Internet BGP Comparison

BGP is a path-vector protocol, meaning it is based on distance-vector. It calculates a spanning, shortest-path tree from every destination advertised. A DV protocol requires at least Ω (dm) messages to converge, where m is the number of links in a simple graph of d destinations. This is trivial to see as, for every link, at least one of the two nodes must inform the other of each destination. This best case can only happen when there is no more than one path between any two nodes, i.e. where the network is a tree. However, in practice on better connected networks, more messages will be sent for transitory states, as BGP “explores” the additional paths available. In the worst case exploring O ((n − 1)!) states [46]. With the addition of sequence numbers, entire classes of transitory states could be suppressed and it is claimed BGP could then be made to converge in O (diam (G)) time [57].

The Cowen Landmark Routing based routing scheme outlined in this Chapter should be no less practical than BGP is, on a growing Internet. For each of its components scales no worse than BGP. Those components are the landmark selection protocol, the landmark routing protocol, and the local cluster routing protocol. The landmark selection protocol further consists of the

DYNAMIC−k−CORE algorithm as the node ranking algorithm, a distance-vector based spanning tree and a counting protocol.

As shown in Chapter 4, the DYNAMIC−k−CORE algorithm converges in

O (diam (G) ∆ (G)) steps in the worst case, and O|E|2 messages in total from an

initial, fully unconverged state. This a little worse than BGP. However, the observed convergence time of DYNAMIC−k−CORE on Internet AS graphs in simulation scaled no worse than linear with the number of ASes. Further, the convergence time as well as the number of messages sent in response to the addition of a link in an otherwise stable state was shown to be minimal, for the overwhelming majority of cases of randomly chosen links. The performance of DYNAMIC−k−CORE should therefore be acceptable.

The spanning tree algorithm used for the counting protocol is distance-vector based. It therefore can not do worse than BGP, as BGP is also DV based. However, where BGP must converge on shortest-path spanning trees rooted at every destination in the network, the spanning tree need only converge on trees rooted at one node (or a small subset of roots, for redundancy). Roots are chosen by node ranking priority using k-core, which ranks nodes in more densely

inter-connected subgraphs more highly, and this may benefit convergence. The counting protocol, assuming a converged spanning tree, requires twice the diameter of the network steps to converge, at worst, and so is O (diam (G)). I.e., it

scales better than current BGP, and similar to BGP with sequence numbers, in the worst case.

The landmark routing protocol can essentially be BGP as it is today, and so can not have worse properties. However, the landmark set should grow more slowly than the network. So, in time, it will be distributing fewer routes than global BGP would. Which would be an improvement in state, processing and messaging costs. Similarly, the local routing cluster protocol can be BGP based, but its operation is restricted to a much smaller subset of the network than global BGP today

operates over.

The landmark selection protocol’s convergence time likely will be dominated by DYNAMIC−k−CORE, in the worst case. The spanning tree should converge quickly after DYNAMIC−k−CORE does, and the counting protocol soon there-after. In the normal case, the network is stable, and the protocols are converged, and further link addition or removals events should generate few messages. The

DYNAMIC−k−CORE can in a large majority of cases can react to link additions in just a few messages, in which case the spanning tree should remain undisturbed, and the counting protocol may send messages over the tree to the roots and back down to all nodes. The spanning tree may result in the most messages in link events from the stable state, but these need still not be onerous compared to BGP, where a new node joining would result in at least one message being sent over every link.

In summary, the overall routing scheme should be practical, compared to BGP. The landmark selection protocol may converge slightly slower than ideal-case BGP. That convergence time should be dominated by the DYNAMIC−k−CORE

algorithm, which has been observed to be no worse than linear on AS graphs. Further, this brings benefits in allowing other parts of the routing protocol’s

messages to be much more constrained in scope than global BGP, which may offset the landmark selection protocol’s additional communication costs. To what degree would require further investigation, as noted in Section 6.4 on page 157.

Finally, this Cowen Landmark Routing protocol allows the overall routing state to be compact, and grow sub-linearly relative to the growth in the size of the network while still offering good routing connectivity. Internet BGP has not been able to deliver this. This Cowen Landmark Routing protocol allows nodes to have good, shortest-path routing information to those nearby nodes to which such routes are beneficial, without imposing such routes on the routing table of every Internet router globally, as current BGP does.