Overview of Directed Diffusion Professor: -Dr Ajay Gupta Presented By: -Vivek Kinra CS691 Spring2003 vivek kinra CS-WMU 1 Note: -Various slides of this presentation are created with the help of presentation slides of UCLA ,USC and various other sources vivek kinra CS-WMU 2 History Research started to investigate the design of localized algorithm using the Directed Diffusion model The idea was developed in the context of a DARPA study by D.Estrin Example of posing query for tanks/vehicles…….. vivek kinra CS-WMU 3 Design Features Data centric: -Routing is based on data contained in sensor node and may not need ID Application focus on the data generated by sensors. Data is named by attributes and applications request data matching certain attribute values. Motivated by robustness, scaling and energy efficiency vivek kinra CS-WMU 4 Directed Diffusion Developed by ISI/USC and UCLA is a novel network protocol built for info retrieval and data dissemination. Data generated by nodes => attributes(A1) Sinks/nodes request data=>Interest into n/w If A1 == Interest then(gradient setup in n/w) (Pedestrians) vivek kinra CS-WMU 5 contd Data pulled towards sinks =>receiver Initiated routing protocol Example target tracking Intermediate node might aggregate data Since all nodes in directed diffusion are application aware so It is completly application oriented. vivek kinra CS-WMU 6 contd It is significantly different from IP style communication Not infeasible with IP or Ad-hoc routing Imp Feature: - interest, data aggregation and propagation are determined by localized interaction vivek kinra CS-WMU 7 Expected Architecture of Sensor Network Required capabilities of sensor node: A Match box sized form factor Battery power source Power conserving processor clocked at several hundred Mhz Memory Radio modem vivek kinra CS-WMU 8 contd Energy efficient MAC layer Can have more than 1 or more sensors e.g seismic geophones, infrared dipoles etc The Atod conversion on such system produce 70ksamples/sec and 12 bit resolution vivek kinra CS-WMU 9 For power issue, common signal processing functions offloaded to low power ASIC Processor woke up only when event of Interest A Sensor Node have a GPS receiver The adv. Of these sensors is with very cheap in cost they obtain high SNR (attenuate with distance). Also can be deployed in huge amount vivek kinra CS-WMU 10 Energy concern Sensors Deployment falls in two ways: Large complex system deployed far. Short range hop-hop communication is preferred over direct long range. Local computation to reduce data before transmission vivek kinra CS-WMU 11 Contd In this organization, individual nodes reduce the sampled waveform generated by target (tank etc) into a relatively coarse grained “event” description. Description =>”codebook value” (event code) Code->a timestamp,…… Nodes exchanged this event code vivek kinra CS-WMU 12 Method description Task conveyed to sensor N/W Nodes tasks it’s sensors Matches sampled wave form against locally stored library Sensors in region may coordinate to pick best estimate. Packet:-Attributes (type, amplitude, Intensity, region, time stamp……) vivek kinra CS-WMU 13 Naming Given Set of Tasks supported by sensor network selecting a naming scheme is first step in designing sensor networks. Basically list of attribute value pairs. E.g. For tracking animal its attributes should describe tasks like, type of animal, geographic location to track, interval for sending updates, duration for which it was recorded (event occurrence time) vivek kinra CS-WMU 14 Data sent in response to Interest Type = four legged animal Instance = rabbit//instance of type location = [125,220]/node location Intensity = 0.6/signal amplitude Confidence = 0.85//confi.. in match Timestamp = 01:20:40//event generation time vivek kinra CS-WMU 15 Sink periodically broadcasts an interest message to each of its neighbors. Initial interest specifies a low data rate (e.g 1 event/sec) Interest are diff based on type, rect or interval Every node maintains a interest cache. Interest entries in cache do not contain info about sink vivek kinra CS-WMU 16 Interest entry Time stamp (last received matching) Gradient field (up to 1/neighbor) G.F => data rate field (requested by neighbor)=>interval attribute Duration=timestamp – expiresAT No Entry No gradient vivek kinra CS-WMU 17 Event interests Sink Have u seen any four