Research on Ad Hoc Networking at UC Santa Cruzs iNRG'

1
Research on Ad Hoc Networking at UC Santa Cruzs
i-NRG.

• Katia Obraczka
• Department of Computer Engineering
• UC Santa Cruz
• katia_at_cse.ucsc.edu
• www.cse.ucsc.edu/katia

2
Whats i-NRG?

• inter-Networking Research Group.
• Research in design, evaluation, and
  implementation of network protocols consisting of
  wired as well as wireless networks, from I-NRGs
  Web page.I

3
Ad-Hoc Networking Research

• Multi-hop, wireless ad hoc network (MANET)
  operates without any fixed infrastructure.
• Attractive choice for scenarios where fixed
  network infrastructure is non-existent or
  unusable.
• Example applications search and rescue, disaster
  recovery, digital battlefield, remote sensing
  etc.,

4
Challenges raised by MANETs

• Time-varying nature of the wireless channel
  mobility, physical environment (e.g.,
  interference, fading and shadowing, etc.).
• Heterogeneous device capabilities and varying
  requirements from applications.
• Energy efficiency.

5
Ongoing Research at i-NRG

• MAC.
• Multicast routing.
• Interconnection protocol.
• Data propagation.
• Reliable multicast transport.
• Energy models.
• Sensor networks for habitat monitoring.

6
Medium Access Control
7
Why work on MAC?

• Medium access control (MAC) protocols are
  critical for the overall performance in MANETs.
• Strict layered approach for network protocol
  stack is no longer applicable.
• Cross layer optimization and adaptiveness to the
  operating environment are desired.

8
Importance of Cross-Layer Optimization
9
Cross-Layer Optimization in MANETS

• Group communications is an efficient means for
  supporting group-oriented services like data
  dissemination, teleconferencing etc.,
• IEEE 802.11 does not have support for
  multicasting.
• Heterogeneous loss characteristics pose a
  significant challenge on the design of reliable
  multicast transport protocol without support from
  underlying layers.

10
Why support from MAC Layer?

• Unlike wired networks, losses are due to various
  factors in MANETs.
• Initiating congestion control for losses that are
  due to transmission errors can decrease the
  throughput.
• Mechanisms to differentiate losses at higher
  layers need support from the MAC layer that
  closely interacts with the wireless medium.

11
Reliable, Adaptive, Congestion-Controlled
Multicast Transport Protocol ReACT

• ReACT classifies packet losses into
• Congestion losses (buffer overflow).
• Non-congestion losses (transmission errors etc.,)
• Congestion is detected using the MAC layer queue
  size (above a certain threshold is marked as
  congested).

12
Effect of Congestion
13
Summary

• Significant improvement in goodput with
  Cross-layer optimizations.
• Lower control packet overhead.
• Better reliable delivery ratio.

Venkatesh Rajendran, Katia Obraczka, Yunjung Yi,
Sung-Ju Lee, Ken Tang and Mario Gerla, "Combining
Source- and Localized Recovery to Achieve
Reliable Multicast in Multi-Hop Ad Hoc Networks,"
Proceedings of the Networking' 04, May 2004.
Exploring the Design Space of Reliable
Multicast Communications in Ad Hoc Networks,
Journal Submission Under Prepartion.
14
Future Research Directions

• Develop an efficient MAC protocol with support
  for group communications and cross-layer
  optimization.
• Study the impact of multicasting support at MAC
  in higher layer performance.

15
Multicast MAC Related Work

• Kuri et al Reliable medium access protocol for
  multi-access wireless LANs.
• Tang et al Broadcast Medium Window (BMW).
• Sun et al - Batch Mode Multicast MAC.
• Chaporkar et al Adaptive strategy for
  maximizing throughput in MAC layer wireless
  multicast.

16
Our approach

• Extend collision avoidance approach (RTS/CTS
  handshake) to multicasting.
• Have a subset of receivers that entirely cover
  the hidden nodes for the multicast group, reply
  with CTS.
• Design a policy to decide on when to transmit
  data.

