On 14 May 2018 at 16:25, Christopher Cox <c...@endlessnow.com> wrote:

> In the ideal case, what you'd have:
>
>        |---- Single virtio virtual interface
>        |
>  VM ---- Host ==== Switch stack
>                  |
>                  |------- 4x 1Gbit interfaces bonded over LACP
>
> The change: virtio instead of "1 Gbit"
>

It's using virtio. My confusion came from the VIF's speed being reported by
the Engine & within the guest.


> You can't get blood from a stone, that is, you can't manufacture bandwidth
> that isn't there.  If you need more than gigabit speed, you need something
> like 10Gbit.  Realize that usually, we're talking about a system created to
> run more than one VM.  If just one, you'll do better with dedicated
> hardware.  If more than one VM, then there sharing going on, though you
> might be able to use QoS (either in oVirt or outside). Even so, if just one
> VM on 10Gbit, you won't necessarily get full 10Gbit out of virtio.  But at
> the same  time bonding should help in the case of multiple VMs.
>
> Now, back to the suggestion at hand.  Multiple virtual NICs.  If the
> logical networks presented via oVirt are such that each (however many)
> logical network has it's own "pipe", then defining a vNIC on each of those
> networks gets you the same sort of "gain" with respect to bonding.  That
> is, no magic bandwidth increase for a particular connection, but more pipes
> available for multiple connections (essentially what you'd expect).
>
> Obviously up to you how you want to do this.  I think you might do better
> to consider a better underlying infrastructure to oVirt rather than trying
> to bond vNICs.  Pretty sure I'm right about that.  Would think the idea of
> bonding at the VM level might be best for simulating something rather than
> something you do because it's right/best.
>

Oh, I'm certain you're right about that! My current budget's focused on
beefing up the resilience our storage layer, however the network is next on
my list. For the moment though, it's a case of working with what I've got.

Bandwidth has only really become an issue recently, since we've started
streaming live events, and that's a simple case of (relatively) low
bandwidth connections. The only place that might get much benefit from
single 10Gbit links would be on our distributed storage layer, although
with 10 nodes, each with 4x1Gbit LAGGs, even that's holding up quite well.

Let's see how the tests go tomorrow...

On 05/14/2018 03:03 PM, Doug Ingham wrote:
>
>> On 14 May 2018 at 15:35, Juan Pablo <pablo.localh...@gmail.com <mailto:
>> pablo.localh...@gmail.com>> wrote:
>>
>>     so you have lacp on your host, and you want lacp also on your vm...
>>     somehow doesn't sounds correct.
>>     there are several lacp modes. which one are you using on the host?
>>
>>
>>   Correct!
>>
>>       |---- Single 1Gbit virtual interface
>>       |
>> VM ---- Host ==== Switch stack
>>                 |
>>                 |------- 4x 1Gbit interfaces bonded over LACP
>>
>> The traffic for all of the VMs is distributed across the host's 4 bonded
>> links, however each VM is limited to the 1Gbit of its own virtual
>> interface. In the case of my proxy, all web traffic is routed through it,
>> so its single Gbit interface has become a bottleneck.
>>
>> To increase the total bandwidth available to my VM, I presume I will need
>> to add multiple Gbit VIFs & bridge them with a bonding mode.
>> Balance-alb (mode 6) is one option, however I'd prefer to use LACP (mode
>> 4) if possible.
>>
>>
>>     2018-05-14 16:20 GMT-03:00 Doug Ingham:
>>
>>         On 14 May 2018 at 15:03, Vinícius Ferrão wrote:
>>
>>             You should use better hashing algorithms for LACP.
>>
>>             Take a look at this explanation:
>>             https://www.ibm.com/developerworks/community/blogs/
>> storageneers/entry/Enhancing_IP_Network_Performance_with_LACP?lang=en
>>             <https://www.ibm.com/developerworks/community/blogs/
>> storageneers/entry/Enhancing_IP_Network_Performance_with_LACP?lang=en>
>>
>>             In general only L2 hashing is made, you can achieve better
>>             throughput with L3 and multiple IPs, or with L4 (ports).
>>
>>             Your switch should support those features too, if you’re
>>             using one.
>>
>>             V.
>>
>>
>>         The problem isn't the LACP connection between the host & the
>>         switch, but setting up LACP between the VM & the host. For
>>         reasons of stability, my 4.1 cluster's switch type is currently
>>         "Linux Bridge", not "OVS". Ergo my question, is LACP on the VM
>>         possible with that, or will I have to use ALB?
>>
>>         Regards,
>>           Doug
>>
>>
>>
>>             On 14 May 2018, at 15:16, Doug Ingham wrote:
>>
>>             Hi All,
>>               My hosts have all of their interfaces bonded via LACP to
>>             maximise throughput, however the VMs are still limited to
>>             Gbit virtual interfaces. Is there a way to configure my VMs
>>             to take full advantage of the bonded physical interfaces?
>>
>>             One way might be adding several VIFs to each VM & using ALB
>>             bonding, however I'd rather use LACP if possible...
>>
>>             Cheers,
>>             --
>>             Doug
>>
>>
>>         --         Doug
>>
>>
>>
>>
>> --
>> Doug
>>
>>
>> _______________________________________________
>> Users mailing list -- users@ovirt.org
>> To unsubscribe send an email to users-le...@ovirt.org
>>
>> _______________________________________________
> Users mailing list -- users@ovirt.org
> To unsubscribe send an email to users-le...@ovirt.org
> oVirt Code of Conduct: https://www.ovirt.org/communit
> y/about/community-guidelines/
> List Archives:
>



-- 
Doug
_______________________________________________
Users mailing list -- users@ovirt.org
To unsubscribe send an email to users-le...@ovirt.org

Reply via email to