1. 1 An Evaluation of Multi-resolution Storage for Sensor Networks D. Ganesan, B. Greenstein, D. Perelyubskiy, D. Estrin, J. Heidemann ACM SenSys 2003

2. 2 Papers DIMENSIONS: Why do we need a new Data Handling architecture for Sensor Networks Hotnets-I 2002 An Evaluation of Multi-resolution Storage for Sensor Networks ACM Sensys 2003 Multi-resolution Storage and Search in Sensor Networks ACM Transactions On Storage 2005

3. 3 Outline Introduction Dimensions Architecture Aging Problem Formulation System Implementation Experimental Evaluation Future Work and Conclusion

4. 4 Introduction

5. 5 Dimensions Architecture Spatial and Temporal Summarization

6. 6 Dimensions Architecture Drill-down Querying

7. 7 Hierarchy Construction From the view of communication cluster head

8. 8 Hierarchy Construction From the view of local storage cluster head

9. 9 Hierarchy Construction Load-balancing Scheme cluster head

10. 10 Hierarchy Construction Processing at each level … local storage data retrieval x y time At level i… compressed summaries from children node... Reconstructed Data Cube for future query…

11. 11 Storage Utilization Circular Buffer All available storage gets filled…  When to drop these summaries?  How to drop these summaries?  Graceful query quality degradation. local storage capacity Resolution 4 Resolution 1Resolution 2 Resolution 3 Local Storage Allocation

12. 12 Graceful Degradation Long-term Storage vs. Query Quality (1/2) Level 0 Level 1 Level 2 Time presentpast Query Accuracy High query accuracy Low compactness Low query accuracy High compactness low high

13. 13 Graceful Degradation Long-term Storage vs. Query Quality (2/2) Example: gracefully degrading storage the coarsest summaries, the longest period of time How long should a summary be stored in the network? progressively shorter time period

14. 14 Aging Problem Communication Overhead communication rate at level i  total amount of data from level i to i+1  level i level i+1

15. 15 Aging Problem Query Quality and Storage Overhead  Query accuracy if a drill-down terminates at level i   The amount of data that each node allocates for summaries from level i

16. 16 Aging Problem Approximate User-specified Aging Function Query Accuracy Time Quality Difference present past 95% 50% user-desired quality degradation system-provided step function Objective: minimize the worst case quality difference

17. 17 Aging Problem Given Other Constraints Drill-down constraint  Storage constraint  S : local storage constraint

18. 18 Choosing an Aging Strategy Prior Information Full No available Omniscient Algorithm Training-based Algorithm Greedy Algorithm Solve: Constraint Optimization Problem Use all data to determine optimal storage allocation. Use training dataset to determine aging parameter. resolution bias: coarsefiner finest 1 prior information

19. 19 Experimental Evaluation Implementation and Parameter Settings Present the design and implementation on Linux platform  Emstar, a Linux-based emulator/simulator for sensor networks  Query surveys on an iPAQ-based implementation Geo-spatial precipitation dataset  15 x 12 grid, 50 kilometers apart  Precipitation data from 1949 to 1994 System Parameters  = 3 epochs * 365 samples/epoch * 2 bytes/sample = 2190 bytes  c 0 : c 1 : c 2 : c 3 = 6 : 12 : 24 : 48

20. 20 Experimental Evaluation Implementation Block Diagram Construct the summaries. Allocate storage to summaries. Hierarchical storage and drill-down search. 9/7 wavelet filter

21. 21 Experimental Evaluation IPAQ Wavelet Codec y x time 3D DWT Quantization RLE Encoder Huffman Encoder Transmission over the air Huffman Decoder RLE Decoder y x time Reconstructed 3D Array Coding Decoding level i cluster head level i+1 cluster head

22. 22 Experimental Evaluation Communication Overhead Communication Rate per Level 6 12 24 48 input compression parameter The dimensions of the grid are not perfectly dyadic (power of 2).

23. 23 Experimental Evaluation Drill-down Query Performance Query Types  GlobalDailyMax  GlobalYearlyMax  LocalYearlyMean  GlobalYearlyEdge Temporal Scale Spatial Scale All Nodes Single Node Daily Yearly Not evaluated Daily Max Yearly Max Yearly Edge Local Mean

24. 24 Experimental Evaluation Drill-down Query Performance Query Error vs. Terminate Level 40 % - 50 % 0 % - 10 % GlobalYearlyEdge?

25. 25 Experimental Evaluation Aging Performance Evaluation Error Comparison between different Aging Strategies Omniscient (entire) vs. Training (first 6 years) Dataset The predicted error of the Training Scheme is within 5%.

26. 26 Experimental Evaluation Aging Performance Evaluation Aggregate results over a range of storage sizes and query types. Storage Sizes: 0 – 100 KB, four query types less than 1% worse than the optimal solution! optimal

27. 27 Experimental Evaluation Aging Performance Evaluation Comparison of Aging Strategies for GlobalYearlyMax Increasing the storage size reduce fraction error!

28. 28 Future Research Problems Irregular Node Placement Micro-climate monitoring sensor network at James Reserve How to handle irregularity?

29. 29 Future Research Problems Performance of Daily Max Query The quality does not always improve! Level 2Level 3

30. 30 Conclusion An ideal search and storage system for sensor networks.  Low communication overhead  Efficient search for a broad range of query  Long-term storage capability  DIMENSIONS Long-term storage and query processing.  Progressive aging of summaries  Load-sharing by cluster-rotation