Rx3
Rx1
C
Tx
E
Rx3
D
17
Energy-Aware Channel Access Protocols
18
Energy-Aware Channel Access

• Sensor networks are a special class of multi-hop
  wireless networks.

Ad-hoc deployment. Self-configuring. Unattended.
Battery powered.
19
Motivation

• Energy conservation is critical in extending the
  life-time of the network.
• Radio is the major source of energy consumption.
• Low-power sleep mode.

20
Sources of Energy Inefficiency
1. Hidden Terminal Collisions
21
Sources of Energy Inefficiency
22
Sources of Energy Inefficiency
Transmission to sleeping nodes leads to increased
retransmissions.
23
Related Work

• Singh et al PAMAS.
• Wei et al S-MAC.
• Sohrabi et alTDMA-based self organizing protocol
  for sensor networks.
• Chunlong et al Low-power sensor MAC.
• Bao et al Node Activation Multiple Access
  (NAMA).

24
Achieving Energy Efficiency

• When a node is neither transmitting or receiving,
  switch to low-power sleep mode.
• Prevent collisions and retransmissions.
• Need to know Tx, Rx and when transmission event
  occurs.

25
Neighborhood-aware contention resolution
(NCR)Bao et al., Mobicom00

• Each node maintains two-hop neighbor information.
• Based on the time slot ID and node ID, node
  priorities are calculated using a random hash
  function.
• A node with the highest two-hop priority is
  selected as the transmitter for the particular
  time slot.

26
NCR - Example
NCR does not elect receivers and hence, no
support for radio-mode control.
27
Dynamic Energy-Aware Node Activation (DEANA)

• Time slotted using dynamic scheduling for channel
  access.
• Elect transmitters using NCR.
• Announce receiver(s) using small control packets

28
Radio Mode Control

• During the random access period all nodes must
  have their radios on.
• During the scheduled access period
• All nodes must have their radios on during the
  control slot.
• During the data slot only nodes participating in
  the data exchange should have their radios on.

29
DEANA Performance

• Developed an analytical model to quantify energy
  savings.
• Significant energy-savings (up to 95) could be
  achieved.

V. Rajendran, J.J. Garcia-Luna-Aceves, and K.
Obraczka, "An Energy-Efficient Channel Access
Scheduling for Sensor Networks," Proc. The Fifth
International Symposium on Wireless Personal
Multimedia Communication (WPMC), October 27--30,
2002, Honolulu, HI.
30
DEANA Limitations

• Involves frequent radio mode switching.
• Overhead due to control slots.
• Transient power consumption might degrade energy
  savings.SolutionDistributed, Traffic-adaptive
  Channel Access Scheduling.

31
TRaffic-Adaptive Medium Access (TRAMA) Goals

• Establish transmission schedules in a way that
• is self adaptive to changes in traffic, node
  state, or connectivity.
• prolongs the battery life of each node.
• is robust to wireless losses.

32
TRAMA - Overview

• Single, time-slotted channel access.
• Transmission scheduling based on two-hop
  neighborhood information and one-hop traffic
  information.
• Random access period
• Used for signaling synchronization and updating
  two-hop neighbor information.
• Scheduled access period
• Used for contention free data exchange between
  nodes.
• Supports unicast, multicast and broadcast
  communication.

33
TRAMA Components

• Neighbor Protocol (NP).
• Gather 2-hop neighborhood information.
• Schedule Exchange Protocol (SEP).
• Gather 1-hop traffic information.
• Adaptive Election Algorithm (AEA).
• Elect transmitter, receiver and stand-by nodes
  for each transmission slot.
• Remove nodes without traffic from election.

34
TRAMA Performance

• Performance analysis by extensive simulation
  using QualNet.
• Comparison against both contention-based
  (IEEE802.11, CSMA and S-MAC) and scheduling-based
  (NAMA) protocols.
• Analytical model for quantifying delay
  performance.

