MVSink: Incrementally Building In-Network Aggregation Trees

In-network data aggregation is widely recognized as an acceptable means to reduce the amount of transmitted data without adversely affecting the quality of the results. To date, most aggregation protocols assume that data from localized regions is correlated, thus they tend to identify aggregation points within these regions. Our work, instead, targets systems where the data sources are largely independent, and over time, the sink requests different combinations of data sources. The combinations are essentially aggregation functions. This problem is significantly different from the localized one because the functions are initially known only by the sink, and the data sources to be combined may be located in any part of the network, not necessarily near one another. This paper describes MVSink, a protocol that lowers the network cost by incrementally pushing the aggregation function as close to the sources as possible, aggregating early the raw data. Our results show between 20% and 30% savings over a simplistic approach in large networks, and demonstrate that a data request needs to be active only for a reasonably short period of time to overcome the cost of identifying the aggregation tree.