153 Jordanian Journal of Computers and Information Technology (JJCIT), Vol. 2, No. 3, December 2016. GEOGRAPHIC GREEDY TRIPLEWISE GOSSIP ALGORITHM FOR WIRELESS SENSOR NETWORKS Maha I. Raheem Department of Computer Engineering, Baghdad University, Baghdad, Iraq maha.i.raheem@ieee.org (Received: 18-Jan.-2016, Revised: 01-Apr.-2016, Accepted: 17-Apr.-2016) ABSTRACT A novel gossip algorithm for distributed averaging with fast convergence and reduced cost of communication over wireless sensor networks (WSNs) is proposed in this paper. This algorithm is proved to improve the behaviour of the standard gossip algorithm (SGA), triplewise gossip algorithms (TGAs) and the geographic gossip algorithm (GGA) by exploiting the geographic information of the network. An analysis of convergence time and cost of communication of the proposed algorithm is performed and a comparison with other existing methods is provided. KEYWORDS Wireless Sensor Networks, Distributed Computing, Distributed Processing, Gossip Algorithms, Routing. 1. INTRODUCTION Agreement/consensus of sensed information is one of issues of distributed signal processing in WSNs. Averaging the initial value of all the nodes in the network is an example of aggregate problem. Distributed averaging methods are widely used to solve agreement problems [1]. Gossip algorithms are widely used in distributed signal processing. Centralized computing, on the other hand, involves collecting data from all network nodes. In centralized computing, computations are performed at a fusion center. Distributed networks consume more power than their centralized counterparts do; the energy consumption depends on the number of radio transmissions and the total number of iterations until convergence. Distributed averaging algorithms have to be designed to avoid unnecessary waste of power and time. Among the advantages gained, gossip algorithms are robust against link failures and a communication bottleneck near the fusion center is avoided [2]. Sums and averages constitute building blocks for many signal processing applications, such as Gram-Schmidt orthogonalization [2]-[3], WSN node localization [4] and linear parameter estimation [5], to name just a few. Gossiping is a modified version of flooding, where the nodes do not broadcast a packet, but send it to a fully or not fully randomly selected neighbour/s. Gossiping avoids the problem of implosion of the network due to collision, but it takes a long time for message propagation throughout the network [1]. Though gossiping has considerably lower overhead than flooding, it does not guarantee that all nodes of the network will receive the message. It relies on the random neighbour selection to eventually propagate the message throughout the network. Gossip algorithms are employed to calculate the average of measurements of a WSN. In a typical pairwise gossip algorithm such as SGA [6]-[7], one node i wakes up at each iteration with probability P=1/N, where N is the number of sensor nodes, and performs averaging with one of its neighbors j at random with probability Pij; iterations continue with slow convergence [1], [5]- [8]. SGA has another disadvantage in that it wastes a lot of energy among all gossip algorithms because of significant recalculation of redundant information. This result motivated Dimakis et