1. TSAR: A Two Tier Sensor Storage Architecture Using Interval Skip Graphs Peter Desnoyers, Deepak Ganesan, and Prashant Shenoy Department of Computer Science University of Massachusetts The 3rd ACM Conference on Embedded Networked Sensor Systems, SenSys 2005 Speaker: Shin-Wei Ho

2. Outline Introduction Design Considerations Data Structures Data Storage and Summarization Experimental Evaluation Conclusion

3. Introduction Typical sensor applications need access to both Live sensor data Monitoring Surveillance Past sensor data Mining of sensor logs to detect unusual patterns Analysis of historical trends

4. Introduction Archival storage of past sensor data requires a storage system. The key attributes are Where the data is stored How the application can access this data in an energy-efficient manner.

5. Introduction In the simplest, sensors stream data or events to a server for long-term archival storage. Centralized Reads are efficient and cheap Writes are expensive

6. Introduction An alternate approach is to have each sensor store data or events locally. All writes are local and incur no communication overheads. Flooding.

7. Introduction The geographic hash table (GHT) Each data item has a key associated with it. A distributed or geographic hash table is used to map keys to nodes.

8. Introduction Most of these approaches assume a flat, homogeneous architecture in which every sensor node is energy-constrained.

9. Introduction Any sensor storage system should also carefully exploit current technology trends. Moore ’ s Law It will soon be feasible to equip each sensor with 1 GB of flash storage for a few tens of dollars.

10. Introduction A typical sensor network deployment will contain multiple geographically distributed proxies. Building monitoring application Camera-based Sensing Application One sensor proxy might be placed per floor or hallway.

11. Introduction Goal Design a storage system Exploits the relative abundance of resources at proxies to mask the scarcity of resources at the sensors. Focus on applications that require access to past sensor data.

12. Design Considerations -- System Model Low power sensors with a significant amount of flash memory (e.g., 1GB). Power-rich sensor proxies

13. Design Considerations -- Usage Models Queries specify time interval value range spatial region E.g. “ find all instances where the temperature exceeded 100 o F at location R during the month of August. ”

14. Design Considerations -- Design Principles Principle 1: Store locally, access globally A sensor storage system should leverage the flash memory on sensors. Efficient mechanisms for retrieval.

15. Design Considerations -- Design Principles Principle 2: Distinguish data from metadata Data must be identified so that it may be retrieved by the application without exhaustive search. associate metadata with each data record leverage the presence of resource-rich proxies to index metadata for resource-constrained sensors.

16. Design Considerations -- Design Principles Principle 3: Provide data-centric query support The authors expect that applications will be best served by a query interface which allows them to locate data by value or attribute (e.g. location and time).

17. Design Considerations -- System Design The key features of the system design are as follows: Writes occur at sensor nodes consist of both opaque data as well as application-specific metadata. Each sensor periodically sends a summary of reported metadata to a nearby proxy.

18. Design Considerations -- System Design Summary of metadata The sensor ID. The interval (t1, t2) over which the summary was generated. A handle identifying the corresponding data record (e.g. its location in flash memory). A coarse-grain representation of the metadata associated with the record.

19. Design Considerations -- System Design The proxy uses the summary to construct an index. Applications see a unified view of distributed data, and can query the index at any proxy to get access to data stored at any sensor.

20. Design Considerations -- System Design The TSAR summarization mechanism guarantees that index lookups will never yield false negatives. I.e. It will never miss summaries which include the value being searched for. However, index lookups may yield false positives, wasting network resources.

21. Design Considerations -- System Design The more coarse-grained the summary lower the update overhead greater the fraction of false positives TSAR employs a novel adaptive technique dynamically varies the granularity of sensor summaries to balance the metadata overhead the overhead of false positives.

22. Data Structures Interval skip graphs The goal is to index all intervals such that the set that overlaps a query interval can be located efficiently. Skip graph Interval Tree

23. Data Structures -- Skip Graph Overview

24. Properties of a skip graph: Ordered index In-place indexing Log n height Redundancy and resiliency

25. Data Structures -- Interval Skip Graph Extends skip graphs to store intervals [low i, high i] Allows efficient searches for all intervals covering a value v.

26. Data Structures -- Interval Skip Graph

27. Data Structures -- Sparse Interval Skip Graph The cost in network messages of an operation may be reduced. most structure traversals occur locally on a single node.

28. Data Structures -- Sparse Interval Skip Graph Skip graphs have a full search structure rooted at each element. However, in order to provide load distribution and failure resilience it is only necessary to provide a full search structure for each system. proxies << data elements

29. Data Structures -- Sparse Interval Skip Graph

30. Data Storage and Summarization Local Storage at Sensors Each sensor node in TSAR maintains an archival store of sensor data. A temporally ordered store is suitable.

31. Data Storage and Summarization

32. There might be a need to overwrite older data in cache. High data rate applications multi-resolution storage of data

33. Data Storage and Summarization -- Adaptive Summarization There is a trade-off between the energy used in sending summaries. frequency resolution

34. Data Storage and Summarization -- Adaptive Summarization If most queries result in true hits the sensor determines that the summary can be coarsened further If many queries result in false hits the sensor makes the granularity of each summary finer

35. Experimental Evaluation Proxy tier: Crossbow Stargate nodes 400MHz Intel XScale processor with 64MB RAM. Radio: Cisco Aironet 340-based 802.11b(inter- proxy) radio and a hostmote bridge to the Mica2 sensor nodes using the EmStar transceiver. Sensor tier: Crossbow Mica2s / Mica2dots 915MHz CC1000 radio, a BMAC protocol stack, a 4 Mb on-board flash memory.

36. Experimental Evaluation

39. Conclusion This paper presents TSAR, a novel two-tier storage architecture for sensor networks. Multi-Tier Sparse Interval Skip Graph Our experimental evaluation of TSAR demonstrated the benefits and feasibility.

40. Thank you !

41. Experimental Evaluation

42. Data Structures -- Skip Graph Overview Maintaining Skip Lists To insert an element we ascend levels starting at 0. randomly choosing one of the two possible chains at each level, an stopping when we reach an empty chain.