Topology information plays an important role in network management. The existing methods for topology inference based on end-to-end measurements need a threshold for general topologies, which is difficult to select to ensure the inference accuracy. In this paper, we propose a sort-based approach, named SBA, to infer the general topologies without using a threshold. First, a sort-based clustering algorithm, named SBC-AL, is proposed to cluster a group of nodes in which every node has at least one sibling. In the SBA, the nodes are classified into disjoint groups by a fan-out decrement mechanism. Then the SBA uses the SBC-AL to cluster the nodes group by group from the bottom up to infer the topology. We prove that the SBA is consistent and suitable for general topologies. The simulation results show that the SBA has a good performance in both accuracy and efficiency.
I. INTRODUCTIONTopology information plays an important role in the resource management, routing update and architecture design of the network. When the topology is unknown, tools such as traceroute are often used to discovery it by collecting the responses from the network elements such as routers. However, as the network grows fast in size and diversity, the network elements may not corporate with such tools. Therefore the network tomography [1] has been proposed to infer the network topology without any corporation of internal network elements. It is used to infer the internal performance characteristics [10] [11] and discovery the network topology [3] by multicast probe packets from the source node to the receivers in the tree-structured network. The basic principle underlying network tomography is to extract the hidden information from active end-to-end measurement data by statistic inference. The observations of the receivers such as packet losses and delays that share the same parent or ancestor have relatively strong correlation, because packets have similar experience on the shared portion of the probe path. In network tomography, the topology discovery is formulated as the problem of hierarchical clustering of the receivers based on the correlation of the received probe packets.