[ovs-dev] OVN - L3 Gap between NB schema and Neutron

Amitabha Biswas abiswas at us.ibm.com
Mon Aug 3 07:36:21 UTC 2015


Hi,

I think Gal is saying the following - ml2 allows the following:

VM (10.0.0.2) --- Logical_Switch ---- (10.0.0.2) LogicalRouter
                              |                 |
                              |                 |
VM (10.0.1.2) ----+                +-----(10.0.1.2) LogicalRouter

The Logical Switch (OVN_NB) has a one to one mapping to a Logical Datapath 
(OVN_SB).

An alternate schema could be to map Subnets to a Logical Switch instead of 
mapping a Neutron Network to a Logical Switch. This would preserve the 
requirement of a single router port in the Logical Switch table.

Thanks
Amitabha



From:   Aaron Rosen <aaronorosen at gmail.com>
To:     Ben Pfaff <blp at nicira.com>
Cc:     dev <dev at openvswitch.org>, Eran Gampel 
<Eran.Gampel at toganetworks.com>, Aaron Rosen <arosen at vmware.com>
Date:   07/30/2015 07:27 PM
Subject:        Re: [ovs-dev] OVN - L3 Gap between NB schema and Neutron
Sent by:        "dev" <dev-bounces at openvswitch.org>



Hi Gal,

So you're saying that ml2 allows you to configure a topology like this?


VM (10.0.0.2) ----Logical_Switch----(10.0.0.2)LogicalRouter
                                      |
                                      |
 +------(10.0.0.3)--Logical-Router--WAN


And then the vm would be responsible for having specific routes to each gw
ip?

I think you're right that this will work with the current L3 agent. That
said, I wondering if it's even worth supporting this topology if it's
complex to implement and there are not many use cases for it (or being
requested by users). I haven't heard anyone asking for this before (and 
nvp
doesn't implement this either fwiw). As an alternative to accomplishing 
the
same thing one could use a VM with two ports.

Aaron






On Thu, Jul 30, 2015 at 11:28 AM, Ben Pfaff <blp at nicira.com> wrote:

> [also adding Salvatore]
>
> On Thu, Jul 30, 2015 at 11:27:57AM -0700, Ben Pfaff wrote:
> > If both the router ports point to the same router, then I am not sure
> > why this would need to be two ports.  Maybe the schema is not 
sufficient
> > to report both IPv4 and IPv6 addresses on a single router port; if so,
> > then I would support enhancing the schema to fix that.
> >
> > I suspect that for connecting to two different routers, it is possible
> > to instead connect one router and then connect that router to others 
in
> > a way that accomplishes an equivalent goal.  I haven't thought it
> > through though.
> >
> > On Thu, Jul 30, 2015 at 09:12:14PM +0300, Gal Sagie wrote:
> > > Yes, i checked this on my setup.
> > > For example, you can have both IPv6 and IPv4 subnets per the same
> network
> > > (which maps to a logical switch)
> > > and connect both as two different router ports (to the same router)
> > >
> > > You can also connect the same network to two different routers, i am
> not
> > > sure if you need the extra route extension for that or not, i think 
you
> > > could
> > > configure it as default gateway with out this extension, but with 
the
> > > extension you
> > > can define routing between the two routers.
> > >
> > >
> > >
> > >
> > >
> > > On Thu, Jul 30, 2015 at 9:03 PM, Ben Pfaff <blp at nicira.com> wrote:
> > >
> > > > [adding Aaron Rosen]
> > > >
> > > > On Wed, Jul 29, 2015 at 12:20:30PM +0300, Gal Sagie wrote:
> > > > > Currently Neutron support defining few subnets (IP cidrs) on a
> network
> > > > > (logical switch)
> > > > > and connecting them to the same router (or different routers).
> > > > > Currently in the NB schema, the logical switch can be connected
> only to
> > > > one
> > > > > logical
> > > > > router port.
> > > > >
> > > > > This needs to be extended so a logical switch can have more then
> one
> > > > > logical router
> > > > > port reference to support the above use case.
> > > >
> > > > Limiting a logical switch to a single router port is an 
intentional
> > > > design decision.  It means that a packet traverses at most two
> logical
> > > > switches (one at ingress, one at egress), which simplifies some of
> the
> > > > logical switch design, and it prevents loops.
> > > >
> > > > VMware's NVP controller uses the same design, for those reasons 
and
> > > > others.  The NVP paper from NSDI 2014 (see
> > > > http://benpfaff.org/papers/net-virt.pdf) puts it this way:
> > > >
> > > >     As an optimization, we constrain the logical topology such 
that
> > > >     logical L2 destinations can only be present at its edge[6]. 
This
> > > >     restriction means that the OVS flow table of a sending 
hypervisor
> > > >     needs only to have flows for logical datapaths to which its 
local
> > > >     VMs are attached as well as those of the L3 routers of the
> logical
> > > >     topology; the receiving hypervisor is determined by the 
logical
> IP
> > > >     destination address, leaving the last logical L2 hop to be
> executed
> > > >     at the receiving hypervisor.
> > > >
> > > >     [6] We have found little value in supporting logical routers
> > > >         interconnected through logical switches without tenant 
VMs.
> > > >
> > > > Are you sure that Neutron supports multiple router ports per 
switch?
> > > > Russell Bryant (in IRC) and Aaron Rosen (in a quick in-person 
chat)
> > > > seemed doubtful.
> > > >
> > >
> > >
> > >
> > > --
> > > Best Regards ,
> > >
> > > The G.
> _______________________________________________
> dev mailing list
> dev at openvswitch.org
> http://openvswitch.org/mailman/listinfo/dev
>
_______________________________________________
dev mailing list
dev at openvswitch.org
http://openvswitch.org/mailman/listinfo/dev





More information about the dev mailing list