The paper "ISP and Egress Path Selection for Multihomed Networks" written by Dhamdhere and Dovrolis and presented at IEEE INFOCOM'06 discusses a solution to select the best ISPs for multihommed stub depending on traffic patterns of the stub.

The idea is simple, first, determine the traffic top X prefixes of the stub. Determine the possible ISPs for the stub and use stochastic techniques to detemine the subset of ISP to subscribe to and how to perform the best egress path selection.

Several techniques have already been proposed and tools names IRC (Intelligent Route Control) are proposed to dynamically route the traffic in multihomed stubs. The paper proposes two devide the problem in two steps.

First, determine the best ISPs among the list of all the possible available ISPs for the stub and subscribe to the subset. The authors propose a brute force heuristic algorithm that aims at finding the subset of ISPs that optimises cost, performances and robustness. To do so, they define three cost metrics: $c_m$ is the monitary cost (depends on the ISP billing), $c_p$ is the path length cost and $c_d$ is the cost for path diversity. The total cost is defined as a weighted sum of the three metrics. To estimate these costs, they propose to use ISPs looking glass as the stub is not yet a custommer of these ISPs and thus could not send probes on the ISPs.

Second, when the ISPs are chosen, it is require to determine which ISP used for which egress prefix. To do so, a simulated annealing algorithm is used. The algorithm gives a mapping solution that minimises the cost.

The technique proposed in the paper could not be used on demand but the authors proposes to recompute the mapping when the traffic changes. During all the paper, it is supposed that a non-congested path exists and that the bandwidth is determined by the link from the stub to ISPs (bandwidth of the core is big overprovisionned).

The paper discusses an important topic for the future Internet and gives some good references to related works. Maybe more work can be done to propose a more light algorithm that could allow dynamic mapping.

This work is related to our IDIPS researches (http://inl.info.ucl.ac.be/idips).