[Goat] routing

[William Waites]

On Mon, Apr 26, 2004 at 10:17:29AM -0400, Benoit Grégoire wrote:
> >
> > What happens if C is plugged into a hub with some servers that
> > don't participate in the protocol? 1 doesn't work because there's
> > no way to tell the other nodes about those servers "care of" this
> > gateway, 2 doesn't work because C doesn't exist.
> 
> Off course it works, as long as C knows that B and D are gateways to the 
> address space of the mesh.  Neither 1 and nor 2 removes the need for basic 
> network configuration.  However both 1 and 2 assume that the mesh protocol 
> knows something about subnets.

You have to think about the path in both directions. C knowing about B and
D being gateways means that it can /send/ packets /into/ the mesh. But
packets arriving /from/ the mesh are harder: each mesh node must know
that the address of the server is reachable through C -- and there
is no mechanism for doing that.

But I think the mechanism for doing that should not be SrcRR. The reason:
SrcRR is a MAC layer protocol -- it deals with reachability /within/ a
subnet. It doesn't address reachability /between/ subnets. Rather than
changing (breaking) SrcRR to make it do something it wasn't designed 
to do, it makes sense to use protocols like OSPF that are designed for
exactly this.

> I suppose the gateway could have multiple adresses on RoofNet (one for each 
> possible IP in it's subnet(s), but that doesn't solve the multiple gateway 
> problem either.  

And that would also mean that NAT would be required (although in
the slightly better scenario of 1:1 NAT)

> > What address do you give the computers on the LAN?
> 
> Same ones as in the Mesh obviously, otherwise you need NAT.

So it's all one big subnet?

> Covering the whole island with a grid of nodes spaced by 1000 meters on 
> average would require only less than a thousand nodes.  Is scaling to 1000 
> nodes a problem for Dijkstra's algorithm?  

You're pushing the limits in two ways here. First, 1000m
between antennas is close to the reasonable limit for omni
antennae in an urban area -- it is possible under ideal
conditions with optimum antenna placement which I doubt 
will be the case in all but a small percentage of installations.
Second, it'll be far more than 1000 nodes when you take
into account all LANs plugged into the Mesh, but even
at 1000 nodes, the fact that an adjacency change has to
be flooded through the whole network puts this near the
limit of a Dijkstra protocol. There is no network running
OSPF, for example, that I know of that has scaled to 1000
nodes without using areas.

> > This works because SrcRR makes the wireless meshes
> > look reliable to OSPF. I'd like to test this out,
> > perhaps at the next hack night.
> 
> Defining clouds, and thus cloud edges in the MESH may be extremely 
> challenging.  Some nodes at the edges may see very far, while other will have 
> very narrow coverage.  But it's worth discussing. 

I don't think they have to be defined a priori. In the process
of building it the "natural" boundaries will become clear and
will be defined by a combination of topographical factors and
availability of places to put equipment.

One good way to define an edge node is as a node that has
one interface in the Mesh, and one interface outside the
Mesh. There are two types of such edge nodes -- ones that
provide connectivity to stub networks (i.e. a single LAN)
and ones that provide transit to other networks beyond
their local segment.

> Well, it's quite likely that none of the Routing protocols will fill be 100% 
> of the requirements without modifications.  In fact my shoping list is:
> 
> - -the EAX (or similar) metric algorithm
> - -Support for IP subnets
> - -A reactive protocol if the MIT paper's findings against proactive protocols 
> hold.
> 
> and isn't directly met by any protocol that I know of. 

Agreed.

Just remember that RoofNet's SrcRR is not an IP routing 
protocol, it's a MAC layer protocol. It's a good building
block, and I support using it, but it's far from the whole
story.

Cheers,
-w
-- 
The "cutting edge" is getting rather dull.
                -- Andy Purshottam

_______________________________________________
Goat mailing list
Goat at isf.waglo.com
http://isf.waglo.com/mailman/listinfo/goat_isf.waglo.com