Presentation on theme: "Dynamic Data Compression in Multi-hop Wireless Networks Abhishek B. Sharma (USC) Collaborators: Leana Golubchik Ramesh Govindan Michael J. Neely."— Presentation transcript:
Dynamic Data Compression in Multi-hop Wireless Networks Abhishek B. Sharma (USC) Collaborators: Leana Golubchik Ramesh Govindan Michael J. Neely
Data collection application in sensor networks Sensor nodes collect measurements that must be delivered at a sink. Multi-hop routing over a tree. Radios have limited transmission range Energy constrained Nodes are battery powered. 2 SIGMETRICS/Performance'09
3 Wireless sensor network platforms: Radio is the energy hog Figure from Sadler and Martonosi (SenSys 2006) Sensor network radios Transmission range: increases # CPU cycles for same energy as 1 byte transmitted Processor: MSP430 Data transmission is expensive.
Energy efficient data collection applications Need to transmit data using small energy budget. Challenge: Transmission costs lots of energy. Data is transmitted across multiple hops. Solution: Send less. compress data before transmitting Energy cost of compression. Not just CPU computations. Memory access, FLASH access SIGMETRICS/Performance'09 4 Transmission vs. Compression energy trade-off.
Related work: Single vs. multi-hop routing (Sadler et al., SenSys’06). Evaluating the energy efficiency of various algorithms. (Barr et al., MobiSys’03). Designing “light” yet energy efficient compression algorithms (Sadler et al., SenSys’06). Sadler et. al., SenSys’06 Single-hop: data compression does not save energy Multi-hop: data compression saves energy. “always compress” is not optimal. Energy trade-off was not explored in a “dynamic” environment. Data compression: Exploring the energy trade-off SIGMETRICS/Performance'09 5
System dynamics SIGMETRICS/Performance'09 6 Sink AB Energy w/o comp.comp. Sink AB w/o comp.comp. Energy Don’t compressCompress System dynamics impact the energy savings from compression. Sink AB w/o comp.comp. Energy Don’t compress
Compression decision in a dynamic environment Compression decision: “When to compress?” Must adapt to system dynamics. 1. Network dynamics: Link quality, topology 2. Application-level: sampling rate 3. Platform upgrade: low power radios, compression algorithm “When to compress” is not straight forward to determine. “Always compress” policy may not work well. SIGMETRICS/Performance'09 7
Data compression in a dynamic environment: Stochastic Network Optimization The application data arrival process and time varying link qualities are modeled as ergodic stochatic processes. Goal: Minimize the total system energy expenditure. System energy expenditure: total energy expenditure across all the nodes. Constraint: Network is “stable” bounded average queue size at all the nodes. implies finite delay in delivering data to the sink. SIGMETRICS/Performance'09 8
Stochastic Network Optimization: Lyapunov Optimization technique 1 Lyapunov drift analysis Arrival process Link dynamics Stability “Backpressure” based transmission decisions Compression at the source Arrival process Link dynamics Lyapunov drift analysis + Utility (energy cost) Stability Energy- efficient “Backpressure” based transmission decisions Compression decision algorithm Lyapunov Optimization: joint decision 1 Georgiadis, Neely and Tassiulas. Resource Allocation and Cross Layer Control in Wireless Networks, Foundations and Trends in Networking.
“Joint” compression and transmission decisions SIGMETRICS/Performance'09 10 Transmission Decision Algorithm Compression Decision Algorithm Data transfer rate Lots of retransmissions Application data rate
Our contributions 1. Stochastic network optimization formulation First to consider data compression for multi-hop networks in a dynamic environment. 2. Derive a “joint” congestion and transmission decision algorithm. 3. Prove stability and analytical performance bounds. 4. Propose and evaluate a distributed version. Works with CSMA MACs: 802.11, 802.15.4 SIGMETRICS/Performance'09 11
SIGMETRICS/Performance'09 12 SEEC: Stable and Energy Efficient Compression System Model Compression Module Transmission Module Application Data l [t] = C(link quality, trans. power) Data from other nodes U n [t] U n [t]: Queue backlog Maintains a table of avg. compression ratio and avg. energy cost for each comp. option k. Node n m U l [t] = U n [t] - U m [t] Decisions (every time slot t): Compression decision: whether to compress ? which option? Transmission decision: which nodes should transmit data?
