined vlan. It implies to me to reconfigure the 
>>>network to ping something on the wanted vlan. Really crazy.
>>
>>This could be explained by previous statement: bonding PFs at
>>hypervisor level and then assigning VFs to a VM can most likely cause
>>undefined behavior.
>The issue is the same when the PF is not attached to a bond, so in an 
>expected working situation.

Interesting, could be something regarding mac anti-spoofing. I've
found something for Mellanox cards (you didn't specify which NIC do
you use) - https://community.mellanox.com/docs/DOC-2461 -- could that
be the source of your issues?

>>
>>>Other question is : In which case can it be useful to be able to 
>>>bond 2 VF? UI let us to do so, but it is impossible to add any 
>>>bridge on that virtual bond.
>>
>>At hypervisor level? I believe it doesn't make sense.
>I wonder this because UI allows to do it. The same as above, user 
>shouldn't be allowed to bond two VFs, and not allowed to add virtual 
>network to a VF

Agreed.
(also adding networking people)

>>
>>If you require bond between 2 PFs, you can assign 2 VFs each from
>>different PF to a VM and bond them within the guest.
>>
>>>Comparing to a large number of restrictions (migration and 
>>>others), my opinion is that this feature seems to be very 
>>>difficult to use in production...
>>
>>The use case for SR-IOV is maximum performance at the cost of
>>convenience while still (somewhat) allowing you to scale.
>>
>>>-- 
>>>Nathanaël Blanchet
>>>
>>>Supervision réseau
>>>Pôle Infrastrutures Informatiques
>>>227 avenue Professeur-Jean-Louis-Viala
>>>34193 MONTPELLIER CEDEX 5
>>>Tél. 33 (0)4 67 54 84 55
>>>Fax  33 (0)4 67 54 84 14
>>>blanchet at abes.fr
>>>
>>
>>>_______________________________________________
>>>Users mailing list
>>>Users at ovirt.org
>>>http://lists.ovirt.org/mailman/listinfo/users
>>
>
>-- 
>Nathanaël Blanchet
>
>Supervision réseau
>Pôle Infrastrutures Informatiques
>227 avenue Professeur-Jean-Louis-Viala
>34193 MONTPELLIER CEDEX 5 	
>Tél. 33 (0)4 67 54 84 55
>Fax  33 (0)4 67 54 84 14
>blanchet at abes.fr
>