1. Challenges in Sensor Network Query Processing Sam Madden NEST Retreat January 15, 2002

2. Outline Background Server Side Solutions Fjords, Sensor Proxies, CACQ Sensor Side Solutions Catalog Management Aggregation Future Work

3. Background: Query Processors

4. What is a Query? Declarative statement requesting a subset of data Possibly transforming or computing statistics about that data Data independent Query can apply to any data

5. What is a Query Processor? Converts declarative queries into flow of data operators, a query plan Relational Operators: Project, Select, Join ‘Scans’ read data from base relations, indices, etc. Traditional Flows: Pull based, ‘iterator model’ Higher level operators call ‘getNext()’ to extract data from lower level operators

6. Query Optimizer Given a declarative query, build the ‘best’ query plan Choose which operators to run What order to run them in Where to run them In distributed databases

7. Why Databases and Sensors? All applications depend on data processing Declarative query language over sensors attractive Application specific solutions difficult to built and deploy Want “to combine and aggregate data streaming from motes.” Sounds like a database…

8. New Problems In Sensor Databases Sensors unreliable Come on and offline, variable bandwidth Sensors push data Sensors stream data Sensors have limited memory, power, bandwidth Communication very expensive Sensors have processors Sensors very numerous

9. Components of A Sensor Database Server Side Query Parser Catalog Query Optimizer Query Executor Query Processor Sensor Side Catalog ‘Advertisements’ Query Processor Network Management

10. Outline Background Server Side Solutions Fjords, Sensor Proxies, CACQ Sensor Side Solutions Catalog Management Aggregation Future Work

11. Fjords Query Plan Abstraction to handle lack of reliability and streaming, push based data Combine push and pull in arbitrary combinations Use connectors between operators to isolate them from flow direction “Bracket Model” – Graefe ‘93

12. Fjords (Continued) Operators assume non-blocking queue interface between each other. Queues implement push vs. pull Pull from A to B : Suspend A, schedule B until it produces data. A cannot go forward until B produces data. Push from B to A : A polls, scheduler thread invokes B until it produces data. A can process other inputs while waiting for B. Supports parallelism between operators via queues, state machines, and OS (e.g. NIC buffers, DMA) in operator transparent way.

13. Fjords Example   PushPush PushPush Pull Samuel Madden, Michael J. Franklin. Fjording The Stream: An Architecture For Queries Over Streaming Sensor Data. International Conference on Data Engineering, 2002. To Appear, Feburary 2002.

14. Fjords Example   PushPush PushPush Pull Samuel Madden, Michael J. Franklin. Fjording The Stream: An Architecture For Queries Over Streaming Sensor Data. International Conference on Data Engineering, 2002. To Appear, Feburary 2002.

15. Fjords Example   PushPush PushPush Pull Samuel Madden, Michael J. Franklin. Fjording The Stream: An Architecture For Queries Over Streaming Sensor Data. International Conference on Data Engineering, 2002. To Appear, Feburary 2002.

16. Fjords Example   PushPush PushPush Pull Samuel Madden, Michael J. Franklin. Fjording The Stream: An Architecture For Queries Over Streaming Sensor Data. International Conference on Data Engineering, 2002. To Appear, Feburary 2002.

17. Fjords Example   PushPush PushPush Pull Samuel Madden, Michael J. Franklin. Fjording The Stream: An Architecture For Queries Over Streaming Sensor Data. International Conference on Data Engineering, 2002. To Appear, Feburary 2002.

18. Fjords Example   PushPush PushPush Pull Samuel Madden, Michael J. Franklin. Fjording The Stream: An Architecture For Queries Over Streaming Sensor Data. International Conference on Data Engineering, 2002. To Appear, Feburary 2002.

19. Fjords Example   PushPush PushPush Pull Samuel Madden, Michael J. Franklin. Fjording The Stream: An Architecture For Queries Over Streaming Sensor Data. International Conference on Data Engineering, 2002. To Appear, Feburary 2002.

20. Fjords Example   PushPush PushPush Pull Samuel Madden, Michael J. Franklin. Fjording The Stream: An Architecture For Queries Over Streaming Sensor Data. International Conference on Data Engineering, 2002. To Appear, Feburary 2002.

21. Fjords Example   PushPush PushPush Pull Samuel Madden, Michael J. Franklin. Fjording The Stream: An Architecture For Queries Over Streaming Sensor Data. International Conference on Data Engineering, 2002. To Appear, Feburary 2002.

