1
play

1 In-network data aggregation What is directed diffusion? Old way - PDF document

Nov 17, 2023 •258 likes •363 views

Motivation Directed Diffusion for Wireless Content and data-centric Sensor Networking Where are nodes with X data over the next 5 minutes, at 1 second intervals? Where is an object within some region? With material from the

  1. Motivation Directed Diffusion for Wireless � Content and data-centric Sensor Networking � Where are nodes with X data over the next 5 minutes, at 1 second intervals? � Where is an object within some region? With material from the � Challenges: subsequent paper, Building � Scalability Efficient Wireless Sensor Networks � Energy efficient with Low-Level Naming � Fault tolerant By Eric Siegfried CSE 521 11/29/2004 1 11/29/2004 2 Building Efficient Wireless Sensor ID-based communication Networks with Low Level Naming � Requires unique host ID addressing � Cost of computation vs. communication � Application is end to end � 3000 instructions vs. 1 bit 100m � Attribute based communications � Integrating nodes into the network � Linda, LIME (tuple space) � Tendency for less robust routes � Typically expensive, not good for resource constrained networks 11/29/2004 3 11/29/2004 4 Routing issues Filters � DSR and AODV recreate IP style � Access information about diffusion network (gradients, list of neighbors) for in- network aggregation � Preferable to be based on attribute- value-operation tuples? � Can help to avoid flooding � Can deal with rectangular region � Peer to peer behavior, all can filter � Can deal with interests, non specific node � Power efficient 11/29/2004 5 11/29/2004 6 1

  2. In-network data aggregation What is directed diffusion? � Old way � Interests � Binding a service to a geographical region � Data which lists the node identifiers � Messages � Elect one or more nodes to aggregate � Gradients � New way � Reinforcements � Name nodes via geographic attributes � Run application specific filters, and inject application specific code into the network 11/29/2004 7 11/29/2004 8 Naming tasks Interest propagation Interest Reply � Interests diffuse through the network, periodically refreshed by the sink type = four-legged animal type = four-legged animal � Interests only contain information about interval = 20 ms instance = [125, 220] duration = 10 sec intensity = 0.6 the previous hop rect = [-100, 100, 200, 400] confidence = 0.85 � Checks cache, may create an entry timestamp = 01:20:40 � Checks for gradient, may add one � Reinforced path, delivery of data 11/29/2004 9 11/29/2004 10 Gradient Establishment Data propagation � value, and direction � Seeing data which meets a certain intensity and confidence � Rate per hour � Timestamp � Referenced in interest cache � expiration � Gradient data rate vs. event data rate 11/29/2004 11 11/29/2004 12 2

  3. Reinforcement, making paths Exploratory Gradient Exploratory Request � Exploratory events Gradient � Positive reinforcement Event Event � Dealing with multiple sources and sinks � Include figure 3c, 3d Low Low Low Bidirectional gradients established on all links through flooding 11/29/2004 13 11/29/2004 14 Illustrating Directed Diffusion Local Behavior Choices 1. For propagating interests 3. For data transmission Different local rules can result In our example, flood in single path delivery, More sophisticated striped multi-path delivery, behaviors possible: e.g. single source to multiple based on cached sinks and so on. information, GPS 4. For reinforcement 2. For setting up gradients Sink reinforce one path, or part Highest gradient towards thereof, based on observed neighbor from whom we losses, delay variances etc. first heard interest other variants: inhibit certain Others possible: towards paths because resource neighbor with highest levels are low energy 11/29/2004 15 11/29/2004 16 Illustrating Directed Diffusion Illustrating Directed Diffusion Sink Sink 11/29/2004 17 11/29/2004 18 3

  4. Illustrating Directed Diffusion Illustrating Directed Diffusion Sink Sink 11/29/2004 19 11/29/2004 20 Illustrating Directed Diffusion Illustrating Directed Diffusion Sink Sink 11/29/2004 21 11/29/2004 22 Illustrating Directed Diffusion Illustrating Directed Diffusion Source Sink Sink 11/29/2004 23 11/29/2004 24 4

  5. Illustrating Directed Diffusion Illustrating Directed Diffusion Source Source Sink Sink 11/29/2004 25 11/29/2004 26 Illustrating Directed Diffusion Illustrating Directed Diffusion Source Source Sink Sink 11/29/2004 27 11/29/2004 28 Illustrating Directed Diffusion Illustrating Directed Diffusion Source Source Sink Sink 11/29/2004 29 11/29/2004 30 5

  6. Illustrating Directed Diffusion Illustrating Directed Diffusion Source Source Sink Sink 11/29/2004 31 11/29/2004 32 Illustrating Directed Diffusion Illustrating Directed Diffusion Source Source Sink Sink 11/29/2004 33 11/29/2004 34 Illustrating Directed Diffusion Illustrating Directed Diffusion Source Source Sink Sink 11/29/2004 35 11/29/2004 36 6

  7. Illustrating Directed Diffusion Illustrating Directed Diffusion Source Source Sink Sink 11/29/2004 37 11/29/2004 38 Negative reinforcement (path Local repair for failed paths truncation) � Reaction to corruption, degradation � Include figure 4a � Include figure 3d � Soft state to time out data gradients � Wasting resources to find the lossy link � Reset gradients to be exploratory 11/29/2004 39 11/29/2004 40 Negative Reinforcement (loop removal) Simulation � Not always removed � Figure 4b vs 4c � Some useful paths even though looped 11/29/2004 41 11/29/2004 42 7

  8. Simulation(2) Simulation(3) 11/29/2004 43 11/29/2004 44 Simulation(4) Conclusions � Nature of directed diffusion paper � Weaknesses � Did non develop software architecture for realizing attributes and in-network processing in an OS � Simulation did not include radio propagation � Costly to have exploratory packets 11/29/2004 45 11/29/2004 46 Micro-Diffusion Nested Queries Preference to second � 5.5 KB of memory method, more power efficient. � Supports 5 gradients and 10 packets (2B per packet) per node � Reduced in size, but header format is compatible with full diffusion implementation � Filters currently not implemented for micro-diffusion 11/29/2004 47 11/29/2004 48 8

  9. Energy Savings Power Savings(2) � If no aggregation, each source pays cost of full path, whereas aggregation Suppression uses pays only one hop prior to aggregation less data per event with � Understanding aggregation is an multiple sources important area of future work � Saves around 40% traffic to aggregate � Nested queries reduce loss by 15-30% 11/29/2004 49 11/29/2004 50 Future work Future work(2) � How diffusion’s parameters map to � Energy aware MAC protocol needed different needs � New applications, and how signal � Trade offs between overhead and processing interacts (for example) with reliability in exploratory messages, in network processing and filters interests, and reinforcements � Impact on wired networks � Adding feedback and congestion control to diffusion 11/29/2004 51 11/29/2004 52 Credits � Dushanth � Sandeep Gupta 11/29/2004 53 9

Recommend

More recommend