The Routing Over Large Clouds Mailing List Archive by date

[Dimitry Haskin <dhaskin@baynetworks.com>]

Matthew,

> > 
> > One example:
> > 
> > Let's assume a forwarder A wants to build a single shortcut to a default
> > router which mainains a complete routing table and can act as a traffic
> > de-multiplexer for A.  The domain topology may look like that:
> > 
> >                      E
> >                      |
> >       A -- B -- C -- D -- F
> >                      |
> >                      G
> > 
> > where D is a proxy router for E, G, and F exit forwarders.  D injects a default
> > into IGP.  To resolve the shortcat destination, A sends request for the default 
> > route (i.e. 0.0.0.0/0) along ABCD path.  When request reaches D, D would be obliged
> > to reply to A since it has more specific routes than the requested default route. D
> > specifies itself as the shortcut endpoint.
> > 
> > Dimitry
> 
> 	I'm assuming that all the routers in the above diagram are on the 
> same NBMA. Is this correct ?

Correct.

> 	If so, then the behaviour suggested isn't quite what I expected
> but can be made to happen. I'm interested in a fairly simple multi-area
> OSPF arrangement of routers as the "underlay" to NHRP. A diagram can be
> made available if you like, but for now I'll try and describe my scenario:
> 	Network is broken up into OSPF areas of approx. 50 routers each. 
> Two ABRs are allocated for each _pair_ of areas. Each ABR links to area 0
> and to _both_ these areas, thus providing redundant connection between
> each area and the backbone.
> 	Each area is in effect a smaller "large cloud"  and so in the
> above diagram routers B and C don't exist.
> 	The ABRs perform address aggregation and consequently other
> routers do not know what size LANs are attached to routers in other
> networks. Just sending NHRP requests for a single host address is liable
> to be inefficient. On the other hand, establishing a "short-cut"  to the
> ABR is not normally useful since NHRP has then done nothing. 
> 
> 	From the current spec. (Yakov's email of 20th Oct.) I'm under the 
> impression that if A sends a Request for a Target covering (say) E,F and 
> G then D should in effect return an error:
> 	"If the Second NHRP target constraint is violated then the router
> [D in this case] sends back an NHRP Reply and terminateas the query. The
> Reply should indicate that the second NHRP target constraint was violated.
> The Reply contains IP and NBMA addresses of the router." 
> 	The upshot of this is that the "error" reply lets you use D as a
> forwarder for E,F and G if you really want.
> 	However, please could you give some examples where this is
> desirable (I'm under the impression that it's advisable to avoid sending
> data via D since sending traffic between D and the NBMA network will often
> cost money ). 
>

If you look at the current practices in legacy networks you may find a lot
of devices that send all their traffic to a router even if a more
direct path is available.  I saw it used on a large scale when I supported
a huge, goverment run X.25 based network where a good number of peripheral
routers and hosts sent their traffic to a few "core" routers.  The reason
it is used exactly the same that you think it should be avoided - economical.
In some cases it is more economical to default traffic to a few high-end
routers than to attempt to maintain necessary states and routing
information at peripheral devices.

As to your case I think you'd use R2R in the following way:

 When A wants to send trafic to a destination for which it doesn't have
 a specific route, it sends a Request with a mask of all ones (or prefix
 length of 32 as Yakov suggested). If this desination is reached via an exit
 router E, E replies with the prefix length of the route to the desination.
 Now A may use the shortcut to E to send traffic to all destinations covered
 by the route that E provided.  Desitnations through F and G are resolved
 in the same way.

I hope it helps.
 
> 	Matthew.
> 
> 

Dimitry