Ricerc@Sapienza

The internet is an enormous highly structured, multi-owner network. The absence of central control and the heterogeneity of the adopted technologies pose two major challenges in network management. Commercial and administrative factors discourage different organizations from sharing details of their network topology and performance. This severely limits the availability of up to date global information, hindering the effectiveness of network management.
In contrast, an up to date view of the status of network components (nodes and links) is necessary for security and performance management. This proposal focuses on inference of internal node and link metrics through end-to-end measurements, an approach known as Network Tomography (NT). NT does not require administrative access privileges on the monitored network, as internal network components are monitored indirectly, from the network edges. Moreover, NT is agnostic to network heterogeneity.
However, NT hits the snag of the huge dimension and intractability of the solution space. The inherent variability of traffic scenarios and the failure dynamics pose additional challenges to the problem, making it necessary to take account of information obsolescence.
In this project, we aim at designing a novel approach to network monitoring, where samples (monitoring paths) are selected on the basis of their capability to contribute information to reconstruct the network state. The monitoring system makes decisions on which paths should be probed next, on the basis of their probability to contribute new information, through Bayesian reasoning.
We envision the design of an always-on monitoring system, which progressively determines the most suitable paths to probe, taking account of past observations and their obsolescence. The proposed approach to Bayesian Network Tomography allows to guide decisions in a complex space and to continuously monitor the network in order to localize node failures and link congestion.