Scalable Network Tomography System

Abstract: Network Tomography enables network operators to measure end-to-end network metrics from several locations at the same time which are useful to a wide range of applications. In previous studies, network tomography was applied to attain a full mesh with active probing but most of the systems have central control and thus lack scalability. There are also a few which are distributed in nature but do not ensure full mesh coverage. They perform measurements of a subset of paths and try to use inference for remaining. But in case of measuring available bandwidth, inference is not a feasible technique. In contrast, we set out to achieve a full mesh-probing scheme which scales to the nation-wide Internet without using any inference technique. In this thesis report a novel network tomography system that uses probabilistic distributed scheduling for peer selection to probe for measurements is proposed. The system ensures that eventually the solution tends to full mesh probing. Bandwidth measurements congests the network and utilizes more system resource compared to measuring metrics like link loss or latency. This is a constraint which is also taken into account. In pursuit of providing full mesh coverage in a distributed system, each node needs to know about each other, and hence a group membership management protocol is introduced. A simple way to store the probed data in a distributed storage which can be queried in message hop is also presented. The system is designed as a plugin system which allows any existing measurement tool to be plugged in and perform measurements to extract the metrics of the network by active probing.