Design of Energy Efficient Computations and Protocols for Wireless Sensor Networks Ashfaq A. Khokhar Multimedia Systems Lab (http://multimedia.ece.uic.edu/) University of Illinois at Chicago Wireless Sensor Networks • A large number of sensor nodes with limited capability of computation, communication and sensing. • Nodes collaborate with each other through a wireless channel to accomplish an assigned task. Sample Sensor Nodes Modern Sensor Nodes UC Berkeley: COTS Dust UC Berkeley: COTS Dust UC Berkeley: Smart Dust Courtesy UC Berkley UCLA: WINS Rockwell: WINS JPL: Sensor Webs Features of WSN Traffic rate is generally low – Sensor nodes are battery powered – – typical communication frequency is in seconds or minutes. recharging is usually unavailable energy is an extremely expensive resource Sensor nodes in the network coordinate with each other to implement a certain function, – traffic is not random as in mobile ad hoc networks. Motivation Wireless Sensor Networks is one of the top 10 Technologies that will change the World in 21st Century According to MIT Technology Review Pervasive computing environments are increasing Abundance of defense, scientific, and commercial applications Wireless Medium Popularity Phenomenal growth in – – – New High Bit Rate Wireless Services – – – Mobile Communications Internet/Intranet, E-Commerce Use of Laptops, Palmtops, and PDAs Intranet/Internet Multimedia: Integrated Voice, Data, Video High quality voice and Videoconferencing New Technology means new products/services – Revenue opportunities Market Estimation WSN: $150 Million in 2003, $7 Billion estimated in 2010 (ON World) Mobility infrastructure market expected to grow from $25.7 Billion in 2004 to $34.8 Billion in 2008 (Dell’Oro Group) Today more wireless connections than wired lines Typical WSN Applications Environment monitoring Transportation Habitat monitoring Office security Industrial monitoring Fire detection Challenges in Wireless Sensor Networks Software Systems – – – – – Networking and Communication – – – – Signal processing Classification Devices – VLSI integration Architectures Deployment Signal and Systems – Routing, Data Gathering, Data Dissemination Hardware: – Computing Control Databases Fusion Knowledge Extraction Sensor technologies User Interfaces and Development Environment Research at Multimedia Systems Lab Software Systems – Networking – – Power-Time Efficient Algorithms MAC layer Routing Layer Protocols Signals and Systems – – – Field Estimation Localization Classification Collaborative Computing over Sensor Networks Sensors are smart: – Exploit these resources and communicate with sink only when necessary – Similar arguments hold for computing among sinks Develop distributed algorithms which are power-time efficient: – processing, storage and communication capabilities Power-time product is comparable to sequential counterpart Contradicting goals: – – Exploit distributed computing resources Avoid redundant computations Example: Conventional Distributed FFT 0 0 1 1 2 3 x2 x3 2 3 x2 + w x3 4 5 6 7 4 5 6 7 x2 - w x3 0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7 Unbalanced Power-Aware FFT: (Ramesh et al -- Milcom 2003) 0 1 2 x2 0 1 2 x 2 + w x3 3 4 5 6 7 4 5 6 7 w*x3 3 x 2 - w x3 0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7 Power–Time Efficient Distributed 1-d FFT Algorithm 0 1 2 x2 0 1 0 1 3 4 5 6 7 8 91 w*xx33 2 x2 + w x3 3 x2 - w x3 2 3 4 5 6 7 8 9 4 5 6 7 8 9 1 x2 0 1 2 0 x2 + w x3 1 3 1 x3 w*x 3 1 3 1 x2 - w x3 1 0 1 1 1 2 1 5 3 1 4 1 7 5 1 2 1 3 1 6 4 1 7 5 1 2 1 3 1 4 1 5 Inverse-Shuffle Complement Permutation 1 5 7 1 1 3 1 13 3 5 9 1 1 4 6 1 0 2 1 12 2 4 8 0 1 5 7 1 1 3 1 13 3 5 9 1 1 4 6 1 0 2 1 12 2 4 8 0 1 5 7 1 1 3 1 3 5 9 1 1 4 6 1 0 2 1 2 4 8 0 RESULTS [SenSys04, AICSS 2006] Proposed/Conventional 1.4 1.2 Normalized Radio Cost 1 0.8 0.6 0.4 0.2 63 60 57 54 51 48 45 42 39 36 33 30 27 24 21 18 15 12 9 6 3 0 0 Sensors The ratio of total number of send/receives in the proposed and conventional 64-Sensored FFT algorithm •The Actual communication cost improvement based on our experiments result is 36%. •Theoretical cost improvement is 42%. •The discrepancy arises from packet dropping due to collisions. Energy(mJ) RESULTS 1200 Proposed 1000 Original 800 600 400 200 0 8 16 32 Number of Sensors 64 Energy consumption of the FFT computation as the number of sensors increased •Approximately 20% energy cost