Jigar Doshi

1
In-Network processing

• Jigar Doshi

2
What is TAG

• Tiny Aggregation for Sensor Networks
• Provides an SQL interface to query sensor
  networks with aggregates like count, max, min etc
• Sensitive to constraints of ad-hoc sensor
  networks
• Query inserted into network over an existing
  routing protocol
• Aggregation done along the reverse path

3
TAG - Motivation

• An aggregation service for Berkeley Motes
• Sensor Networks will be used by practitioners in
  variety of fields
• Civil engineers to monitor buildings during
  earthquakes
• Biologists for habitat monitoring
• All applications depend on extracting info from
  the network.
• Thus it must be a core service and easy to use.
• TAG fills this void

4
Requirements of the Routing Algorithm

• Deliver Query requests to all nodes
• Route from every node to root
• No Duplicates !
• Does it violate the end-to-end principle?
• A simple example proposed tree based routing

5
Tree Based Routing

• One root
• Any interior node sets sender as parent and sets
  its level to that of parent 1
• Rebroadcasts
• Message sent by node to its parent eventually
  reaches root
• Reselect parent after k silent epochs

Query
1
P0, L1
2
3
P1, L2
P1, L2
4
P2, L3
6
P3, L3
5
P4, L4
6
Query Model

• One Table sensors
• SELECT AVG(volume), room FROM sensorsWHERE floor
  6GROUP BY roomHAVING AVG(VOLUME) gt
  thresholdEPOCH DURATION 30s
• In generalSELECT agg(expr), attrs FROM
  sensorsWHERE selPredsGROUP BY
  attrsHAVINGhavingPredsEPOCH DURATION I
• Difference between TAG SQL Continuous Output

7
Aggregate Structure

• Standard SQL supports the basic 5
• MIN, MAX, SUM, AVERAGE, and COUNT
• TAG supports any function conforming to
• Merging function f. Merges two partial state
  records
• Initializer i Instantiates a state record for a
  single sensor value
• Evaluator e Computes the actual value of the
  aggregate from a partial state record
• Example - average
• fltS1, C1gt, ltS2, C2gt ? lt S1 S2 , C1 C2gt
• iv ? ltv,1gt
• eltS1, C1gt ? S1/C1
• TAG supports MEDIAN, HISTOGRAM and COUNT DISTINCT
  also

8
Classifying Aggregates

• Duplicate Sensitive (yes/no)
• Exemplary/Summary
• Monotonics f(a,b) e(s) gt MAX(e(s1),e(s2))
  OR e(s) lt MIN(e(s1),e(s2))
• Decides whether predicate can be applied in
  network
• Partial State
• Distributive
• Algebraic
• Holistic
• Unique
• Content Sensitive

9
Aggregate Taxonomy
10
The TAG Algorithm

• 2 Phases
• Distribution Queries are pushed down the
  network.
• Parents broadcast queries to their children
• Collection Aggregate values continuously sent
  from children to parents
• Reply from all children required before
  forwarding an aggregate value
• TDMA like partitioning
• Children must deliver records during a
  parent-specified time interval
• Parent collects all values (including its own)
  and sends the aggregate up the tree
• Result a single aggregate value per epoch
• Question Contention ?

11
Flow of partial State

• Parent reception interval must be chosen
  carefully
• All children must be able to report
• Cannot exceed end of epoch
• However we can always make the algorithm pipelined

12
Illustration Pipelined Aggregation
Epoch 3
4
SELECT COUNT() FROM sensors
1
3
Sensor
2
1 2 3 4 5
1 1 1 1 1 1
2 3 1 2 2 1
3 4 1 3 2 1
Epoch
1
13
Illustration Pipelined Aggregation
SELECT COUNT() FROM sensors
Epoch 4
5
1
3
Sensor
1 2 3 4 5
1 1 1 1 1 1
2 3 1 2 2 1
3 4 1 3 2 1
4 5 1 3 2 1
2
Epoch
1
14
Illustration Pipelined Aggregation
SELECT COUNT() FROM sensors
Epoch 5
5
1
3
Sensor
1 2 3 4 5
1 1 1 1 1 1
2 3 1 2 2 1
3 4 1 3 2 1
4 5 1 3 2 1
5 5 1 3 2 1
2
Epoch
1
15
Grouping

