Directed diffusion: A scalable and robust communication paradigm for sensor networks
Chalermek Intanagonwiwat, Ramesh Govindan and Deborah Estrin
In Proceedings of the Sixth Annual International Conference on Mobile Computing and Networking (MobiCOM '00)

• The paper [pdf]
• The presentation [ppt]

Summary:

Directed diffusion is a new communication paradigm intended for sensor network applications. As such, it aims to optimize robustness, scalability, and energy efficiency. It abandons traditional end-to-end philosophies in favour of a localized, hop-to-hop methodology, in which information propagation is determined by localized interactions. Such a structure reduces the amount of state maintained by each node and improves scalability.

Communication in directed diffusion is data-centric: interests, which specify events that are to be recorded, are propagated by sink nodes throughout the network via flooding. These initial requests for information stipulate that responses should be sent at a low data rate. Upon receiving responses from multiple sources, a sink can use positive and negative reinforcement to increase and decrease gradients’ (i.e. paths’) data rates, respectively. Reinforcement mechanisms can vary; this paper uses interest messages with modified interval values for this purpose. Rules governing the propagation of control messages are executed locally—that is, each node decides individually, without global state, the policies determining how it will react to control messages.

The adoption of a hop-to-hop paradigm obviates the need for complex routers and universally unique identifiers. As well, it can result it highly adaptive networks, in which broken links can be avoided automatically according to local policies. In ns2 experiments, the authors found that directed diffusion dissipated less energy than both flooding (a watermark) and omniscient multicast (a representation of attainable performance with traditional end-to-end architectures). As well, it incurred comparable delay to omniscient multicast, and substantially out-performed flooding. Two particularly important performance-enhancing factors of directed diffusion are negative reinforcement, which allows for the pruning of superfluous gradients, and duplicate suppression, which takes advantage of the incorporation of application layer semantics in the communication protocol (made possible by directed diffusion) to avoid transmitting redundant messages.

Discussion:

• The military recently expressed a goal of deploying video camera equipped sensor nodes in less than ten years
• Determining geographic location can an important (and, in this paper, neglected) aspect of sensor networks
1. Events’ locations will often be important information to collect
2. Geographic knowledge could be used to reduce redundant flooding
• Bluetooth (with TDMA) or token schemes are perhaps better than 802.11 (with CSMA) for conserving power in wireless networks. However, 802.11, or ‘wireless Ethernet,’ seems to have been adopted as a path of least resistance to reusing IP in wireless settings.
• Process for the acquisition of locally unique identifiers not explained in paper; could use ad hoc techniques such as having elected cluster heads assign local IDs—adds overhead
• Localized interaction paradigm similar to anonymous networking work, in which nodes are restricted from acquiring global state for privacy rather than scalability issues
• No mention made of congestion