reduction for N=32 and 64. •The energy improvement is due to reduction in the FFT computation time, as sensors are all on during the computation. •Except for the network of 8, the overhead of id shuffle phase is justified from the overall savings. RESULTS Conventional FFT Proposed FFT The number of packets transmitted at each time-interval during the conventional and proposed FFT computation over a 64-sensor network sensor networks •The Congested intervals are reduced significantly. •Packet collision probability is reduced. •MAC protocols can shut down the radio circuitry aggressively. Example: Classification Apriori determination of categories Train for each category – – Test – – Expose to sample measurements Compute Mean (i,j) & Covariance (i,j) Classifier Algorithms Develop false positives, beliefs, etc Deploy – Identify detected object w.r.t. categories ? Classifier Algorithms Traditional Signal Processing (SP) algorithms – – Pro: used extensively for its known accuracy Con: computationally intensive Novel WSN classifiers Sub-optimal Classifiers [Kotecha & et. el. 2005] Influence fields [Arora & et. el. 2004] Differentiated Surveillance [Yan & et. el. 2003] – – Maximum Likelihood (ML) [Li & et. el. 2002; Duarte & Hu, 2003] k Nearest Neighbor (kNN) [Li & et. el. 2002; Duarte & Hu, 2003] Pro: simpler computationally Con: novice, accuracy? Our Goal: adapt existing SP algorithms for efficient classification in sensornets What’s the Problem? Events are f × d matrices f – modalities/features sensed d – temporal processing dimensions Typically f 50 & d 512 Mean is f × d and is f × f matrices -1 is also f × f Large matrix computations for every detected event – Matrix multiplication Inverse computation ! – – Unstable Expensive Power Consumption Comparison [HiPC 2006] log of power consumption Power consumption comparison of MAP with Jacobi (MAP-J) and MAP with LU (MAP-L) 10 8 6 MAP-J 4 MAP-L 2 0 10 20 30 40 50 60 features 70 80 90 100 d = 512 Communication Protocols A fresher look was needed Energy efficiency and network life time as the main goal – – – MAC Layer Network Layer: Geographical Routing Application Layer: Data Collection Geographic Routing in WSN Geographic routing: – – – Greedy forwarding: At each hop, data packet is relayed by a neighbor which is geographically closer to the destination than the packet holder in terms of Euclidean distance Communication void(hole): a network area where greedy forwarding fails to locate the next hop node in a packet holder’s neighborhood Some strategies must be used to handle the void to guarantee the packet delivery if a real path exist Existing Geographic Routing Protocols Beacon-based(state-based): PGSR…. Neighbor information collection: periodical beacon messages exchanged among neighbor nodes Next hop node is chosen from the neighbor table by the packet holder Right/left hand rule based on planar graph is used to handle void High communication overhead due to the beacon messages Beacon-free(state-free): IGF, PSGP, SIF… No neighbor information maintained in each node Next hop node is chosen based on a competition, which is triggered by a packet holder, among the packet holder’s neighbors Right/left hand rule or increasing transmission power methods are used to handle the void: Motivation for Void Handling Right/left-hand rule is impractical since it’s hard to build planar graph in beacon-free geographic routing protocols Increasing transmission power is not energy efficient History of void handling should be maintained and used for future data delivery Proposed Contention-based Geographic Routing Protocol (CGR-D) Beacon-free A depth first search (DFS) method is used to locate the next hop node at each hop An integrated cost function which combines forwarding area determination, void handling and load balancing is defined as the metric used in DFS method Each node maintains a local variable (void cost) which is an indication whether a node is in or close to a void area The void cost is updated per sink basis and periodically reset to 0 Cost Function If a node C is closer to the sink D than the current packet holder F: f (d, E, r, H) = [* (1-d/R) + * (1-E/E0) + *r + K*(1+H)]*T0 (1) Otherwise: f (d, E, r, H) = [* (1-d/R) + * (2-E/E0) + *(1+r) + K*(1+H)]*T0 Where (2) d = dist(F, D) – dist(C, D) dist(x,y)means distance between x and y H = voidCost(C) - voidCost(F) voidCost(x) is the void cost value of