Network Topology Inference Using Higher-Order Statistical Characteristics of End-to-End Measured Delays

Abstract

Network topology is important information for many network control and management applications. Network tomography infers network topology from end-to-end measured packet delays or losses, which is more feasible than internal cooperation-based methods and attracts many studies. Most of the existing methods for network topology inference usually function under the assumption that the distribution of packet delay or loss follows a given distribution (e.g., Gaussian or Gaussian mixture), and they estimate network topology from the parameters of the given distribution. However, these methods may fail to obtain an accurate estimation because the real distribution of packet delay or loss usually cannot be described by a certain distribution. In this paper, we present a novel network topology inference method based on the unicast end-to-end measured delays. The method abandons the assumption of packet delay distribution and constructs network topology by inferring the higher-order cumulants of internal links from the end-to-end measured delays. The analytical and simulation results show that the proposed method offers over 10% improvement in accuracy compared with that of the state-of-the-art works. INDEX TERMS Higher-order cumulant, topology inference, end-to-end measurement, network tomography.

Network Topology Inference Based on Subset Structure Fusion

“…Malekzadeh et al [27] proposed a new probe scheme named traceroute with sandwich probe based on end-to-end unicast delay measurement, which combined the delay-based "sandwich" packet unicast probe model and traceroute. Fei et al [28] use the delay cumulants of second-order and above to infer the topology, so the statistical information of the path delays can be more fully utilized. Although existing network topology tomography methods is relatively mature, it is suitable for small networks.…”

“…We define the tree accuracy as the ratio of the simulation number that the tree is correctly estimated to the total simulation number same as we did in literature [28]. The tree accuracy can intuitively reflect the precision of topology inference.…”