35
Simulation Results
All nodes generate synthetic broadcast traffic
using Poisson arrivals. 50 nodes, 500x500
area. 512 byte data. Average node density 6
Delivery Ratio
36
Energy Savings
Percentage Sleep Time
Average Length of sleep interval
37
Queueing Delay
Synthetic Broadcast Traffic
Delay measured only for delivered packets at the
MAC queue. Scheduling-based protocols deliver
more packets than contention-based. If we
account the delay involved in recovering from
losses at the higher layer, average delay will
increase for contention-based protocols.
38
Analytical Results
100 nodes, grid topology. 650x650 m square
area. Radio Range 104m Contender size 25
(for middle nodes)
Queuing Delay (analytical vs simulation)
39
TRAMA Summary

• Significant improvement in delivery ratio in all
  scenarios when compared to contention-based
  protocols.
• Significant energy savings compared to S-MAC
  (which incurs more switching).
• Acceptable latency and traffic adaptive.

V. Rajendran, K. Obraczka, and J.J.
Garcia-Luna-Aceves, "Energy-Efficient,
Collision-Free Medium Access Control for Wireless
Sensor Networks," Proc. ACM SenSys 03, Los
Angeles, California, 5-7 November 2003. V.
Rajendran, K. Obraczka, and J.J.
Garcia-Luna-Aceves, "Energy-Efficient,
Collision-Free Medium Access Control for Wireless
Sensor Networks", to appear in ACM/Kluwer
Wireless Networks (WINET).
40
TRAMA Limitations

• Complex election algorithm and data structure.
• Overhead due to explicit schedule propagation.
• Higher queueing delay.

41
Flow-aware Medium Access Framework

• Avoid explicit schedule propagation.
• Take advantage of application.
• Simple election algorithm to suit systems with
  low memory and processing power (e.g., 4KB ROM in
  Motes).
• Incorporate time-synchronization, flow discovery
  and neighbor discovery during random-access
  period.

42
Flow Information

• Flow information characterizes application-specifi
  c traffic patterns.
• Flows can be unicast, multicast or broadcast.
• Characterized by source, destination, duration
  and rate.

43
Example Data Gathering Application
Fc
Fd
Fb
Sink
Fe
44
Flow Discovery Mechanism

• Can be combined with neighbor discovery during
  random-access period.
• Mechanism is adapted based on the application.
• for data gathering application, flow discovery is
  essentially establishing the data forwarding tree.

45
Example
Sink initiates neighbor discovery, flow discovery
and time synchronization. Broadcasts periodic
SYNC packets. Potential children reply with
SYNC_REQ. Source reinforces with another SYNC
packet. Once associated with a parent, nodes
start sending periodic SYNC broadcasts.
SYNC
SYNC_REQ
Sink
46
Election Process

• Weighted election to incorporate traffic
  adaptivity.
• Nodes are assigned weights based on their
  incoming and outgoing flows.
• Highest priority 2-hop node is elected as the
  transmitter.
• A node listens if any of its children has the
  highest 1-hop priority.
• Can switch to sleep mode if no transmission is
  started.

wc1
wc1
wa3
Sink ws0
47
Simulation Results (16 nodes, 500x500 area,
CC1000 radio, grid topology, edge sink)
Queueing Delay
Delivery Ratio
48
Energy Savings
49
Test-bed Demonstration

• Mica2 Motes (MPR400) with Chipcon CC1000 radio
  operating at 915 Mhz.
• Data rate 19.2 Kbps.
• Range 100m.
• Modulation method FSK.
• Software Platform TinyOS.
• Compared with S-MAC.

50
Experimental Results
Packet Generation Period
Sink
Queue Drops
Delivery
Savings
2 s
FLAMA
4 s
6s
2 s
S-MAC
Sensors
4 s
Topology
6 s
51
FLAMA Summary

• Simple algorithm that can be implemented on a
  sensor platform.
• Significant performance improvement by
  application awareness.