22. Fjords Applications Combine traffic streams with web-based accident reports Francis Li, Sam Madden, Megan Thomas. Traffic Visualization. http://www.cs.berkeley.edu/~mct/infovis/project/traffic.html

23. Operators for Streaming Data Need special operators for dealing with streams (See P. Seshadri, et al. The design and implementation of a sequence database systems..VLDB ’96 ) In particular, streams can’t be joined or sorted in the traditional sense Solution: Use windows – e.g. “Zipper Join”

24. Sensor Proxy Energy-sensitive database operator Buffer sensor tuples and route to multiple user queries to hide query load from sensors Push aggregation operators into sensors to reduce communications load Dynamically adjust sample rate based on user demand Push results into Fjords so that other operators don’t block waiting on slow or dead sensors

25. Some Results Pushing predicates into sensors can vastly reduce costs: Atmel Simulator 100 samples / sec 5 vehicles / sec 7x power savings

26. CACQ Expect hundreds to thousands of queries over same sensor sources Continuously Adaptive Continuous Queries Continuous Queries: Long running queries which combine selections and joins to improve efficiency ( See Chen, NiagaraCQ, SIGMOD 2000 )   Stocks. symbol = ‘MSFT’ Stocks. symbol = ‘APPL’ Query 2Query 1 Stock Quotes ‘MSFT’ ‘APPL’ Stock Quotes 

27. CACQ (Cont.) Continuous Adaptivity From Eddies Route tuples differently, depending on selectvity and cost estimates of operators static dataflow eddy Diagrams Courtesy Joe Hellerstein

28. CACQ (cont.) Combining CA with CQ is a win: CQ increases number of simultaneous queries Adaptivity well suited to long running queries Eddies allow us to avoid ugly query- optimization phase in traditional CQ Eddies + Streams == few copies, unlike traditional CQ

29. CACQ (cont) Look for a paper in SIGMOD 2002 (fingers crossed!)

30. Outline Background Server Side Solutions Fjords, Sensor Proxies, CACQ Sensor Side Solutions Catalog Management Aggregation Future Work

31. Sensor Side Solutions CACQ + Fjords provides interface + performance on QP, but sensors still need help: Locate / identify sensors Reduce power consumption Take advantage of processors? Improve responsiveness

32. Cataloging Sensors To query sensors, need a way to locate, identify properties, extract values Goal: Drop a bunch of sensors around the DBMS, allow them to be queried without manual effort Idea: Add a layer to each sensor which advertises its capabilities

33. Catalog (Continued) #temperature sensor field { name : "temp" #optional type : int units : celsius min : -20 max : 100 bits : 8 sample_cost : 10.0 J #optional -- for use in costing sample_time : 10.0 ms #optional -- for use in costing input : adc2 #optional : read from adc channel 1 sends : ondemand accessorEvent : GET_TEMPERATURE_DATA responseEvent : TEMPERATURE_DATA_READY } Compiled in 27 bytes of memory Layer to register with Query Processor Can be “push” or “pull”

34. Aggregating Over Sensors Sensor Proxy combines user queries, pushes down aggregates Goal: Save energy, increase efficiency Idea: Take advantage of the routing hierarchy

35. Why bother with aggregation Individual sensor readings are of limited use Interest in higher level properties, e.g. what vehicles drove through, what is the spread of temperatures in the building We have a processor & network on board, lets use it We cannot survive without aggregation Delivering a message to all nodes much easier than delivering a message from each node to a central point Delivering a large amount of data from every node harder still, vide connectivity experiment Forwarding raw information too expensive Scarce energy Scarce bandwidth Multihop performance penalty

36. Aggregation challenges Inherently unreliable environment, certain information unavailable or expensive to obtain how many nodes are present? how many nodes are supposed to respond? what is the error distribution (in particular, what about malicious nodes?) Trying to build an infrastructure to remove all uncertainty from the application may not be feasible – do we want to build distributed transactions? Information trickles in one message at a time Never have a complete and up-to-date information about the neighborhood What type of information should we expect from aggregation Streams Robust estimates

37. What does it mean to aggregate (The DB Perspective) General purpose solution: apply standard aggregation operators like COUNT, MIN, MAX, AVERAGE, and SUM to any set of sensors. Existing solutions are application specific In sensors, operators may be arbitrary signal processing functions By assuming a standard interface, many optimizations are possible Example: TopN queries via hypothesis testing Provide grouping semantics: e.g. ‘select avg(temp) group by trunc(light/10)’ In sensor networks, groups may be random samples t1t1 t2t2 t3t3 t4t4 t5t5 t6t6 t7t7 t8t8 t9t9

38. Outline Background Server Side Solutions Fjords, Sensor Proxies, CACQ Sensor Side Solutions Catalog Management Aggregation Future Work

