Energy-Efficient Data Management for Sensor Networks

Transcription

1 Energy-Efficient Data Management for Sensor Networks Al Demers, Cornell University Johannes Gehrke, Cornell University Rajmohan Rajaraman, Northeastern University Niki Trigoni, Cornell University Yong Yao, Cornell University The Cougar Project

2 Outline 1. Our Model 1.1. Sensor Network 1.2. Data and Queries 1.3. In-network aggregation 2. Energy-efficient data management 2.1. View Selection 2.2. Wave Scheduling 2.3. Multi-Tree Scheduling

3 Sensor Network Small stationary sensor nodes connected through a multi-hop wireless network Communication: Limited bandwidth, variable latency, packet drops Computation: Limited processing power and memory sizes Power: Limited supply of energy B A D E C F

4 Sensor Network - Scaling Properties Near-neighbor wireless communication implies diameter at least sqrt(n) Better scaling requires hierarchy (with some more capable nodes?) Reliable self-organization in such a system? Partial solutions that keep per-node state small?

5 The DB View of Sensor Networks Traditional Procedural addressing of individual sensor nodes; user specifies how task executes; data is processed centrally. DB Approach Declarative querying and tasking; user isolated from how the network works ; in-network distributed processing. SensId Loc Time Type Value 1 (2,5) 3 temperature 60 1 (2,5) 6 pressure 62 SensId Loc Time Type Value 2 (4,2) 3 light 55 2 (4,2) 5 pressure 30 SensId Loc Time Type Value 3 (3,1) 3 humidity 70 User IF

6 Example Queries Ø Ø Snapshot queries: Ø What is the concentration of chemical X in the northeast quadrant? SELECT AVG(R.concentration) FROM Sensordata R WHERE R.loc in (50,50,100,100) Long-running queries Ø Notify me over the next hour whenever the concentration of chemical X in an area is higher than my security threshold. SELECT R.area, AVG(R.concentration) FROM Sensordata R WHERE R.loc in rectangle GROUP BY R.area HAVING AVG(R.concentration)>T DURATION (now,now+3600) EVERY 10

7 How to Execute Aggregate Queries? Two choices: Ø Centralized processing ü In-network processing 5 Why in-network processing? Ø Sensor network is power (and bandwidth) constrained Ø Local computation is much cheaper than communication

8 Outline 1. Our Model 1.1. Sensor Network 1.2. Data and Queries 1.3. In-network aggregation 2. Energy-efficient data management 2.1. View Selection 2.2. Wave Scheduling 2.3. Multi-Tree Scheduling

9 View Selection: Hybrid Pull-Push Model Push Model Pull Model Hybrid Model View Node

10 View Selection: Problem Definition Time is divided into periods. Given: Ø Aggregate Query Workload: QW={<Q1,p1>, } (p i is the probability of asking Q i at the end of the period) Ø Sensor Updates: DW={<s1,u1>, } (u j is the probability that s j has an update at the end of the period) Ø Tree that connects sensors to the query point Ø Cost of communicating b bits on a link is: α+β*b Compute locations for materialized views (sub-aggregates) to minimize total communication cost

11 View Selection: Example 1 Q a b p : probability of Q q : query message cost r : result message cost edge model push pull a-b r q+p r pull if q < (1-p) r a push q p r otherwise r a b b

12 View Selection: Example 2 Q1 a b c d e Q2 Q1: c+d Q2: d+e pr(q1) = pr(q2) = p q : query message cost r : result message cost (q<r) edge model push pull Pull if a-b b-c b-d b-e 2r r r r q+2 p r q+p r q+p (2-p) r q+p r q < 2 (1-p) r q < (1-p) r q < (1-p) 2 r q < (1-p) r

13 View Selection: Example 2 pr(q1) = p pr(q2) = p a b a b a b c d e c d e c d e a-b b-c b-d b-e Pull if q < 2 (1-p) r q < (1-p) r q < (1-p) 2 r q < (1-p) r q<(1-p) 2 r T T T T (1-p) 2 r <q<(1-p)r T T F T (1-p)r<q<2(1-p)r T F F F

14 View Selection: Example 3 a b The aggregate function matters (e.g. MIN vs SUM) c d e Q1: min(c,d) q1:sum(c,d) Q2: min(d,e) q2:sum(d,e) Q3: min(c,d,e) q3:sum(c,d,e) Result messages from b to a 2 messages 3 messages

15 View Selection: Complexity NP-complete problem (Set Basis Problem) Dynamic programming algorithms (exponential worst-case complexity) Approximation algorithms for four problem instances 1. Regular Queries and Sensor Updates 2. Irregular Queries, Regular Sensor Updates 3. Regular Queries, Irregular Sensor Updates 4. Irregular Queries and Sensor Updates

16 Outline 1. Our Model 1.1. Sensor Network 1.2. Data and Queries 1.3. In-network aggregation 2. Energy-efficient data management 2.1. View Selection 2.2. Wave Scheduling 2.3. Multi-Tree Scheduling

17 Scheduling: Assumptions Nodes are location-aware Neighboring nodes in the network are synchronized (GPS, distributed time synchronization algorithms) Recent studies about radio energy consumption idle : receive : transmit We have a few gateway nodes (destination nodes)

