Large scale network measurement
A large network as the Internet is often divided into a number of domains (like ISPs) and all domains are independently managed. One ISP cannot access other ISPs networks. However, we need to find the characteristics of the whole network, such as link-level loss rates, link level delay distribution, available bandwidth, network topology. etc. to understand the network behaviours. Using the characteristics, we can further to set up models and study the interactions between different components. The key question here is how to find those characteristics without direct measurement, simply rely on end-to-end measurement. The research at this moment are focused on the following three areas:
a) Loss tomography, which aims to estimate/infer loss rates of each link by end-to-end observation. We have some good results for a tree topology and is going to extend them into a general topology
b) Delay tomography, which is for estimating delay distribution. In this area, we also have some good results by using sequential imputation. The work needs to be extending to general topology.
c) Temporal tomography, which aims to find the temporal/spatial correlations, such as transition matrix, of the links. In this situation, observations are related to each other, as Markov chain. With transition matrix, we will be able to understand the temporal correlation between network states. Hidden Markov model, etc. are under investigation to uncover the temporal information.
d) Spatial tomography, in contrast to temporal tomography, aims to find spatial correlation between links in a network, such as a link has heavy delay, whether this delay will have any impact to its neighbours and by how much.
Topology estimation/identification can use in some bioinformatics techniques, such as inferring motif.