Network Tomography with Partial Topology Knowledge and Dynamic Routing

Abstract

Networks are always evolving to meet the needs of progressing and novel applications. To this end, improved network capacity, latency and security are required. Monitoring is key to achieving these objectives. It is a cornerstone for the uninterrupted network operation and the applications’ QoS management. Network tomography uses a subset of monitoring information, corresponding to partial view of the network state, to estimate wide-sense network performance metrics. In this paper, we present a novel machine learning (ML) formulation for Network Tomography. It is novel in that its features are designed under the assumption that: (i) the existence of certain links of the network is not known (e.g., due to security reasons), (ii) the routing is dynamic (non-deterministic), i.e., for the same origin–destination node pair, a different route may be selected depending on the state of certain links. These assumptions are typically present in modern networks. The formulation can be used to estimate both additive and non-additive performance metrics. We evaluate our proposal using different ML algorithms: neural networks (NNs), Gaussian regression and linear regression with interactions. Our simulations indicate that our ML formulation has better estimation accuracy compared to traditional algebraic or other ML approaches that cannot or do not take into account these two hypotheses. Regarding the accuracy of the examined ML algorithms, the differences mainly depend on the additive or not nature of the network metrics.