250 likes | 350 Views
Lance: Optimizing High-Resolution Signal Collection in Wireless Sensor Networks. Geoffrey Werner-Allen, Stephen Dawson-Haggerty, and Matt Welsh School of Engineering and Applied Sciences, Harvard University, Cambridge SenSys 2008. Outline. Introduction Lance System Architecture
E N D
Lance: Optimizing High-Resolution Signal Collection in Wireless Sensor Networks Geoffrey Werner-Allen, Stephen Dawson-Haggerty, and Matt Welsh School of Engineering and Applied Sciences, Harvard University, Cambridge SenSys 2008
Outline • Introduction • Lance System Architecture • Policy Module • Case Study • Implementation • Evaluation • Deployment • Conclusion
Introduction • Acquisition of high resolution signals using low-power wireless sensor nodes. • EX: Acoustic, seismic, and vibration waveforms (high data rate) • Constraint : (1). Radio bandwidth (2). Energy usage • Typically: Only focus on “interesting” signal • Managing: • Limited energy capacity • Severely constrainedradio bandwidth. • Depending on the (1). Sampling rate and (2). Resolution. • Low-power sensor node radios single-hop throughput : 100 Kbps • The best reliable protocols :Less than8 Kbps for a single transfer over multiple hops.
Introduction • Lance -- General approach to bandwidth and energy management for reliable signal collection. • Application data unit (ADU) : • Summary (node) • Value (base station) • download schedule (lance) • Incorporates a cost estimator that predicts the energy cost for reliably downloading each ADU from the network. • Cost estimator : Load-balancing download operations. • Policy modules optimization metrics • lifetime targeting • acquiring temporally- or spatially-correlated data
Lance System Architecture • ADU : Unit of data storage and retrieval • Each unique aiconsists of a tuple {i, ni, ti, di, vi, ¯ci} • i : ADU identifier • ni : The node storing the ADU • ti : Timestamp • di : Raw sensor data • vi :Application Specific value • ¯ci : Energy requirement to download the ADU from the network • ¯ci : vector , the estimated energy expenditure of node j when ADU i is retrieved. • Assume : ADUs are of uniform size and that nodes have sufficient flash storage to buffer collected signals.
Lance System Architecture • Energy model: • Cost for downloading the ADU. • Cost to nodes that overhear transmissions by nodes participating in the transfer. • a priori assumption : Battery capacity C joule • Lifetime target : L (each node) • discharge rate : no more than C/L (joule/per unit time) • High Level Goal : • Download the set of ADUs that maximizes the total value, subject to the lifetime target. • epoch duration ∆ • : Multidimensional knapsack problem
Lance System Architecture • Design Principle • Decouple mechanism from policy • Simplicity through centralized control • Low cost for maintenance traffic • System Overview:
Lance System Architecture • Two application-provided components • Summarization Function (node) : Local Information • Chain of Policy Module (base) : Global Information • Constraint on Summarization Function • Small Summary (a few bytes) limits the overhead for storing and transmitting • Run Efficiently as ADUs are sampled • Example: Seismic events • Commonly used measure : RSAM(Real-Time Seismic Amplitude Measurement)
Lance System Architecture • Cost Estimation • Compute Download Energy cost vector ¯ci for each ADU sampled by the network. • Assumption : Spanning tree topology rooted at base station. • Cost Function (factor) = (reliable transmission protocol )+ (node’s position in routing path)+ (radio link quality) + (MAC protocol) • Complex Dynamics in Sensor Network • Empirical Model—Three primitive energy cost: • Ed :reading data from flash + sending multiple radio pkt (including retransmit) to next hop • Er: intermediate node (forward message) • Eo: overhear transmission
Lance System Architecture • Lance Optimizer • Scheduling ADUs for download • Reliable Transmission Protocol—Fetch or Flush • Adhering life target L, maximize the values of ADUs retrieved. • Greedy heuristic approximation of the multidimensional knapsack. • Procedure • Step 1. Exclude ADUs without enough energy to perform a download. • Step 2. Determine next ADU to download (Scoring Function) • Scoring Function • 1. Value Only • 2. Cost Total • 3. Cost Bottleneck
Policy Module • Application-supplied function : Input ADUs • Produce new ai’ with a possibly modified value vi’ • Linear chain of policy modules m1,m2,m3… • Standard tool kit of policy module 1. Value Thresholding 2. Value Adjustment and noise removal 3. Value Dilation 4. Correlated Event Detection 5. Cost Based Filtering
Case Study • Geophysical monitoring • Volcano monitoring at Reventador • Seismic and acoustic data at 100 Hz per channel with a resolution of 24 bits/sample • Deficiencies • 1. System could not prioritize certain events over others. • 2. Following each trigger, the network initiated a nonpreemptive download from every node in the network in a round-robin fashion. • 3. No attempt to manage energy. • Adaptation to Lance • Node-level event detector ADU summarization function • Base Station Lance’s optimizer and policy modules
Case Study • Exponentially weighted moving averages (EWMA) of the seismic signal. • Short term average and Long term average • Ratio of two averages Summarization Function • Filter, Correlated, and SpacespreadPolicy module • Report max-ratio over ADU allowing Lance to prioritizing different events. • Download management is value-driven rather than FIFO • Avoiding the nonpreemptive download
Implementation • TinyOS 2.x for TMote Sky and iMote2 sensor nodes. • 1 MB flash memory (ST M25P80) divided into 16 sectors of 64 KB each. • Per ADU/ 64KB each • Collection Tree Protocol (TinyOS 2.0) • Nodes send a periodic storage summary to the base station. • Reliability consideration : last 5 ADUs / each summary • “Fetch” reliable transfer protocol
Evaluation • Simulator and Synthetic Data Set • Optimal solution : Maximize data value subject to bandwidth and energy constraints. • Optimality : the fraction of the data value downloaded by Lance compared to the optimal solution. • MoteLab • 10-node linear topology • 25-node realistic tree topology • Download speeds : Based on empirical measurements. • Three value distributions are used: • (1). Uniform random • (2). Exponentially distributed • (3). Zipf with exponent α= 1.
Evaluation • 10-node linear topology with exponentially-distributed ADU values.
Evaluation • Different lifetimes and value distributions, run on the 25-node tree topology.
Evaluation • Bandwidth adaptation • 25-node tree topology • cost-bottleneck scoring function • target lifetime at 8 months
Evaluation • Stress the system in a realistic setting subject to • radio interference and congestion • exercise the multihop routing protocol • Fetch reliable data-collection protocol • ADU summary traffic generated by the nodes • cost-bottleneck scoring function.
Deployment • Tungurahua Volcano • Lance was used to manage the bandwidth resources • Seven of the nodes were deployed in a three armed “star” topology radiating away from a central hub node
Deployment • RSAM-based summarization • Sensitive to DC bias(causing Lance to generally prefer downloading ADUs from one or two nodes (those with the largest positive bias).) • Introduce Policy Module • computing the median RSAM • subtracting the median
Conclusion • Wide range of application-specific resource management policies. • Lance achieves near-optimal data retrieval under a range of energy and bandwidth limitations, as well as varying data distributions. • Study the use of more sophisticated node-level data processing, including feature extraction, adaptation to changing energy availability, and data summarization.