Abstract-Achieving cost-effective systems for network performance monitoring has been the subject of many research works over the last few years. Most of them adopt a two-step approach. The first step assigns optimal locations to monitoring devices, whereas the second step selects a minimal set of paths to be monitored. However, such an approach does not consider the trade-off between the optimization objectives of each step, and hence may lead to sub-optimal usage of network resources and biased measurements. In this paper, we propose to evaluate and reduce this tradeoff. Toward this end, we come up with two ILP formulations for a novel monitoring cost model. The aim is to jointly minimize the monitor location cost and the anomaly detection cost, thereby obtaining a monitoring solution that minimizes the total monitoring cost. Our formulations apply for both active and passive monitoring architectures. We show that the problem is NP-hard by mapping it to the uncapacitated facility location problem. Simulation results illustrate the interplay between the optimization objectives and evaluate the quality of the obtained monitoring solution.

No abstract

Abstract-To reduce monitoring cost, the number of monitors to be deployed have to be minimized and the overhead of monitoring flows on the underlying network have to be reduced. In a recent work, we demonstrated, using ILP formulations, that there is a trade-off between theses two minimization objectives. However, we have shown that the trade-off could be efficiently balanced by jointly optimizing monitor location and anomaly detection costs. The problem is NP-complete, hence ILPs could not deliver solutions for large networks. In this paper, we address the scalability issues. We propose two greedy algorithms that optimize monitor location cost and anomaly detection cost jointly. The first algorithm is based on an exhaustive heuristic that explores all paths that are candidate to be monitored, in order to select a subset of paths that reduces the total monitoring cost. On the opposite, the second algorithm is based on a selective heuristic that avoids exploring all the candidate paths to further improve scalability. The main challenge of this heuristic is to not degrade the solution quality. The two algorithms have been evaluated through extensive simulations on networks of hundred of billions of paths. The comparison of the solutions delivered by the two algorithms to each other and to the solutions delivered by the ILP demonstrates that the selective algorithm provides near-optimal solutions, while achieving a desirable scalability with respect to the network size and significant reduction of the computation time.