• Simple aggregation mechanism
• Complicated by HAVING clause
• Group eviction to solve storage problem
• Evicted tenant sent to parent

16
Simulation Environment

• Java-based simulation visualization for
  validating algorithms, collecting data.
• Coarse grained event based simulation
• Sensors arranged on a grid, radio connectivity by
  Euclidian distance
• Communication model
• Lossless All neighbors hear all messages
• Symmetric links
• No collisions, hidden terminals, etc.
• Realistic
• Number of hops related to distance but not
  proportional

17
Simulation Screenshot
18
Simulation Results

• 2500 nodes, d50
• TAG outperforms centralized approach by an order
  of magnitude in most cases
• Does equally well in the worst case
• Actual benefit depends on the topology

19
Experiment Basic TAG

• Dense Packing, Ideal Communication

20
Experiment Hypothesis Testing

• Uniform Value Distribution, Dense Packing, Ideal
  Communication

21
TAG Loss Tolerance

• Maintain List of the link signal quality to the
  neighbors and if better shift parent
• If no hello for a ?. Assume that the parent has
  moved away. Pick a new parent
• You can pick a node below you in the tree so
  child may have to reselect their parent

22
Experiment Effects of Loss
23
Experimental Results

• 16 nodes, depth 4 tree, COUNT aggregate of
  messages
• Centralized 4685
• TAG 2330
• Quality of aggregate is better with TAG
• Centralized case high loss due to contention

24
Experiment Benefit of Cache
25
Summary

• TAG is based on a declarative interface
• Uses aggregation extensively
• Makes network tasking easier for the end user
• TAG outperforms centralized approaches in most
  cases
• Aggregation again
• Relies heavily on the underlying routing layer
• Placement of query tree is constrained by the
  characteristics of routing tree

26
Critique

• Fault Tolerance too simplistic
• How high is the failed node on the tree
• Mobility
• Multiple Queries ?
• Query Optimization
• Cougar
• Congestion
• Generic ?

27

• Questions ?

28
Time Indexing in Sensor Networks

• Rong Zheng, GuangHui He, Indy Gupta, Jennifer C
  Hou, Lui Sha
• UIUC

29
What is Time Indexing

• Sensor networks collect periodic data
• Therefore lot of the queries of the data are
  temporal
• What is the average temperature last saturday
• Time Indexing refers to storage of Data based on
  the time attribute
• Data is stored on a cluster head depending on the
  time it has occurred
• Allows easy retrieval of temporal information

30
Storage of Information in a sensor network

• Can store data either
• Externally i.e Centralized or along perimeter
• Locally at the sensor node
• In-Network i.e intermediate super nodes
• Many queries temporal in nature thus info can be
  time indexed
• To provide time indexing rendezvous points are
  proposed

31
The Problem

• Time Indexed Storage maps and stores time indexed
  data x(v,t)v e V, t e (m?t, m1?t)to a
  subset of nodes in the network . A time-indexed
  query retrieves the individual or aggregated
  sensor reading during the same time
• Time Indexing Problem is concerned with efficient
  storage and retrieval of info based on the time
  attribute.

32
Requirements of the Solution

• Distributed time indexed storage
• Retrieve info based on time attribute
• So what are the requirements of any solution
  applicable to sensor nets?
• Low protocol Overhead
• Low Communication query overhead
• Energy balancing
• Solution Use Dominating Sets !

