Ouroboros Frequently Asked Questions

General Remarks

A new model should demonstrate new capabilities instead of just existing ones.

Of course, we understand the sentiment, and we are convinced the Ouroboros prototype will prove to have some exciting capabilities, such as hot-swapping a complete underlying network without breaking an end-to-end application connection. But that is not our primary objective -- yet. As of now, we don't have enough critical components fully implemented to demonstrate such advanced features, let alone evaluate them copared to currrnt technology.

At this point, we are more concerned with building a solid basic understanding of the fundamental principles underlying networks. Finding the minimal conditions that make or break certain network functionalities. The current model (OSI / TCP layering) has little predictive value about what the consequences are going against the 5/7 layers, and the current Internet does not neatly consist of 5 or 7 layers.

The model we are developing should first explain _why_ the networks that we already have actually work. The Ouroboros model has a different conceptual view, different basic building blocks, put together in a conceptually different way. This is a bit orthogonal to the work carried out by engineers building features. Erwin Schrodinger put it as follows: The task is ... not so much to see what no one has yet seen; but to think what nobody has yet thought, about that which everybody sees.

Paper Remarks

Ouroboros claims Unicast and Multicast must be distinct mechanisms. However, these mechanisms interact seamlessly for existing protocols such as IP/ARP and IPv6 (where, in particular, IPv6 packets are transmitted as Ethernet multicast until MAC learning occurs). Isn't this contradictory?

It's not because they are distinct mechanisms, that they can't interact seamlessly. In the case of Ethernet MAC learning, IPv6 doesn't expose this multicast operation to the application. Ethernet itself doesn't even reveal this to IPv6. On the one hand, as far as IPv6 is concerned, Ethernet is always sending that packet in unicast, but before MAC learning, Ethernet is doing it in a very inefficient way (sending to all, and dropping at all but one). This is just a hidden implementation aspect of IPv6/Ethernet. On the other hand, some parts of IPv6 multicast will leverage Ethernet broadcast, but that is a known intentional optimization of which the IPv6 implementation is aware.

The claim which we are making is that there is no single abstraction that unifies multicast with unicast from the point of view of the process that sends the packets. Multicast and unicast are not transparently interchangeable. Taking an arbitrary unicast program and swapping out unicast for multicast or vice-versa will in almost all cases break that program. It should not be a controversial statement: the proof is pretty straightforward: try to run TCP on top of IP multicast: it won't work because its acknowledgement mechanism depends on one-to-one communication.

Giving things with very different semantics (unicast and multicast) a similar syntax, like Ethernet and IP multicast try to do with a "broadcast address", is just a bad idea that causes more problems (and confusion) than it solves. Ouroboros states that a broadcast address is equivalent to a "layer name", and the construction of "multicast groups" is equivalent to adding nodes to a layer and is governed by the layer management primitives, which are common to both unicast and broadcast layers.