V. Rajendran, K. Obraczka, and J.J.
Garcia-Luna-Aceves, "Energy-efficient, Flow-Aware
Medium Access for Sensor Networks," submitted to
ACM SenSys 04.
52
Building Blocks for Distributed Scheduling

• Neighbor Discovery Protocol.
• Gather 2-hop neighborhood.
• Time synchronization.
• Flow Discovery (traffic awareness).
• Gather traffic or flow information in the
  neighborhood.
• Election Algorithm
• Fair, distributed and traffic-adaptive.

53
Future Research Directions

• Framework for other sensor applications like
  control system applications, data aggregation
  etc.,
• Support for multicast and broadcast applications.
• Multi-channel support for improved channel
  utilization.

54
Flexible Interconnection Protocol (FLIP)
55
What is FLIP?

• FLIP uses customizable headers.
• Fields can be turned on and off selecting the
  exact functionality needed and incurring minimum
  overhead.
• Beneficial in power-constrained networks allowing
  efficient protocols without loosing
  functionality.

56
FLIP Headers

• Headers composed of a meta-header and header
  fields.
• The meta-header specifies which fields will
  follow.
• The meta-header is subdivided into bytes and a
  continuation bit flag is used to indicate
  subsequent meta-header bytes.

57
Current FLIP Header
GTP Header
FLIP Header
GTP Flags
58
Sample FILP Packet
FLIP ESP (Extra Simple Packet)
FLIP/GTP packet
59
Using FLIP Directed Diffusion
Flexible Header
Can we optimize Directed Diffusion by using FLIP?
Diffusion Static Header
60
Scenario 1 Monitoring
61
Scenario 2 Data Gathering
62
Data Gathering
63
Scenario 3 Data Aggregation
64
More details

• inrg.cse.ucsc.edu/flip.
• Implementation of FLIP for Linux and TinyOS.
• Publications
• Ignacio Solis, Katia Obraczka and Julio Marcos,
  "FLIP A Flexible Protocol for Communication
  Between Heterogeneous Devices", Proceedings of
  the IEEE ISCC 2001, July 2001.
• Ignacio Solis and Katia Obraczka, The Case for a
  Flexible-Header Protocol (FLIP) in Power
  Constrained Networks, Proceedings of the WCNC
  2003.
• Ignacio Solis and Katia Obraczka, FLIP A
  Flexible Interconnection Protocol for
  Heterogeneous Internetworking, in ACM/Kluwer
  Mobile Networking and Applications (MONET)
  Special on Integration of Heterogeneous Wireless
  Technologies.

65
In-Network Data Aggregation

• Process data as it flows from sensors to sink(s).

66
Data Aggregation

• Process data as it flows from sensors to sink(s).

Sink
Sensor
67
Data Aggregation

• Process data as it flows from sensors to sink(s).

68
Data Aggregation

• Process data as it flows from sensors to sink(s).

69
Data Aggregation

• Process data as it flows from sensors to sink(s).
• Energy savings by transmitting less data.
• More manageable data streams.
• Less processing required at the sink.

70
Observations

• Energy savings depend on aggregator.
• Aggregation by averaging versus aggregation by
  concatenation.
• Trade-off between energy efficiency and data
  quality (accuracy, freshness).

71
Goal

• Study the impact of timing on the performance of
  aggregation algorithms.
• How long should a node wait to clock out data?
• Focus periodic data generation applications.
• Example application continuous monitoring
  (environmental conditions, industrial process,
  etc.).

72
Outline

• Cascading Timeouts.
• Other Timing Models.
• Performance Evaluation.
• Results.
• Conclusions.

73
Cascading Timeouts

• Initial tree establishment.
• Sink as root.
• Sensors as nodes.
• Nodes aggregate data from their children and
  forward result onto sink.
• Data generated periodically.

74
Tree Establishment

• Sink broadcasts request for data.

Sink
Request
Sensor
75
Tree Establishment

• Nodes send reply to parent.
• Discover how many children.

