Resilient Overlay Networks

Summary:

Resilient Overlay Network (RON) is an application-layer overlay on top of internet routing allowing an improved detection and recovery of path outages (or serious performance degradation) for distributed internet applications.

The main goal of RON is to protect applications running on a group of cooperating nodes against the failures of underlying internet paths. BGP-4 has been optimized for scalability and has a poor performance with respect to this criteria. The reason that RON is resilient against fault much more than BGP-4 scheme lies in the fact that RON does not have to consider scalability (2-50 number of nodes).

RON achieves this goal by aggressively probing and monitoring the paths among RON nodes. It detects when the default internet path is the best path and uses it, otherwise data is routed through an intermediate node. RON nodes continuously monitor, measure and share the path properties. An interesting aspect of this architecture is a distributed performance data base in which supports summarization and efficient data exchange.

Interestingly, RON control packets are sent along the RON data path in order to ensure resilience against the mentioned possible internet path failures.Therefor RON routing protocol is itself a client of RON.

Latency, packet loss and throughput are the three default maintained information about each virtual link within RON while clients can overwrite a different metric and influence the routing with their parameters of interest. This is an important characteristic of RON architecture which enable an improved level of integration with applications.

In order to demonstrate these concepts, two distinct experiments are done over a 12 node and 16 node RON structures, demonstrating fault avoidance rate of 60-100% with 18 seconds average recovery time.

Critique:

RON is an interesting approach in grabbing the internet resources to build a geographically distributed small sized resilient network (up to 50). (As indicated by authors) I wonder what could happen if the density of RON networks is increased to a non-negligible portion of the internet? Will the overhead or special path optimizations done by RONs start to introduce new behaviors in the internet not seen before?

Another interesting point of discussion can be the scalability issue. There seems to be a range of applications in which the number of involved nodes is beyond 50 (maybe order of 100s) but not high enough to justify a dedicated network. Is there any hope for such applications?

I enjoyed this paper but will vote for dropping it just because I felt it was too lengthy.