Directed Diffusion: A Scalable and Robust Communication Paradigm for Sensor Networks
by Chalermek Intanagonwiwat, Ramesh Govindan, Deborah Estrin
url show details
Details
| month: | oct # "~31 | | type: | misc | | booktitle: | MOBICOM | | abstract: | Advances in processor, memory and radio technology will enable small and cheap nodes capable of sensing | | year: | 2000 | | address: | N. Y. | | pages: | 56--67 | | publisher: | ACM Press | | annote: | Ramesh Govindan (USC/Information Sciences Institute); Deborah Estrin (USC/Information Sciences Institute; and University of California , Los; Angeles); |
|
|
You need to log in to add tags and post comments.
This paper presents a new data dissemination paradigm for wireless sensor networks called directed diffusion. The main idea of the approach is that data is propagated through local interactions of neighboring sensor nodes. This approach differs from traditional networking in several important ways: e.g., it is data-centric, all communication is neighbor-to-neighbor (as opposed to "end-to-end"), there are no "routers", sensor nodes do not have globally unique identifiers/addresses, etc.
The paper focuses on the specific application of location tracking, although the authors argue that the example is representative of a large class of remote surveillance sensor networks. Simulations using ns-2 suggest its potential effectiveness in practice: "Even with relatively unoptimized path selection, it outperforms an idealized traditional data dissemination scheme like omniscient multicast."
The paper definitely made it sound like this was a revolutionary new idea that really deviated from all standard networking approaches; however, alot of these ideas seem somewhat similar to other protocols we've seen in the class, so I'm not really sure. For example, in EXOR/COPE we saw the idea of broadcasting a message to all the neighbors in hearing distance. I'm not sure of the chronology of this paper wrt the other wireless papers we read, so maybe someone wants to post on that. We also saw the idea of "caching" packets and reverse path propagation with other protocols. The reinforcement/negative reinforcement component of the algorithm did seem new.
The experimental results did seem very encouraging wrt energy efficiency, as directed diffusion performed "noticeably" better than omniscient multicast and much better than flooding, which I presume are the two standards of comparison. They also ran several other experiments changing various aspects of the network/algorithm to isolate the factors contributing to increased energy, although the conclusions weren't entirely clear to me.
The fact that the motivating application for this paper is surveillance does nothing to disuade me from the impression that Systems is another way of supporting "The Man" (see also pay-for-QoS arguments).
That aside, the notion of "soft" patch choosing (that is, having several paths active, but running only one at a high data rate) seems interesting. I would be very interested in seeing what sort of behaviour actually crops up in non-simulated situations; specifically, interference between paths may mean that non-optimal paths are chosen (because a high-data-rate path will probably tend to interfere with low-data-rate neighbors). Also, it would be interesting to investigate power consumption a bit more -- specifically, it seems like nodes along high-rate paths will use more power than fringe nodes. In a uniform-node envirnoment, this might effectively shorten network life by draining the batteries of all the "core" nodes.
I might be missing something, but I don't understand what are outstanding points of this paper compared to the ones that we have read. (considerations on energy efficiency and delay?) Also, as Cody pointed out, comparison analysis given in the paper seems simple.
Answering Sam's question on the chronology: This paper appeared in Mobicomm 2000, while ExOr and Xors in the air appeared in SigComm 2005 and 2006 respectively, so it is possible that some of the ideas were novel when this paper was published.
This paper is very interesting for me. We have to be aware that a power consumption is the most critical issue for any wireless sensor network. So, most nodes in the network have to turn off its wireless component and only turns on its wireless component when it has to send or receive data. I think this routing protocol is designed based on this assumption, so it seems to be different from other protocol we have read so far.
This paper presents an interesting idea for balancing transmission quality with power conservation: Initially, "interests" are broadcast throughout the graph, and data is transmitted with redundancy backwards along interest edges to the source. However, the empirically "best" path is repeatedly reinforced, until redundancy is minimized. The "interest" framework also provides an interesting way of thinking about the sensor-net problem, as compared to the sensornet-as-database view in the other paper we read.
This combines routing and data delivery by storing "interest" queries along the path. This approach can be not too scalable if there are many different "types" of queries.
But I think this approach will be very useful in combat fields where each unit acts as both source and sink. So both source and sink moves around and a sink wants to detect events. Also if an event occurs (engagement, etc), they want frequent update on that.
I think Direct Diffusion is a well-suited data dissemination for wireless sensor networks, basically because 1) sensor network is energy-limited, 2) wireless channel is lossy, 3) there is no router in sensor networks, 4) there is some levels of dynamic in the network, i.e. some neighboring nodes might die because of running out of battery.
From the paper, we've found that while both algorithms incur comparative average delay, direct diffusion is more energy conservative than omniscient multicast. However, it might be interesting to see a comparison between direct diffusion and omniscient multicast on delivery ratio, as well.
I agree with Sam that we have seen a number of similar ideas like flooding-type propagation in other, more recent paper. Probably because the paper is proposing a new data-centric scheme (rather than node-centric), it introduces a number of fancy names such as interests (tasks?), gradients (reverse paths?). I doubt that applications such as location tracking is representative. At least, I think questions like "how many people do you observe in a region?" may be too high level (includes pattern recognition?) to match real industrial application needs.
Its interesting to note the expectations of the author from "future sensor nodes". They are almost exactly what we have now, except that instead of a tdma mac layer, csma is being used widely. (Although we could have a tdma mac). So todays researchers can expect the future sensors to be something like "sensor dust"??
It is interesting to see all this kind of work focus on the energy-efficient, loss-tolerant. This direct diffusion in this paper is a data-centric approach, and it works.