A case study

1
A case study

• Describing a street testbed we recently built for
  studying the use of wireless mesh network for
  adaptive traffic control system
• Discuss some initial measurement results
  regarding link characteristics of
• 802.11
• 900Mhz
• Ethernet over powerline
• and Unwired (a WiMax variant)
• Discuss some of our experience in building a
  testbed in a real-life environment

• Describing a street testbed we recently built for
  studying the use of wireless mesh network for
  adaptive traffic control system
• Discuss some initial measurement results
  regarding link characteristics of
• 802.11
• 900Mhz
• Ethernet over powerline
• and Unwired (a WiMax variant)
• Discuss some of our experience in building a
  testbed in a real-world environment

2
Adaptive Traffic Control

• How it works
• Road-side sensors detect the states of
  vehicle/road
• e.g loop detector under the pavement for vehicle
  counting
• Sensor data is fed to traffic light controller
• Sensor data can be also fed to variable speed
  limit sign
• the controller uses the sensor data to make
  decision about the duration of green/red lights

3
Traffic server (Regional Computer)
loop detector
Traffic controller
CC1
4
Communication for traffic control system

• Traditionally rely on wired connections
• Private or leased lines
• High operating cost, inflexibility
• People have started looking at using public
  shared network
• eg. ADSL, GPRS
• Inconsistent delay jitter and reliability issues
• e.g. GPRS can have high RTT (gt1sec), fluctuating
  bandwidth and occasional outage

5
Sydney Coordinated Adaptive Traffic System (SCATS)

• A popular traffic management system (used by gt100
  cities)
• Created by Sydney RTA (Road and Transport
  Authority)
• Serial point-to-point communication over
  voice-grade telephone line, using 300bps
  modem
• Hierarchical structure
• TMC (Traffic Management Center)
• Regional Computers (RC)
• Traffic controllers

TMC
RC
RC
RC
controller
controller
controller
6
Second-by-second SCATS messages
Regional Computer
Controller
Loop detector
7
SCATS protocol

• Periodic message exchanges every sec
• If RC does not receive ACK within 1 sec retry
• If the ACK fails to arrive the 2nd time link
  failure
• Controller enters self-controlling mode and
  stays in this mode for 15 min
• Uncoordinated traffic control
• extend by another 15 min if another communication
  failure happens in this mode

control command
controller
RC
Sensor data ACK
8
Summary for communication layer of traffic control

• Wired connections are typically used
• private or leased from public telecommunications
  operators
• Traffic signal data demand is light
• Low-bandwidth dial-up network is commonly used
• But reliability and latency are critical issues

9
What are the problems?

• High cost
• High front-end cost
• RTA pays 14 millions each year to Telstra for the
  leased lines
• High maintenance cost
• Installation or relocation is expensive
• Very inflexible
• installation/relocation incur long delays
• Low bandwidth
• RTA uses 300 bps dial-up lines!
• Difficult to integrate other sensors/equipment
  (e.g. video cameras)

10
Wireless mesh network

• Getting increasing popularity recently
• Trial deployment in several major cities
• Strix, Tropos, LocustWorld, etc
• A competitive last-mile solution
• Application residential broadband, public safety
• Our research
• Using mesh network for a mission-critical system
  such as traffic control
• Can we use low-cost, standard-based wireless
  technology (such as 802.11, 802.16) to build a
  dedicated RTA wireless network?
• Different requirement from prior work
• Trade throughput for latency and reliability

11
Research Challenges

• Scalability
• Connecting numerous road-side devices to SCATS
• Reliability
• Mission-critical data (e.g. accident detection,
  traffic signal control, etc)
• Requires timely routing that is robust against
  faults in nodes or links
• Low latency
• SCATS is a real-time traffic control system (lt 1
  sec)

12
Todays talk The testbed

• Collaborate with New South Wales Road and Traffic
  Authority (NSW RTA)
• Study the feasibility of using wireless mesh
  network for traffic control

13
Outline

• Background
• Site survey for the testbed
• What is typical node distance?
• Traffic controller map
• Feasibility of using off-the-shelf hardware?
• Intersection-pair measurements
• Wireless mesh testbed
• Preliminary results
• Experience we learned and conclusion

14
Typical distance between two adjacent traffic
lights?

• Q What is the degree of connectivity between
  traffic controller for a given radio range?
• Data source traffic controller map for Sydney
  CBD area (2787 traffic controllers)
• 354 traffic controllers have their closest
  neighbors within 100m
• 2407 traffic controllers have their closest
  neighbors within 500m
• 2701 traffic controllers have their closest
  neighbors within 1000m

15
Degree of connectivity between traffic controllers
to ensure that 90 (2500) nodes each has at
least 3 neighbours (e.g. for fault tolerance)
requires a radio range of at least 1km.
16
Wireless survey

• Building a testbed in real world can involve lots
  of
• NTD 500K for only 7 nodes
• Not to mention the numerous man-hours
• To understand the feasibility of using
  off-the-shelf 802.11 radios products
• What is the performance of 802.11 with different
  parameter settings?