33
Dominating Sets

• For a graph G and a subset S of the vertex set
  V(G), denote by Ng(s) the set of vertices in G
  which are in S or adjacent to a vertex in S. If
  Ng(s) V(G), then S is said to be a dominating
  set (of vertices in G).
• Applied to a lot of problems like adhoc routing.
  There we need one such set. Here we need multiple
  sets
• Can be one hop or k hop. Only one hop used here
  but can be extended to k-hop.
• But finding the dominating set here is not very
  straight forward

34
(No Transcript)
35
Selecting the Nodes - Overview

• Each node has a schedule vector where 1 indicates
  that it is a leader for that cycle
• If node is a leader it uses more power. Therefore
  goal is to minimize average power consumption of
  the network
• This can be rewritten as

36

• However previous case not always optimal as the
  selected nodes will run out of power
• Thus We need load balancing, Therefore we
  introduce a new function U(x)
• The function U(x) determines behavior of
  algorithm
• Now problem can be defined as

37
Choices For U(x)

• U(x) -x. Minimizes total number of 1
  slotsProblem Dominating nodes run out of power
• U(x) -(-log(x/T))? - max min allocation tries
  to maximize time till first node runs out of
  energy
• U(x) log(1-x/T) Load balancing

38
Localized Algorithm for RP election

• The problem is NP complete
• However if we restrict to integers problem can be
  solved in polynomial time
• Do a newton like method to reach solution
  (Gradient descent too slow)
• Algorithm is localized as constraints are
  localized
• Only p and y is exchanged at every step
• We derive a heuristic which is simpler later

39
Performance
40
The integer Programming version
41
Information retrieval

• Three steps
• Inject Query
• RP to RP forwarding
• Accumulation at a point
• Overlay Construction
• Send messages with source ip and sequence number
  and the broadcast it. Thus each nodes which
  overlay stores what time.
• To find data send an overlay message to the
  person who has the data routed using the routing
  table
• Send data back once it reaches target O(V)
• Total query Messages

42
Performance Evaluation
43
Performance Evaluation
44
Future Directions

• Weighted Utility Functions
• K Domination instead of 1 domination
• Variable query Interval
• Failure of Nodes? Fault Tolerance
• Node Movement?
• Messages have to be re-routed
• Fisheye Routing

45
Localized Algorithms In Wireless Ad-Hoc Networks
Location Discovery And Sensor Exposure

• Seapahn Meguerdichian, Sasha Slijepcevic,Vahag
  Karayan1, Miodrag Potkonjak
• UCLA

46
What are Localized Algorithm

• Generic technique for solving distributed
  optimization problems in wireless networks
• Take into account topology, available energy,
  power, privacy requirements, etc.
• Obtain only needed information and use it to
  guide optimization
• Take into account problem properties

47
Localized Algorithm Components

• Data acquisition mechanisms
• Optimization mechanisms
• Search expansion rules
• Bounding conditions
• Termination rules

48
Generic Localized algorithm

• Initiate Search
• Request info from neighbours
• Loop while !terminate
• Form Partial Solution
• Decide which nodes to contact.
• Decide Whom to terminate
• Contact Nodes
• End Loop

49
Sensor Coverage

• Given
• Field A
• N sensors
• Initial and final points I and F
• How well can the field be observed
• Closest Sensor (minimum distance) only
• Multiple Sensors speed and path considered
• We have to find the Minimal Exposure Path PminE
  in A, starting in I and ending in F
• PminE is the path in A, along which the exposure
  is the smallest among all paths from I to F.

50
Sensor Exposure
Suppose S(s ,p) represents the non-negative
sensibility of sensor s to the point p. For
example
The Exposure for an object in the sensor field
during the interval t1,t2 along the path p(t)
is
51
Localized Exposure Field Partitioning

• Voronoi Partitioning
• Advantages
• One sensor per Polygon
• Node can calculate its VP by knowing only its
  immediate (Delaunay) neighbors
• Smaller VPs in high node density areas
• Drawbacks
• One sensor potentially in charge of large area
• Paths likely to be close to border edges
• How to find Delaunay neighbors?

