A prediction criterion for selecting popular destinations

Inter-domain traffic engineering is being increasingly integrated in network management systems. Yet, scalability issues of traffic engineering may limit its applicability to large networks due to excessive computational costs and overhead of routing changes. One documented means of reducing these overheads is to focus the routing optimizations on the paths to popular destinations, and thus be able to shift a large volume of traffic with a small number of path switches, rather than shifting the traffic for all the prefixes. However, there is a lack of simple and pragmatic methods for selecting popular destinations. This paper seeks to address this problem by introducing a practical criterion to define a threshold to categorize the popularity of the traffic. This is based on the analysis of the errors of commonly used predictors for traffic tracking. Our results show that by applying this criterion, we can reduce the number of target prefixes to a small fraction of the total number, while ensuring stability in the traffic engineering, which results from an effective predictability of the traffic headed to these prefixes.