(Even using MPLS, you can partly re-use the code. )
- incremental deployment looks hard
- ITR only gets to choose destination ETR, not alternative ways to get there
- packet overhead is relatively large
What I propose is use MPLS instead. The MPLS equivalent of the ITR/ETR functions are widely available in implementations today. MPLS has the potential to have relatively little packet overhead, and it allows for the possibility to specify paths rather than just exit points. Because MPLS uses a label rather than a tunnel endpoint address, it's much easier to decapsulate packets at arbitrary locations rather than at one exactly specified point.
It's fairly easy to deploy MPLS today and put labels on packets when they enter an ISP network, tunnel them through the core and pop the label at a border router and egress the packet from the AS.
Next step would be something along the lines of distributing labels in eBGP between consenting neighboring ASes, so that AS A can tell AS B to put label L on packets towards prefix X, so that B's ingress router doesn't have to do a full table lookup, either. Fairly boring stuff.
But it gets more interesting if we give sources (could be site exit routers or middleboxes, but ideally source hosts) the ability to put a label stack on packets. That would work like this: at some point, a source decides that it's sending so many packets to a destination that it's useful to make these packets go through the core faster, so it sends out a bubble packet. This is an IP packet with as its destination address the destination in question, preceded by an MPLS header with a special "bubble" label value. If the ISP doesn't support this mechanism, the packet is simply dropped and nothing happens. (Remember, this packet is not part of the ongoing communication with the destination.) If the ISP supports the mechanism, it returns to the source a set of label stacks with priorities. The source now encapsulates packets to the destination in one of the label stacks. It monitors return traffic to see if the path works, and switches to a different one if it doesn't. (So this requires state, want this to happen close to the source.) The source can then send a packet that has one of the thusly discovered label stacks followed by a bubble to discover label stacks for paths through the next AS. And so on until it knows a full path or possibly a number of full paths.
This mechanism allows for multihoming and network (even host?) mobility by having a destination declare that it's currently residing at an alternative address. The source then sends a bubble for that address, but when it has a full path, it encapslates packets with the original destination address using the MPLS stack for the path towards the care of address.
The thing that I really like about this is that it allows for a source to find different paths towards the same destination, which gives the host ways to optimize its communication that don't exist today.
--
word 'unsubscribe' in a single line as the message text body.