Request
Reply
76
Tree Establishment

• Nodes send reply to parent.
• Discover how many children.
• Establish reverse paths to sink.

Request
Reply
77
Tree Establishment

• Nodes send reply to parent.
• Discover how many children.
• Establish reverse paths to sink.

Request
Reply
78
Timeout Scheduling

• Nodes set timeouts based on position in the tree.
• Cascading effect data wave reaches sink in one
  period.

79
Timeout Scheduling

• Once sink request is received, node schedules its
  timeout after 2e.
• Subsequent timeouts
• scheduled every t interval.

80
Observations

• Timeouts depend on single hop distance (shd).
• It has been shown that shd does not change
  considerably over time.
• Relatively constant offered load.
• Relatively stable data propagation tree.

81
Other Timing Models

• Periodic simple.
• Periodic per-hop.

82
Periodic Simple Aggregation

• Periodic simple nodes wait a pre-defined period
  of time, aggregate all data items received, and
  send out aggregate.
• Aggregation period set to data generation period
  (T).
• Directed diffusion.

T
83
Periodic Per-Hop Aggregation

• Periodic per-hoponce all data items from
  children are received, node aggregates them and
  forwards result.
• Timeout equal o aggregation period.

T
84
Evaluation

• Simulations using ns-2.
• Compared against
• No aggregation.
• Periodic simple.
• Periodic per-hop.

85
Simulation Setup

• 100 nodes.
• 500m2 field.
• 100m transmission range.
• 115 Kbps data rate.
• CSMA MAC.
• Power levels TX at 24.75mW, RX at 13.5mW, and
  idle at 0.675 mW.
• Data generated every second collection tree
  established at 1s and data collection started at
  3s.
• Simulation ran for 20s.

86
Performance Metrics

• Energy consumed.
• Data accuracy and freshness.
• Accuracy ratio between total number of readings
  received at sink to total number of readings
  generated.
• Freshness Time difference between when data is
  generated and when it is received at sink.
• Overhead.
• Communication complexity.

87
Results

• Sink placed in corner, center, or random.
• Data points obtained by averaging over 20 runs.

88
Results Energy Efficiency
89
Results Energy Efficiency
Algorithms
Sink Placement
None Simple Per-Hop Cascading (W)
0.1425 0.0485 0.0485 0.0431
Corner Random Center
0.1293 0.0486 0.0488 0.0425
0.1122 0.0488 0.0489 0.0412
90
Results Accuracy and Freshness
91
Conclusions

• Timing models are critical to balance the energy
  efficiency/data quality trade-off of aggregation
  mechanisms.
• By carefully choosing when to aggregate and
  forward data, considerable energy savings can be
  achieved (as much as 6 times less traffic) while
  maintaining data freshness and accuracy.

92
Collection under Packet Loss

• Can't spend time recovering.
• Loosing aggregated packages is a risk.
• Proactive recovery.
• Double-Send
• Max-Send
• Adaptive-Send

93
Packets Sent per Round
94
Readings Collected per Round
95
Packets per Reading
96
Spatially-Correlated Data Aggregation
97
Aggregating Data by Grouping

• When to aggregate is not enough for collecting
  all the data on the network
• We need to exploit the correlation between the
  data being collected.
• Can we group nodes? How do we determine the
  grouping? How do we represent them?

98
Isoclusters Grouping by Value

• Group nodes into isoclusters, where all the
  members have a sensed variable in the same range
  (i.e. contour maps, isotherms)
• When collecting data we can focus on isolines
  (isopleths)

99
Grouping Related Work

• Isobars - (TAG)
• Divide world into grid
• Group into discrete polygons and aggregate as
  reports flow
• Optimize/Deal with losses by cutting parts from a
  bounding box.
• E-Scan
• Aggregate into polygons as data flows through
  tree
• Join polygons if close by and if they meet range
  criteria

100
Isolines

• Isolines are lines which pass through our network
  and have the same value. Nodes detect them by
  comparing the value they are sensing with their
  neighbors.
• When nodes detect a nearby isoline they send a
  report to the data sink.
• Only nodes detecting lines report.

