Link delay estimation under undeterministic routing using neural network

Abstract

Network virtualization allows the provision of various network services and enables flexible control by dynamically changing the path according to the service etc. While conventional network tomography uses path information to estimate the internal network status, such as each link delay, dynamic path changes make it difficult to determine the path that a packet will take. For networks with undeterministic routing, this study proposes a method for estimating the status of each link using a neural network that does not require path information as an input. Instead, it estimates the status of each link using only end-to-end measurements. The neural network is trained using various patterns of individual link statuses as teaching signals on a simulated network where the path changes dynamically. We evaluated the effectiveness of our method through simulations. The results show that the proposed method can identify degraded links with a true positive rate of 98% and false positive rate of 8%.