Intelligent Route Control systems are tools that give the oportunity for multihomed ASes to modify the routes according to some performance criteria.

The paper "Avoiding Oscillations due to Intelligent Route Control Systems" written by Gao et al. at INFOCOM'06 shows, like the paper "Can Coexisting Overlays Inadvertenly Step on Each Other" (Keralapura et al., ICNP'05), that IRCs can lead to traffic oscilliations causing performances degradation.

The paper pin-points that the problem comes from two pitfalls:

1. current IRCs ignores that the performances of a path can change when new traffic is added (after the switching to that path).
2. self-synchronisation can occures between two (related or not) IRCs as the measuring periods overlap).

In the paper, they first show how IRCs are working. After, they presents why there is oscilliation and lower performances and proposes a two fold solution to thwart these issues.

They first propose to add the criteria of available end-to-end bandwidth for the path selection. The idea is that performances would probably fall if traffic is added on a saturated link.

The other part of the solution is to add randomization in the IRCs. Indeed, if the routing periods are randomized, the synchorisation can be reduced and the oscilliation may decreases. IRCs are based on routing period. The routing period can be considered as the cyclic phase of the IRCs. For each routing period, the IRC estimate the performances of the paths and modify the routing table to force the traffic to use the "best" path.

They evaluate that the proposed solution can dramatically reduce the oscilliations. The results are obtained by simulation traffic behaviour for 1s routing periods (which is smaller than traditionnal timers in IRCs) and some different randomization techniques (Deterministic Path Switching, Fixes/Adaptative Switching Probability, Random Routing Periods and Hysteresis Routing Period).

Oscilliations is an important problem in IRCs and the paper has the merit to propose a first solution to tackle it. The principle of randomization seems to be promising and feasible. However, while the available bandwith criteria seems a really good idea. I think there is two main problems with the proposed solution. On one hand, techniques to estimate available bandwidth are often network intrusive and thus open the traditionnal question of scalability. Especially, for the best of my knowledge, available bw tools requires collaboration between the source and the destination of the measurements.

On the other hand, avail bw seems to be usefull when one path is saturated but what appends if the paths still have avail bw? Using the path with the higher avail bw seems not to always be the best choice. For instance, if you have a link with 1 Mbps and the other with 2Mbps and that you must send only 56Kbps both links are good from the bandwidth point of view but maybe that best link is not the one with the higher bw but the other. The problem here is how to know the future traffic (it is hard to estimate the bw required for the future traffic).

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