[ovs-dev] [PATCH v2 1/2] Check and allocate free qdisc queue id for ports with qos parameters

[Ben Pfaff]

On Mon, Jun 06, 2016 at 02:31:03PM +0530, Babu Shanmugam wrote:
> On Saturday 04 June 2016 01:17 AM, Ben Pfaff wrote:
> >On Wed, May 25, 2016 at 03:35:44PM +0530, bschanmu at redhat.com wrote:
> >I'm a little worried about the way that this will re-set all of the
> >queues on every iteration through the ovn-controller main loop.  That
> >could have some performance drawbacks.  However, maybe it's not worth
> >worrying about it until it's a real problem in practice.
>
> I think setup_qos() does not touch the queues whose parameters are
> consistent with the qos_settings on the logical_port.
> It will modify the queues only when there is a change in the logical_ports.

You're right.

It does still go to the kernel to dump all the queues, which is the
expensive part in the common case.  I guess I won't worry about it until
it's a problem.

> >This patch makes ovn-controller configure queues directly on the egress
> >interface.  That is completely fine.  However, it means that if the
> >egress interface is on an OVS bridge (which is often reasonable,
> >especially if bonding is involved), then ovs-vswitchd will fight with
> >ovn-controller for control of the queues.  This is a long-time issue in
> >ovs-vswitchd, and there has been previous discussion:
> >         http://openvswitch.org/pipermail/discuss/2015-May/017687.html
> >(Currently, I favor the solution proposed there of adding a new QoS
> >type.)
>
> I too feel that new Qos type is a better solution. I could not locate the
> code which handles the new type, is it already available?

None of the solutions presented there has been implemented.

> >I don't think it's a good assumption that all the tunnels share the same
> >egress interface.
>
> I agree to this. So, we now have to set queues on all possible egreess
> interfaces. Is my understanding correct?

Yes.