DATA AGGREGATION IN WIRELESS SENSOR NETWORKS: A COMPARISON OF COLLECTION TREE PROTOCOLS AND GOSSIP ALGORITHMS

Decentralized data aggregation is a canonical task in wireless sensor networks (WSNs). Nodes are independently gathering measurements and the goal is to fuse this data into a unified aggregate. In this paper we compare the performance of the Collection Tree Protocol (CTP) with that of two different gossip algorithms, pairwise randomized gossip and broadcast gossip. We measure performance in terms of the number of transmissions required to compute and disseminate the average to all nodes in the network (i.e., distributed averaging). CTP aggregates and disseminates information along a spanning tree; it thus is very efficient for aggregation, but establishing and maintaining the spanning tree in a decentralized manner involves non-negligible overhead. Gossip algorithms are fully decentralized and only use peer-to-peer communications (i.e., no routing); consequently, they involve little overhead for setup and maintenance, but the actual aggregate computation is slower to converge. Our simulations show that broadcast gossip requires significantly fewer transmissions than CTP in networks with more than 100 nodes when network connectivity is dynamic or unrealiable, and CTP and broadcast gossip offer comparable performance in smaller networks.