17
Experiment setup

• 20 intersection pairs
• Two linux laptop
• External antennas
• 8 dBi omni-directional
• 14 dBi directional
• Two wireless interfaces Intel Centrino (RFMON)
  and Senao SL-2511CD (200mW)
• Antenna height 4m
• signal loss over the coaxial cable 2.7dB
• Duration of each experiment 5 min (3 times for
  consistency)
• Use GPS to measure distance between intersections

18
Factors that might affect the performance of
802.11

• Effect of
• Distance
• Transmission rate
• Number of MAC-layer retry
• Type of antenna

19
Effect of the distance

• Pathloss attenuation experienced by a wireless
  signal as a function of distance
• Shadowing amount of variations in pathloos
  between similar propagation scenarios
• prior work suggested pathloss can range from 2 to
  5 for outdoor urban environment
• Using linear regression, we find our environment
  has a pathloos 3.1 and shadowing 7.2
• significantly lower than the suggested urban
  pathloss of 4 in the literature

20
Effect of transmission rate

• Higher transmission rates
• allow high quality links to transmit more data
• but have a higher loss probability on lossy
  links.
• throughput is a function of transmission rate and
  the delivery probability.
• We tried 1Mbps, 2Mbps, 5.5Mbps,11Mbps
• Most of our links have a higher throughput when
  using a higher transmission rate

21
Effect of maximum retries

• MAX-RETRY is one of the wireless card parameters
• A higher retry limit
• Decrease the probability that a packet is dropped
  due to a link error
• potentially increase the probability of network
  interface buffer overflow and the latency
• A optimal setting depends on the channel
  conditions and flow rate
• MAX-RETRY10 seems to work best in our case
• MAC-layer re-transmissions is a norm
• our links have intermediate quality

22
More than 50 are retransmitted at the MAC layer
Link distance 200m MAX-RETRY10
23
Omni-directional vs. directional

• Directional antenna
• increased spatial reuse and improved signal
  quality
• less power consumption while maintaining a
  similar link quality
• higher cost
• Deployment
• Opportunistic forwarding

24
Intersection selection (omni-directional, 11Mbps)

25
Intersection selection (directional, 11Mbps)

26
Outline

• Background
• Site survey for the testbed
• Wireless mesh testbed
• Hardware and software
• Preliminary results
• Experience we learned and conclusion

27
Street testbed
CBD
200m
200m

400m
500m
200m
200m
300m
Univ. of Sydney
NICTA
28
STaRCOMM testbed

• Cover 7 intersections in Sydney CBD (Central
  Business District)
• Inter-node distance 200m 500m
• 500m x 1000m area
• Currently extending to 15-20 nodes
• Nodes are custom-build embedded PCs
• NLOS for all the nodes
• Three types of nodes
• mesh nodes
• gateway node
• Curbside node

29
Node

• mesh nodes
• Each node has 3 radio interfaces
• Two 2.4GHz (802.11) or one 2.4GHz one 900MHz
• One 3.5GHz (WiMax variant) for backhaul
• Connect to traffic controller via powerline
  communication
• gateway node
• located at Sydney U.
• Connect to mesh nodes via 802.11
• Connect to Regional Computer (at NICTA) via
  AARNet
• Curbside node
• Located in traffic controller housing
• One serial interface (to traffic controller) and
  one IP interface (to mesh node via
  ethernet-over-powerline)
• Encapsulate SCATS data into IP packet and
  decapsulate IP packet into serial data

30
Traffic controller
413
414
415
c6
Motherboard
Ethernet switch
Unwired modem
521
522
523
524
m0
Unwired Net
Usyd Net
(Internet)
NICTA
Power line
Mesh node
Wired
RC
Testbed management
3.5GHz
Gateway node
2.4GHz
curbside node
900MHz
31
Mesh node

• VIA MB770F motherboard
• Ubiquity SR2 (400mW)
• w/ 8dBi omni-directional ant.
• Ubiquity SR9 (700mW)
• w/ 6dBi omni-directional ant.
• Uniwred modem
• Diamond digital router
• Netgear powerLAN adapter
• Fault recovery
• Remote switch
• Watchdog timer
• Roof for water/heat proof
• Mosquito mesh for insect proof

32
(No Transcript)
33
Gateway Node
34
Curbside Node
Power-over-ethernet adapter
35
Software

• custom-built Linux OS image
• watchdog timer
• A daemon periodically update the timer to keep
  system from rebooting
• Software from Orbit project
• Including OML for measurement collection

36
Outline

• Background
• Site survey for the testbed
• Wireless mesh testbed
• Preliminary results
• Experience we learned and conclusion

37
Effect of hop numbers on losses
(2.4GHz)
One hop 521-522 Two hop 521-523

• Consecutive loss increases as the number of hops
  increase
• On the same link or from different links?

