In-Network Analytics for Ubiquitous Sensing

Abstract

We address the problem of in-network analytics for data that is generated by sensors at the edge of the network. Specifically, we consider the problem of summarizing a continuous physical phenomenon, such as temperature or pollution, over a geographic region like a road network. Samples are collected by sensors placed alongside roads as well as in cars driving along them. We divide the region into sectors and find a summary for each sector, so that their union is a continuous function that minimizes some global error function. We designate a node (either virtual or physical) that is responsible for estimating the function in each sector. Each node computes its estimate based on the samples taken in its sector and information from adjacent nodes.

The algorithm works in networks with bounded, yet unknown, latencies. It accommodates the addition and removal of samples and the arrival and departure of nodes, and it converges to a globally optimal solution using only pairwise message exchanges between neighbors. The algorithm relies on a weakly-fair scheduler to implement these pairwise exchanges, and we present an implementation of such a scheduler. Our scheduler, which may be of independent interest, is locally quiescent, meaning that it only sends messages when required by the algorithm. It achieves quiescence on every link where the algorithm ceases to schedule pairwise exchanges; in particular, if the algorithm converges, it globally quiesces.

This research was supported by the Intel Collaborative Research Institute for Computational Intelligence (ICRI-CI), the Israeli Ministry of Trade and Labor Rescue Consortium, the Sidney Goldstein Research Fund, the Israeli Science Foundation, the Arlene & Arnold Goldstein Center at the Technion Autonomous Systems Program, a Technion Fellowship, two Andrew and Erna Finci Viterbi fellowships, the Lady Davis Fellowship Trust, and the Hasso-Plattner Institute for Software Systems Engineering.