Distributed Partial Inference under Churn

Effective network measurement can significantly improve application performance. One of the main challenges in obtaining network measurements is to achieve high accuracy while consuming few network resources. To address this problem, several inference mechanisms have been proposed. These mechanisms can provide the O(n²) end-to-end measurements among n nodes using O(n) measurements, with some loss in accuracy. We construct a measurement request graph where a measurement request issued by an application (e.g., for delay between two nodes) is represented by an edge between the nodes. When the measurement request graph is sparse, an inference mechanism operating over all the nodes (complete inference) incurs unnecessary cost. Given a measurement request graph, our goal is to optimize the number of measurements as well as improve overall accuracy by applying inference only to dense sub-graphs, while taking the other measurements directly. We call this technique partial inference. Previous work only considered static measurement request graphs. However, the measurement request graph can be dynamic when nodes frequently join and leave the network. This paper designs and evaluates a distributed algorithm where each node decides if it should participate in an inference mechanism based on limited information.