38
Effect of distance on losses (with 2.4GHz)
521-522 200m 521-523 400m
losses become burstier as the distance increases
39
Effect of number of hops on latency
Latency and its variation increase as the number
of hops increase
40
Effect of distance on latency
Latency is not strongly correlated with distance
41
Effect of distance on loss
Loss is not completely correlated with distance
location-dependant
42
Effect of antenna location
Antenna location makes a difference
43
900MHz vs. 2.4GHz
900MHz has a lower loss rate but higher latency
due to retry?
44
Power-line communication
Powerline communication works pretty well when
distance is within Its operation region
45
Throughput from different technologies

• Larger variation for 900 Hz
• powerline does better than radio when the
  distance is short

46
Latency of Unwired link(round-trip delay from
mesh node to unwired gateway)
A

• High latency
• Large variation
• Outage is common

A
B
47
Latency of backhaul link(round-trip delay from
nicta to mesh node)
A
AB
AB
Almost half of the delay happens on the Unwired
wireless link
48
Clear Diurnal Pattern

• More interference?
• Other user traffic causing network congestion?

49
Outline

• Background
• Site survey for the testbed
• Wireless mesh testbed
• Preliminary results
• Experience we learned and conclusion

50
Deployment

• Protection of antenna connectors is necessary
• Connectors often held on by weak glue or crimp.
• Gradual stress (e.g. vibration) could eventually
  loosen the plug
• degrade the signal before it is transmitted into
  the air
• Make sure that your wireless cards comply to the
  specification before starting using them.
• E.g. some of our Senao wireless cards does not
  output 200mW as they should

51
Deployment

• while the hardware can be identical, different
  firmwares and drivers could introduce inaccuracy
  in the measurement results.
• compare against with a spectrum analyzer if you
  can!
• Antenna locations matter!
• At 2.4GHz, a quarter wavelength is approximately
  30cm
• when multiple antennas are deployed, it is
  essential to have a means for independently
  adjusting their position.

52
Maintenance

• Remote management is important for an outdoor
  testbed
• Access the node
• Unwired link
• 802.11 link
• Ethernet port
• Serial port
• Reboot the node
• Remote switch
• Watchdog timer
• PXE network reboot (configured in BIOS)
• DHCP server by default does not provide PXE boot
  info
• Second image for fallback (via Grub)

53
Security

• A major concern to to any wireless network
• Anybody can sniff the air
• Connected to the Internet via Unwired
• Its real!! Two nodes were hacked.
• integrated with the traffic control system
  security model
• segmentation to contain the damage of a attack
• multiple levels of fallback to local control

54
Interference

• 2.4GHz/900MHz are shared channels
• We saw an average of 50 external APs at any time
  of the day
• A serious problem when WiFi becoming more and
  more pervasive

55
Conclusion

• It is feasible to build a wireless network with
  off-the-shelf hardware/software to control
  traffic lights
• Signal quality and losses are location-dependent
  (but not strongly correlated with distance)
• For a good link, losses are in general uniformly
  distributed
• Larger variation in 900MHz than in 2.4GHz
• Powerline communication is excellent for a short
  distance
• Issues with using public shared network
• Large variations and outages is a norm
• Diurnal patterns

56
Future work..

• By collaborating with NICTA and department of
  transportation _at_ NCKU, we plan to a build a
  similar testbed around NCKU campus
• Vehicle-infrastructure communication
• Multimedia (Video/Audio) over mesh
• Hierarchical mesh-sensor networks

57
..Future work

• Wireless data mining
• Loss Model for mesh links
• Outage prediction
• Dynamic channel assignment
• Multi-path routing

58
Thank you!

• Questions?

59
Why WMN for traffic control?

• Low installation cost
• Low front-end investment
• Easy maintenance
• Robust and reliable
• Reliability increases as the number of nodes
  increase

60
Effect of antenna

• directional antenna exhibits similar performance
  as omni-directional antenna for most of the links
  in our environment
• But directional antenna does help for challenging
  links

61
Testbed location

• A typical suburban area with lots of traffic,
  foliages, pedestrians and high-rise residential
  buildings.
• The 200-500m range is representative of 90 of
  the distance between traffic controllers in the
  Sydney CBD area
• Close to NICTA (for on-site maintenance)