leg animal??? QUERY DIFFUSED IN TO INTEREST WHICH IS LIST OF ATTRIBUTE VALUE PAIRS Interest Propagation (Flooding) vivek kinra CS-WMU 18 YES I HAVE SEEN ONE…. INTIAL GRADIENTS SETUP(VALUE+DIRECTION) Two-way Gradient setup vivek kinra CS-WMU 19 Gradient setup/reinforced path source Sink/Interest I-Propagation Initial grad.. viveksetup kinra CS-WMU Data …..reinforced path 20 Interest/gradient Task ={type,rect,a duration of 10 min}is instantiated at particular node Interval :- event data rate Sink periodically broadcast interest msg (& refresh interest) to neighbors. Initial Interest :-{rect,duration attributes,larger interval attribute} Gradient expiration vivek kinra CS-WMU 21 DATA DELIVERY THROUGH REINFORCED PATH SINGLE PATH DELIVERY (CAN BE MULTIPATH ALSO) vivek kinra CS-WMU 22 IN CASE OF NODE FAILURE USE ALTERNATIVE PATHS vivek kinra CS-WMU 23 Reinforcement When to reinforce ?(quality/delay matrices can be chosen) Whom to reinforce ? How many to reinforce? When to send negative reinforcement vivek kinra CS-WMU 24 When?? Sink initially diffuses a interest for a low event-rate. Once sources starts detect a matching target they send low rate events. After the sink starts receiving these low data rate events it reinforces one particular neighbor to draw down higher quality. vivek kinra CS-WMU 25 Whom?? To reinforce this neighbor, the sink resends the original interest message but with smaller interval (higher data rate). Two approaches for reinforce Incremental approach:- Add min # of links to existing tree Select links so that min energy is used vivek kinra CS-WMU 26 How Many Node must reinforce at least one neighbor vivek kinra CS-WMU 27 Negative Reinforcement Earlier used A but now B is better One way :- time out all high data gradients in the n/w Sink would periodically reinforce B and cease A that will degrade the path to A to lower data rate Other way-:Degrade the path to A by resending the interest with low data rate vivek kinra CS-WMU 28 Whether to negatively reinforce or not N.R those neighbor from which no new event have been received. Or few events are coming. Significant experiments are required before deciding which local rule achieve an energy efficient global behaviour vivek kinra CS-WMU 29 Issues of Concern Ad hoc, self organizing, adaptive systems with predictable behavior Collaborative processing, data fusion, multiple sensory modalities Data analysis/mining vivek kinra CS-WMU 30 Issues yet to be resolved How to handle congested network? Semantics for gradients. Handling of more than one sources. Negative reinforcement increases delay and contention vivek kinra CS-WMU 31 comments (battery life, size, processing power, memory, etc.)? The paper presents a motion-detection scenario for sensor networks. To identify an event sources must match sampled sensor waveforms against signatures stored in a local library. To be useful, this library may have to store several thousand such signatures or more. We could implement "task-centric" sensor networks, where sensor nodes are focused on one or two type of event detection. vivek kinra CS-WMU 32 Tiny Diffusion Implementation of Diffusion on resource constrained USB motes 8 bit CPU, 8k program memory, 512 bytes data memory Subsets of full system Retains only gradients and condenses attributes to a single tag Entire system runs for less than 5.5 KB memory vivek kinra CS-WMU 33 contd Tiny OS adds ~3.5 KB and 144 bytes of data (inclusive support for radio and photo sensor Diffusion adds ~2k code and 110 bytes of data to tiny OS vivek kinra CS-WMU 34 Tiny Diffusion Functionality Resource Constraint Limited Cache size-currently 10 entries of 2 bytes each Limited ability to support multiple traffic stream. currently support 5 concurrently active gradients vivek kinra CS-WMU 35 TinyOS Implementation vivek kinra CS-WMU 36 vivek kinra CS-WMU 37 Gateway Architecture Photo Data Source Data Sink TinyDiffusion TINYOS Acoustic Data Source Query Data Sink DIFFUSION LINUX Device Driver MOTE ATMEL 8586 4MHz MCU 8K program memory 512 Bytes Data Memory RFM Radio 900 MHz RFM Transceiver TINYOS vivek kinra CS-WMU PC104 AMD Elan™SC400 66MHz CPU 16MB RAM Form Factor: 3.6" x 3.8" x 0.6" 38 Tiered Testbed PC-104+(linux) with MoteNIC Tags, Sensor Card UCB Motes w/TinyOS Yet to come: SmartDust (highly specialized nodes) PS104 TAG USB Mote vivek kinra CS-WMU 39 vivek kinra CS-WMU 40