[Openstack-operators] VXLAN and jumbo frames: performance gains and configuration options

Elena Ezhova eezhova at mirantis.com
Tue Dec 15 08:01:35 UTC 2015


Hi!

Increasing MTU on interfaces would allow to make use of jumbo frames which
should improve the performance. Here [1] you can find what configuration
options you have to change to use jumbo frames.

Hope this helps.

Thanks,
Elena

[1]
https://chruz.wordpress.com/2015/10/09/configuring-openstack-to-use-jumbo-frames-mtu-9000/

On Tue, Dec 15, 2015 at 12:34 AM, Gustavo Randich <gustavo.randich at gmail.com
> wrote:

> Hi,
>
> Scenario: Kilo / VXLAN / DVR
>
> Networking installation documentation mentions that when using tunneling
> protocols, ideally, we could use jumbo frames on the physical network, but
> then describes a "conservative" approach which is to decrease the MTU size
> of the guests' interfaces (i.e. to 1454), using DHCP.
>
> In my current installation, which is the 'conservative' one (guest
> MTU=1454, host MTU=1500) , I'm getting these approximate iperf rates:
>
>    tenant network, guest to guest same host = 9.5 GB/sec
>    tenant network, guest to guest different host = 1.4 GB/sec
>
>
> My questions are:
>
>    - Should there be any performance gains if I opt for guest MTU = 1500,
> and physical/host MTU = 1550 / 1600 / 9000?
>
>    - Which Neutron / ML2 plugin configuration parameters should I modify
> (in addition to changing physical interfaces' MTU), in order to specify an
> MTU greater than 1500 in the physical network that contains my tenant
> networks?
>
>    - Anyone knows of a Neutron installation procedure or guide covering
> the jumbo frames approach?
>
>
> Thanks!
>
>
> _______________________________________________
> OpenStack-operators mailing list
> OpenStack-operators at lists.openstack.org
> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-operators
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-operators/attachments/20151215/c245387b/attachment.html>


More information about the OpenStack-operators mailing list