SEEC: Transmission schedule “Queue differential backlog” based Each link is assigned a weight. Negative weight on a link Either due to a small queue backlog or poor link quality SIGMETRICS/Performance'09 13 Differential backlog Transmission rate Control parameter Transmit power Transmission scheduler Link weights Positive weight links on which data transfer is allowed Scheduling constraints
Transmission decision: Impact on queue backlog A node does not get to transmit till its backlog is greater than transmission threshold [t] = O (V/ [t]). Weight on its outgoing link should be positive. Increasing V results in higher queue backlog. Higher delay in delivering data to the sink. Avg. queue backlog grows will hop-count distance from the sink. SIGMETRICS/Performance'09 14 Sink
Compression decision: Driven by queue backlog A node compresses data only when its queue backlog is greater than compression threshold [t]. Directly proportional to compression energy cost. Inversely proportional to the average compression ratio. Increases as we increase the V. SEEC does not compute these thresholds explicitly. SIGMETRICS/Performance'09 15
Example: SEEC in action Transmit power = P (fixed) Link quality: “Good”= 2 Mbps, “Bad” = 1 Mbps SIGMETRICS/Performance'09 16 Sink AB time Node ANode B Queue backlog A [t] A [t] B [t] B [t] No compression Both links are “Good”Link from A to sink becomes “Bad” Node B starts compressing data
SEEC’s Performance: Energy vs. Delay trade-off SIGMETRICS/Performance'09 17 V (control parameter) P*P* Theorem:
Distributed version: Implementing SEEC’s transmission decision Finding the optimum transmission schedule is NP- complete. Approximation algorithms are known. 1. Global vs. Local information. 2. 802.11, 802.15.4 MACs: CSMA based (no timeslots). Positive queue differential heuristic (Sridharan et al.) Contend if (outgoing) link weight is positive Distributed version: dSEEC. SIGMETRICS/Performance'09 18
SIGMETRICS/Performance'09 19 Evaluation using Simulations Determining the model parameters Compression ratio and energy cost, transmission energy cost Measurements on real hardware: LEAP2 Radio: 802.11b Compressed real-world sensor data from a bridge vibrations monitoring deployment (Paek et al.’ 06). Compression algorithm: zlib compression libraries. Simulator: Qualnet
dSEEC: Summary of simulation results. 1. 10-30% energy savings compared to “always compress”. Tree-topology impacts the savings. SIGMETRICS/Performance'09 20
Compare with “Always compress” Cluster-Tree topology 1 1 Used in several deployments: Paek (WCSCM’06), Hicks (ImageSense’08) Periodic application data arrival Link quality did not change. Never compress dSEEC Always compress 30 % reduction
dSEEC: Summary of simulation results. 1. 10-30% energy savings compared to “always compress”. Tree-topology impacts the savings. 2. Able to adapt to system dynamics. 3. Sensitivity of energy savings to V SIGMETRICS/Performance'09 22 Lots of simulation results in the paper
Conclusion 1. Derived an algorithm for making compression decisions that is stable, energy-efficient, and adapts to system dynamics. Our work is the first to propose an adaptive algorithm for the multi-hop networks. 2. Energy vs. Delay trade-off Proved Analytical bounds 3. dSEEC: distributed version suited for CSMA MACs 4. Significant energy savings compared to simple policies. Future direction: Consider in-network data aggregation and compression. SIGMETRICS/Performance'09 23
SIGMETRICS/Performance'09 24 Algorithm derivation; proofs available in technical report. http://enl.usc.edu/~abhishek http://enl.usc.edu/~abhishek Questions?