Nalanda

Citation: Intanagonwiwat, C., Govindan, R., and Estrin, D. 2000. Directed diffusion: a scalable and robust communication paradigm for sensor networks. In Proceedings of the 6th Annual international Conference on Mobile Computing and Networking (Boston, Massachusetts, United States, August 06 - 11, 2000). MobiCom ‘00. ACM, New York, NY, 56-67. DOI= http://doi.acm.org/10.1145/345910.345920

This paper describes the general concept of directed diffusion for requesting, and retrieving, data from sensor networks. In such a system, users express an interest in certain kinds of data, and this interest then diffuses across the network towards the nodes most capable of responding, and is then relayed back along the same paths, a smaller set of which may be reinforced by the sensor network itself.

Tasks, or interests, are defined as a set of name/value pairs, along with a specification of the region of the network from which data is to be retrieved. Interests are periodically broadcast by the node to which a user connects, called the sink node, initially at a long interval, to probe the network for the existence of nodes which match the interest. Every nodes maintains an interest cache, with gradients on each interest pointing back to the node from which the interest was received and specifying a data rate.

Nodes, starting with the sink, initially broadcast interests to all their neighbours at a relatively low rate. As the sink starts receiving data, it selectively reinforces particualr neighbours, by resending the interest to them with a higher specified data rate. Nodes in turn perform the same process of reinforcement, selecting neighbours based on matches in their data cache to the interest. In addition, neighbour node selection is affected by the delay to respond to an interest: neighbours with lower delays are preferred. As the conditions along paths vary, data rates may also be lowered for negative reinforcement to remove a path from the preferred set.

The approach described here operates based only on local knowledge, which seems an important consideration for networks with power-constrained nodes. The authors suggest, as evaluated more completely in the other reading for today, that intermediate processing of data may yield even greater benefits.