node x R is radio range E is the remaining energy of node C E0 is the initial energy r is a random number between [0,1] + + = 1 and , , > 0 K, T0 are system parameters Data Delivery Example for CGR-D Performance Evaluation General Data Gathering Data gathering problem: Sensing data and transmitting data to a sink. Clustering based solution: LEACH is a typical one. Nodes organized into clusters. Head collects data and transmits aggregated data to the sink. Sink Mobility of Sensor Nodes The nearest cluster head may change especially in a highmobility scenario. Keeping transmitting to the old cluster head consumes more energy. Sink Our Solution: Low-energy Dynamic Cluster Selection (LEDCS) Protocol [VTC 06] As in LEACH, time is divided into rounds and at the beginning of each round, each node i determines whether it is a cluster head in the current round with a predefined probability. Introduce a contention period at the beginning of each time frame. Nodes may join the new cluster head during this contention period. If a node is a cluster head in the current round, it broadcasts this info across the network which also includes its own moving direction and velocity in the current round. Our Solution: Low-energy Dynamic Cluster Selection (LEDCS) Protocol Simulation Results 1000 sensors in a 400m x 400m area Percentage of contention period considering different velocity of sensor nodes. Simulation Results Total number of data packets received by the sink considering different velocity of sensor nodes. Up to 80% higher Simulation Results Total number of data rounds Data Gathering in Event Driven Applications [INSS 07] Bursty traffic in event driven applications. Sink Estimation of Traffic Load Estimate the current traffic load at the beginning of each round. Based on the estimated value, set up mathematical model and determine optimal number of cluster heads to minimize the total energy consumption. the total energy consumption vs. optimal number of cluster heads Comparison of Total Energy Consumption Test case 1 Test case 2 Data Collection via Cross-Layer Optimization Data Collection: A major class of sensor network application ●Generally a spanning tree is used for data collection over a sensor network ● Problems: ● Congestion is major performance bottleneck. Goals: To increase data delivery ratio with simpler MAC protocols ●To mitigate congestion in the network especially near the sink node ●To increase bandwidth utilization ● Sink Sensor nodes MOTIVATION SMAC like contention based MAC protocols are simpler and do not require tight synchronization. ● A B C sleep active active active sleep active Question: In data collection applications is it absolutely necessary for a node to communicate every one of its hop neighbors? sleep sleep Sink -Not if we can find route to sink! -Node B does not have to communicate with C, so it can stop following C's schedule B A C Single Data Collection Tree •Single Data Collection Tree •Multihop communication •Single spanning tree is constructed and nodes forward their readings to their parents •Major performance bottlenecks: • Heavy congestion near the sink node. • High competition for the wireless medium. • High delays due to medium access Multiple Data Collection Trees Phase 1: Construct 2 different trees Multiple Data Collection Trees Phase 2 : Activate each tree at different times •Construct more than 1 collection tree •Active trees at different times than the others •Data collection is possible without communicating with all the neighbors •Routing protocol should find the set of neighbors necessary to communicate with sink. •Decreasing the number of active nodes will mitigate congestion and increase delivery ratio. Single vs. Multiple Data Collection Trees Single data collection tree: active sleep 1 Duty Cycle Advantage: Simple tree construction Disadvantages: ● Network is highly congested near sink ● Bandwidth is not fully utilized ● High collision probability Multiple Data Collection Trees Advantages: •Increased bandwidth utilization •Mitigated Congestion •Energy consumption not increased •Less nodes are active Disadvantage: More complex tree construction sleep active active sleep Simulation Results Delivery Ratio- 10% SMAC 0.85 0.75 0.65 0.55 0.45 0.35 SMAC 50 SMAC 100 0.25 SMAC 200 0.15 1 2 3 4 5 6 Num ber Of Trees 7 8 9 Simulation Results Delivery Ratio Regular Grid 1.2 1 0.8 0.6 50 0.4 100 150 0.2 200 0 1 2 3 4 N um be r Of Tr e e s 5 6 Simulation Results Energy Consum ption 810 790 SMAC 50 Joules 770 SMAC 100 750 730 710 690 670 650 1 2 3 4 5 6 Num ber of Trees 7 8 9 Data Gathering in Vehicular Networks Stationary Internet Gateways (IGWs) along the highway which provide interfaces connecting vehicles to the Internet, a gateway communicate with vehicles out of its transmission range via multihop communications Traditional 802.11 MAC solution: Multi-hopping scheme suffers from low throughput and starvation of packets originating from vehicles far away from gateways due to high collision Example: CVIA Protocol 1. 