[WLANware] GSoC 2010: Project IP/UDP encapsulation Kernel module

[Mitar]

Hi!

On Tue, Apr 6, 2010 at 11:28 PM, ZioPRoTo (Saverio Proto)
<zioproto at gmail.com> wrote:
> So instead of sticking people to a gateway inside the mesh, just make
> all the traffic to Internet be tunneled to a fast server and then from
> there all the traffic exits with the same IP address.
> Users see more bandwidth (especially in upload) and they exploit for
> the better an eventual route flap :)

This is exactly what we are doing in wlan ljubljana. We have a lot of
fiber connections and we make use of them a lot. Currently we are
using OpenVPN to connect to central server but this have big
drawbacks: our connections are mostly CPU limited and central server
is what is then single point of failure.

We were discussing this idea and we really started to like it. I think
this is something wonderful. But as we have quite a lot of experience
with practical deployment of such network combined with WiFi and OLSR
I would like to propose an enhancement to original idea. It makes it a
little bit more complicated but what you get is something great.

Our proposal is to make on a node one virtual network device you
configure (with ioctl for example) with unlimited number of other
nodes (ip:port pairs) and kernel module makes a switch to all those
nodes. So it still does IPinUDP but we generalize the idea to multiple
nodes on other side. And we need bridge logic to not duplicate traffic
too much (only for broadcast data this would be the case then, like
OLSR traffic).

So on one extreme you could add all known other nodes to the list of
other sides to virtual network device on every node. Some of those
connections would work, some of them would not, but this would then
OLSR on top of all this recognize and set routes properly.

Of course this extreme would not be useful for big meshes but having
this possibility some higher logic (from userspace) could take a
subset of all nodes and just those configure on network device. But
having this generalization would be really useful.

So this would make topology completely decentralized and also traffic
which would go between nodes could go directly and not over the
central server.

In this way our networks would really become link-level independent.
Whatever we use, all would work.


Mitar