18 Scheduling: Goal Goal => Save energy Ideas: Minimize collisions at the MAC layer Manage the radio in a power-efficient manner Select energy-efficient routes for message delivery

19 Scheduling: TAG approach (1 tree) a b d c e Sending Receiving Idle f g TAG trees do not avoid interference.

20 Scheduling: Multiple trees If scheduled consecutively delay grows linearly with number of destinations If scheduled concurrently power grows more than linearly in the number of destinations (contention) TAG scheduling is good for only a small number of trees

21 Scheduling: GAF Topology Control Periodically re-elect leader in each cell

22 Scheduling: Trees embedded on a grid

23 Our approach: Wave Scheduling

24 Our approach: Wave Scheduling EAST WAVE Sending Receiving Idle

25 Our approach: Wave Scheduling EAST WAVE Sending Receiving Idle

26 Our approach: Wave Scheduling EAST WAVE Sending Receiving Idle

27 Our approach: Wave Scheduling EAST WAVE Sending Receiving Idle

28 Our approach: Wave Scheduling EAST WAVE Sending Receiving Idle

29 Wave Scheduling: Properties Repeats in all 4 directions (North, East, South, West) Non-interfering edges are scheduled concurrently Simple or Interleaved General-purpose schedule. Every edge of the network is activated exactly once per period Can think of it as a real wave

30 Wave Scheduling: Intuition Sending Receiving Idle

31 Wave Scheduling: Intuition Sending Receiving Idle

32 Wave Scheduling: Intuition Sending Receiving Idle

33 Wave Scheduling: Intuition Sending Receiving Idle

34 Wave Scheduling: Routing Schedules have handed-ness e.g. The (N,E,S,W) schedule favors paths that take right-hand turns Given a schedule, we can select Min-delay Routes or Min-Energy Routes N E W S What about holes in the grid?

35 Wave Scheduling: Routing DESTINATION Long right turn path has smaller latency than short left turn path SOURCE =>

36 Wave Scheduling: Routing Anomalies d c Solution: Proactive distancevector routing scheme SOURCE a b DESTINATION Get min-feasible-hops with min delay or min-feasible-delay with min hops Small routing tables 1 bit per edge per destination node =>

37 Wave Scheduling: Routing Anomalies A B Problem: Infinite Loop (because of heavy traffic) SOURCE DESTINATION C Solution: Decide the next hop for a message at the time it is received by a node

38 Outline 1. Our Model 1.1. Sensor Network 1.2. Data and Queries 1.3. In-network aggregation 2. Energy-efficient data management 2.1. View Selection 2.2. Wave Scheduling 2.3. Multi-Tree Scheduling

39 Multi-Tree Scheduling If scheduled consecutively delay grows linearly with number of destinations If scheduled concurrently energy grows more than linearly in the number of destinations (contention) So look for a combined schedule activating each edge once!

40 Combined Schedule? The combined dissemination graph may not be a tree... It may not even be acyclic!

41 Combined Schedule? A difficult problem... Optimal schedule for combined graph is NP-complete (reduction from FVS) If we consider edge-edge interference (collisions) the problem gets even harder Related to maximum common supersequence problem, but lower bounds do not carry over Closest: the Loading Time Problem (LTP) -- heuristics can be adapted

42 Heuristics... Naive Algorithm = schedule trees in sequence Easy examples make the LTP heuristics as bad as the naive one We have clever heuristics! But we have very clever counterexamples None of our heuristics beats naive in worst case Are the clever heuristics better on average?

43 Evaluating on Real Queries... What is the query workload? Spatial locality? Connectivity? Given some representative queries, how can we compare our heuristics against the optimal? Yong has a few nice theorems that may make this just tractable for realistically large queries

44 Scheduling: Experimental Setup NS-2 Network Simulator Mac layer: IEEE communication range = 250m interference range = 550m transmit:receive:idle = 1.6:1.2:1.0 grid cell has size (100m * 100m) grid: 100 (10 by 10) cells We assume GAF, so we have one node per grid cell Experiments lasted 5000 seconds simulation time The source and destination of a message are randomly selected.

45 Scheduling: Experiments Wave ENERGY: Varying the number of destination nodes Average Energy Consumption Energy (Joules) Messages

46 Scheduling: Experiments Tree ENERGY: Varying the number of destination nodes Average Energy Consumption Energy (Joules) Messages

47 Scheduling: Experiments ENERGY: Energy-based vs Delay-based Routing Average Energy Consumption Energy (Joules) Messages

48 Conclusions Saving energy is a significant consideration in sensor networks Two approaches : 1. We can save energy by reducing the number of messages. Proactively push data to carefully selected storage points. 2. We can save energy by turning off radios whenever possible Coordinate sensor node radios with (global) wave schedules or (global) multi-tree schedules.

49 Future Work Interaction between view selection and dissemination graph selection Efficient schedules given specific message generation patterns Improved fault-tolerance

50 Thank you!

51 Scheduling: Experiments DELAY: Energy-based vs Delay-based Routing Average Message Delay Delay (secs) Messages