The Routing Over Large Clouds Mailing List Archive by date

[(Mike Shand REO2 G/C2 DTN:830-4424) <shand@shand.reo.dec.com>]

> by the way, you don't need _bugs_ in datagram routing algorithms 
> to get loops, you just need transients in distance vector schemes....

Or indeed with most Link State schemes. Unless you go to considerable lengths to
synchronize the state (and the use of that state for forwarding) between routers
then transient loops are inevitable. It is generally a lot easier to allow loops
and make them relatively benign by the use of the ttl, than it is to
guarantee they never occur (and hence be able to do without the ttl). Of course
bridging adopts the oposite tradeoff, but only by severely restricting the use of
alternate paths.

I haven't looked at the Ipsilon stuff in detail, but presumably, provided there is
no loop when the flow is set up then there is no problem. Changing routes subsequenetly 
will make no difference, except that they should eventually be detected and the VCs
torn down and re-established. Presumably the ttl delta is preserved across the ends of
VC anyway.

The case of loops during the VC initialization process is more interesting but
presumably when it attempts to set up the same flow as an existing flow but using
a different ingress interface, then something complains. 

	Mike