Scalable, self-healing, and self-optimizing routing overlays

After Internet routing was shown in a number of classic measurement papers to result in paths that are sub-optimal with respect to a number of metrics, routing overlays were proposed as a method for improving performance, without the need to re-engineer the underlying network. In this paper, we present SMART, a self-healing, self-optimizing and highly scalable routing overlay, which has a number of advantages with respect to existing solutions. First, SMART can run with off-the-shelf applications and does not require any kernel modification. In addition, SMART can be widely deployed over a sizable population of routers, because it can quickly learn and efficiently track the optimal path with a limited monitoring effort. We describe the design objectives, the architecture and the implementation of SMART, as well as the online decision methods used for learning the optimal routes. Experimental results demonstrate significant improvements over native IP routing, both in terms of latency and throughput.