39. DBMS Side Efficient Catalog Management Moving Object Databases Query Optimization Techniques Sensor Side Efficient Grouping Joins over Network Topology Non Standard Aggregate Functions Somewhere In Between Histograms and other Correlations Sampling and Compression for Streams Real Query Language / API Demonstration Apps (SIGMOD Demo)

40. Questions?

41. 2 1 3 45 Scenario: Count

42. 2 1 3 4 5 Goal: Count the number of nodes in the network. Number of children is unknown. 12345 ----- ----- ----- ----- ----- ----- ----- Sensor # Time

43. 2 13 Scenario: Count Goal: Count the number of nodes in the network. Number of children is unknown. 12345 1---- ----- ----- ----- ----- ----- ----- Sensor # Time

44. 2 13 4 Scenario: Count Goal: Count the number of nodes in the network. Number of children is unknown. 12345 1---- 111-- 1 + 211-- ----- ----- ----- ----- Sensor # Time

45. 2 13 45 Scenario: Count Goal: Count the number of nodes in the network. Number of children is unknown. 12345 1---- 111-- 1 + 2111- 1 + ½ 1- ----- ----- ----- Sensor # Time

46. 2 13 45 Scenario: Count Goal: Count the number of nodes in the network. Number of children is unknown. 12345 1---- 111-- 1 + 2111- 1 + ½ 11 1+31+ ½ 1+11 ----- ----- Sensor # Time

47. 2 1 3 4 5 Scenario: Count Goal: Count the number of nodes in the network. Number of children is unknown. 12345 1---- 111-- 1 + 2111- 1 + ½ 11 1+31+ ½ 1+11 1+31+2/ 2 1+11 ----- Sensor # Time

48. 2 1 3 4 5 Scenario: Count Goal: Count the number of nodes in the network. Number of children is unknown. 12345 1---- 111-- 1 + 2111- 1 + ½ 11 1+31+ ½ 1+11 1+31+2/ 2 1+11 1+41+2/ 2 1+11 Sensor # Time

49. Counting Lessons Take advantage of redundancy to improve accuracy (reply to all parents, not just one) Use broadcast to reduce number of messages Result is a stream of values: much more robust to failures, movement, or collision than a single value.

50. Aggregation in network programming Network programming problem Reliable delivery of a large number of messages to all nodes in range, while exploiting the broadcast nature of the medium Basic setup Broadcast a known number of idempotent program fragments Each node keeps a bitmap of fragments received (1=packet received) Two stages of the problem: single hop, and multihop Solutions Single hop, dense cell Broadcasting the program – trivial, the central node broadcasts Feedback from nodes – broadcast a request from the central node: Is anyone missing packets in this packet range? Convergence: no replies to the request

51. Aggregation in multihop network programming Broadcasting the program – use flooding Remember the last 8 packets forwarded, use that cache to decide whether to forward or not Feedback from nodes Distribute requests for feedback using the flooding After some delay, respond if any packets are missing locally Responses from children: AND with the local bitmap, store the result locally, forward the request Suboptimal because there is no local fixups Convergence No replies to the request

52. Aggregation over streams Inherent uncertainty of the system Can nodes communicate, do they have enough power, have they moved? computing a complete single answer can be very expensive, and may not be possible Partial estimates have their own value Aggregation over streams Values reflect the current best estimates Self stabilizing: in the absence of changes converges to a desired value within N steps

53. Identifying Groups Need a way to identify groups Idea: set of membership criteria pushed down Nodes determine their membership set based on those criteria Nodes can be in multiple but not unlimited groups E.g. “Group 1 : 0 <= t < 10, Group 2 : 10 <= t < 20, …” Need a way to evaluate aggregation predicates by group May want to allow grouping and aggregation predicates to be expressed together to take advantage of broadcast effects

54. Local Query Rewrite Intermediate nodes may determine that its faster to evaluate an aggregate by asking children a different question. Example 1: MAX(t). Once we have a guess T for MAX, ask children to report iff t > T, rather than asking all children to compute a local maximum. Example 2: Network programming. Rather than asking nodes what packets they have, ask them to report iff packets missing. Is this a general technique? Maybe: Inform child of guess at aggregate, ask it to refute. Works for average (within error bound), not count.

55. Wins and pitfalls of aggregation Aggregation over natural network topology Aggregation over an arbitrary subset of the network may be a loss Really dense cells Aggregation does not help with the starvation problem Use the message suppression via query rewrite technique Still beneficial in a multihop scenario

56. Advanced Aggregation Tricks Break the Network Protocol Boundary Use analog reading from channel over time to determine aggregates. Simple example: Time Sum Reading = 11 = 110100 Reading = 21 = 101010 Reading = 32 = 2 + 2 + 4 + 8 + 16