Probabilistic Network Fault-Diagnosis Using Cross-Layer Observations

Fault diagnosis is vital to initiate correct recovery actions in order to provide reliable end-to-end services in unreliable networking environments. In this paper we investigate end-node driven fault diagnosis assuming that no support functions in the network exist.Fault diagnosis in networking systems spanning wired and wireless links is complicated as faults are hidden in the network and observations are unreliable. To overcome this, we show how Bayesian networks (BNs) can be applied for probabilistic fault diagnosis.We model a TCP end-to-end connection to estimate the state of the network and diagnose faults in the wireless and wired domain respectively. Estimation accuracy and diagnosis reactivity performance is evaluated from simulations and compared to a simple threshold based approach. We show how multiple unreliable cross-layer observations improve BN diagnosis performance and robustness to changes in the environment. Furthermore, we evaluate the BN approach and methods to extract features from network traffic to suggest improvements.