Performance comparison of randomized gossip, broadcast gossip and collection tree protocol for distributed averaging

Gossip and tree-based aggregation algorithms are two popular solutions for distributed averaging in wireless networks. The former uses only local message exchanges and requires no routing structures whereas the latter requires building a spanning tree. In this paper we provide a detailed comparison of their performance in terms of communication overhead, accuracy, latency and energy consumption using the network simulator Castalia. We use randomized gossip, broadcast gossip and the collection tree protocol as practical representatives in each category. Through simulations, we show that broadcast gossip requires, in general, the least communication overhead and lowest latency and energy at the expense of lower accuracy. Randomized gossip requires more transmissions than broadcast gossip, but has higher accuracy. The collection tree protocol requires, in general, the most communication overhead.