52
Minimal Exposure Path Algorithm

• Use a grid to approximate path exposures.
• The exposure (weight) along each edge of the grid
  approximated using numerical techniques.
• Use Dijkstras Single-Source Shortest Path
  Algorithm on the weighted graph (grid) to find
  the Minimal Exposure Path.

53
Exposure Profile

• Each polygon edge has a corresponding Exposure
  Profile (EP)
• Can use different data structures to store EPs.
• EPs initialized to infinity
• Continuously updated in algorithm by keeping
  smaller values and discarding larger ones

54
Updating EP

• Node s1 updates an EP e13
• s1 sends update message to neighbor node s3
• s3 computes new minimal exposure paths and
  updates all its EPs.
• s3 sends appropriate EP update messages to
  corresponding neighbors

55
Stopping the algorithm

• Algorithm stops when
• Each EP at the search boundary is larger than the
  specified termination condition ? (parameter
  indicating bound on exposure)
• Specified by the algorithm at first
• Periodically set to exposure at destination point
  during the optimization process (broadcast)
• No more edge updates (EP)
• Guaranteed to converge since exposure is always
  increasing.

56
Localized Exposure Message Types

• Path_request Node si receives a request from an
  agent to find PminE from I to D .
• Edge_update Node si receives an update
  notification from a neighbor to continue search
  for PminE(I,D).
• Abort_update Aborting conditions notification.
• Dest_update Destination reached notification

57
Localized Exposure Caveat

• At times, minimal exposure paths may lie close to
  polygon edges
• Numerical errors and data structure limitations
  can cause cycles to develop between neighboring
  nodes
• Each believes the other has better path to
  destination
• Possible Solution
• Accept improvement at each point in an EP update
  message based on a threshold
• Discard small improvements
• Prevents infinite loops and unnecessary update
  messages
• Solution will be slightly inaccurate

58
Performance
59
Performance
60
Performance
61
Localized Minimal Exposure Path
62
Location Discovery Preliminaries

• Multi-lateration (assumed atomic)Node 0
  estimates location based on info from 1,2,3
• Di approximate distance between estimated
  location i and location x,y Ri0 represents RSSI
  based distance

1
2
0
3
63
Location Discovery

• Assumption
• All nodes know their neighbors
• All messages are received correctly
• Node gets information from other nodes and
  updates its estimate accordingly by trilateration
• Order specified by Algorithm
• Node can do this multiple times
• Objective Function used to prioritize inputs
• Nodes variance to the centre of mass measured and
  those with smaller variances preferred
• Nodes who have become references for others also
  perferred

64
Initiation

• Set bits gotFinal and Orphan
• Send to everybody

65
Information Exchange

• Listen to all neighbors
• A node which still does not have final location
  becomes orphan if less than 3 respond since
  trilateration will not work.

66
Trilateration

• A number of trilaterations done and provides a
  estimate of the node
• Calculates Objective Function
• Objective function acts as a weighing function
• Node terminates when it determines location or
  becomes an orphan

67
Performance
68
Critique

• Presents a decentralized approach to solve
  problems in sensor networks
• Analysis
• Overhead / Performance when compared to a
  centralized approach?
• Time taken To Converge?
• Mobility/ Failure of nodes?

69

• Questions ?

70
Dealing with Losses

• Neighbor monitoring
• Node will choose a new parent
• When link quality to existing parent drops
• When existing parent hasnt transmitted for some
  time
• Records can get lost due to parent switching
• But, no duplicates (1 transmission per epoch)
• Caching
• Parents cache children's partial state records
  for some number of rounds
• Memory duration must be shorter than parent
  switching interval
• Redundancy
• If a node has multiple parents, split the
  aggregate and send at both
• Doesnt work for holistic aggregates (e.g. MEDIAN)