101
Isoclustering Example

• ns-2, 400 x 400 meters
• 16 x 16 sensor nodes in grid patter
• 40 meter communication range
• We map reality, no aggregation and isoclusters.
• Isoclusters sends 1/3 of the readings no
  aggregation sends.
• We map with GMT mapping tools

102
Isoclustering Example (cont)
Reality
All nodes reporting
103
Isoclustering Example (cont)
Reality
Isoclusters
104
Isoclustering Example (cont)
Reality
Isoclusters reporting nodes
105
Front Monitoring
T11
T4
T7
106
Front Monitoring _at_ T7
None Optimized
None
Isocluster
Polygon
107
Sensor Networks for Habitat Monitoring
108
Monitoring

• Monitoring the environment (e.g., variations of
  conditions such as temperature, humidity, seismic
  activity, etc.).
• Understanding certain animal species.
• Studying ecosystems (e.g., interactions between
  animals and environment, etc.).

109
The coyote project

• Understand the physiology and behavior of large
  predators and their interaction with their
  ecosystem.
• To date, very little information on terrestrial
  mammals.
• Critical to understand their energetic and
  habitat needs (e.g., ecosystem balance,
  conservationism, etc.)

110
Interdisciplinary research

• Use of sensor network technology to provide
  biologists with information they need.
• Example obtain physiology information by
  correlating movement patterns, eating behavior,
  daily activities, etc.

111
Heterogeneous network

• Static sensors.
• Continuous power.
• High bandwidth.
• Mobile sensors.
• Animal-borne.
• Anemic low power, low bandwidth.

112
Multi-tiered sensor network

• Static sensor lt-gt static sensor.
• Static sensor lt-gt mobile sensor.
• Mobile sensor lt-gt mobile sensor.

113
Self-configuring network

• Network adapts based on what it learns about the
  behavior of target objects.
• Example deployment of additional nodes based on
  acquired knowledge of animals whereabouts tuning
  protocol parameters (e.g., when to transmit data)
  for performance versus energy efficiency balance.

114
Where are we now?

• First version of the mobile sensor node.
• Collar fitted with an accelerometer and GPS.

115
GPS Antenna
Accelerometer
2 Mb Flash Memory
GPS Module
Microcontroller
116
Preliminary experiments
117
(No Transcript)
118
(No Transcript)
119
Energy model
120
Motivation

• Understand tradeoffs between computation and
  communication.
• Energy spent to process data locally.
• Energy required to transmit/receive data.
• Node decides adequate data representation full
  video stream, yes/no, etc.

121
Energy model for communication

• Ad hoc networks are typically energy constrained
  environments (e.g., sensor networks).
• Models in current network simulators (QualNet,
  GloMoSim and ns-2) either do not model all the
  radio states or do not take into account the
  energy consumed by them.

122
Features

• Explicitly accounts for low-power radio modes.
• Considers the different energy costs associated
  with each one of the possible radio states, i.e
• Transmitting,
• Receiving, overhearing and sensing,
• Idle,
• Sleeping.

123
Model

• Energy spent while in a given radio state y is
• Ey Py Ty
• Py represents the power dissipated by the radio
  while in state y, and is given by P V i
• Ty represents the time spent in state y, and is
  given by t PacketSize / TransmissionRate

124
Processing/sensing energy model

• For simple (scalar) sensors (e.g., temperature),
  energy consumed by communication subsystem
  dominates.
• However, for more sophisticated sensors, (e.g.,
  cameras), this is probably not true.
• How do we account for energy consumed by
  processing/sensing activities?

125
Current approach

• Measure energy consumed by various processing
  activities.
• Set of experiments
• Baseline system.
• Simple processing.
• Network transmission and reception.
• Video capturing.
• Video capturing network transmission.