2. 3. Divide the line between gateways into segments , length of segments is equal to the transmission range of a vehicle Assign TDMA slots for each segments such that only segments out of transmission range will transmit in the same time slot– collision free In the figure S1,3,5,7 transmit in a time slot and s2,4,6,8 transmit in the next time slot CVIA Protocol Four phases in each time slot: 1) Temporary Router Selection phase: inbound and outbound temporary routers elected for each segments 2) Inbound router transmits collected data to outbound router 3) Outbound router collects data within its segment 4) Outbound router transmits data to the neighboring inbound router Problems Inter- and Intra-segment contention leads to higher packet losses The time length of a slot for local data gathering phase is uniform, leading to transmission delays Solutions: – ??? Distortion Analysis in Sensing Field Measuring Spatial-temporal Correlated Data Example of field measuring Gaussian correlated data. Consider the real-time data gathering problem in a field that data is both spatial and temporal correlated. Sink will do real-time data reconstruction for the whole field. How many nodes should be put into the field to minimize the total distortion? Number of nodes increases: Spatial distortion decreases while temporal distortion increases. An optimal number of nodes exists to minimize the total distortion. Randomly deployed single-hop sensor networks Nodes randomly deployed in the field following Poisson distribution. 1. One-dimensional case 2. Two-dimensional case Use Voroni Cell partitions to achieve minimal distortion within each cell D( x1 , t1 ; x2 , t 2 ) E[(Y ( x1 , t1 ) Y ( x2 , t 2 )) 2 ] 2 2e n (( x1 x2 ) 2 2 ( t1 t 2 ) 2 ) D1 (t ) Pr{S ( r ) v}2(1 e ( r 2 2 ( iT t ) 2 ) )dr i 1 TDMA protocol used for data collection Randomly deployed one-hop sensor networks 1. An optimal number of nodes always exist to minimize the total distortion D. 2. when correlation intensity α is fixed and time scaling constant βT is increased, the optimal value of n that minimizes the total distortion D decreases. Fixed topology for multi-hop sensor networks We assume that each node has a limited transmission range R A TDMA-based transmission algorithm is designed for collision-free data transmission from nodes to the sink. An Example 1D Transmission Schedule Total distortion varies with different coefficients (one-dimensional case) Total distortion varies with different coefficients (two-dimensional case) a) α=0.1,βT=0.001 b) α=0.1,βT=0.002 c)α=0.1,βT=0.005 d) α=0.02,βT=0.002 e) α=0.05,βT=0.001 f) α=0.05,βT=0.002 Analysis for both cases The total distortion experiences a sudden drop for every increase of five in number of nodes (one-dimensional case) or number of rings (two-dimensional case). With the increase of correlation coefficients α and βT, for a given number of nodes, the total distortion will increase due to the weaker correlation of the field. Minimum number of nodes required given a certain distortion constraint Correlation Coefficients Distortion const raint Minimum number of nodes Correlation Coefficients Distortion const raint Minimum number of nodes n*k(n, k) α=1, βT=0.2 12 35 α=0.1, βT=0.005 120 50(10,5) α=1, βT=0.002 4 6 α=0.1, βT=0.002 60 90(18,5) α=1, βT=0.002 1.5 14 α=0.1, βT=0.001 40 120(24,5) α=0.5, βT=0.02 4 5 α=0.1, βT=0.001 120 45(9,5) α=0.5, βT=0.002 1.5 7 α=0.05, βT=0.005 120 35(7,5) α=0.5, βT=0.002 1 14 α=0.05, βT=0.002 40 70(14,5) α=0.2, βT=0.002 1.5 5 α=0.02, βT=0.002 60 30(6,5) One-dimensional grid network case α=0.2, βT=0.002 1 6 Two-dimensional wheel-based 40 35(7,5) network case α=0.02, βT=0.002 Other Problems Under Investigation Localization Synchronization Upper layer communication protocols Data fusion Knowledge extraction