The Performance of TCPIP over Bluetooth

1
The Performance of TCP/IP over Bluetooth

• Chris Snow
• Supervisors
• Serguei Primak, Electrical Engineering
• Hanan Lutfiyya, Computer Science

2
Introduction

• Ask your questions any time!
• Who am I ?
• What is the point of this research?

3
Topics of Discussion

• A brief intro to Bluetooth
• Wireless channel modeling
• TCP/IP in a wireless environment
• Simulations and results
• So what does this all mean?

4
Bluetooth

• A short range (usually 10m), low-power wireless
  protocol
• Used in mobile phones, PDAs, laptop computers,
  wireless headsets, ..
• Allows for automatic detection and connection
  establishment
• The ultimate plug-and-play solution!

5
Bluetooth usage scenario

• A mobile phone supporting data and a laptop
  computer
• With a Bluetooth connection, the laptop can use
  the phones data connection to get Internet
  access
• No cables needed! Just put the phone beside the
  computer (or even in your pocket up to 10m)

6
In-depth Bluetooth

• Devices form piconets, with 1 master and up to 7
  slaves
• Any Bluetooth device can act as a master or a
  slave! Roles change
• Devices can be part of more than one piconet at
  the same time
• Scatternets can be formed collections of
  piconets

7
Bluetooth Packets

• There are about 15 types of Bluetooth packets
• The important variations between types are size
  (number of time slots) and use of forward error
  correction (FEC)
• The six important types are 1,3, and 5 time
  slots, w/ and w/o FEC

8
Wireless Channels

• The channel is whatever medium the message passes
  through between sender and receiver in this
  case, the air, walls, buildings, etc.
• It is very difficult to characterize a wireless
  channel, because it is both spatially and
  temporally varying

9
Fading Channels

• The error pattern of the channel is not uniformly
  distributed
• Some event may cause increased loss in the
  channel for a period of time, then the channel
  state improves
• This is a fading channel a major area of
  communications research

10
Modeling Wireless Channels

• In order to simulate a fading channel, we need to
  develop a model
• This model should represent the quality of the
  channel as a function of one or more parameters
• A common model is the Gilbert-Elliot (GE) model

11
The single channel Gilbert-Elliot Model

• In the GE model, the channel modeled as a
  two-state Markov process
• Either the channel is good or it is bad
• The model has several parameters correlation,
  probability of a Good channel, and the
  probabilities of error given that we are in the
  Good or the Bad state

12
Gilbert-Elliot (contd)

• The correlation relates successive packets in
  time.
• With non-zero correlation, the channel has some
  tendency to stay in the same state on transitions
• Every time a packet is received, the new state is
  computed (Good or Bad) and the packet reception
  is evaluated based on the relevant probability of
  error

13
Two-channel GE model

• When we have two channels, we can model them in
  several ways
• Uncorrelated the two channels are completely
  unrelated (e.g. they are physically remote)
• Completely correlated identical behaviour on
  the two channels (e.g. they pass through the same
  physical space)

14
Two channels (contd)

• The third option is to provide some level of
  correlation between the channels
• This model defines 4 aggregate states based on
  the state of each channel (e.g. G1G2, G1B2, etc)
• We can then define the transition equations based
  on the correlation of each channel and the
  correlation between channels (alpha)
• The spatially correlated channel can be studied
  to examine the effect of different correlation
  levels

15
TCP/IP

• The TCP protocol provides for connections between
  hosts on an IP network
• Includes many performance enhancers, including
  sliding windows and back-off
• The main problem TCP/IP was designed for a wired
  network such as Ethernet

16
TCP/IP with wireless

• In a wireless environment, many things cause
  packet loss
• TCP/IPs mechanisms for dealing with failed
  transmits are not optimal for a wireless channel
• So why use TCP/IP ? Interoperability.

17
The simulations

• I extended the Bluehoc package (an open-source
  Bluetooth simulator) to use the one- and
  two-channel models described above
• Then I studied the effects of different
  correlations, error rates, etc

18
GE model in Bluehoc

• For performance reasons, the error model in
  Bluehoc is packet-based (either a packet is
  received, or it isnt)
• The previous model just provided a table of
  probabilities of error, based on packet type and
  distance (calculated based on channel
  measurements by Rappaport)

19
Fixing the GE model

• Different packet types and distances have
  different probabilities of error!
• The GE model assumes a fixed probability of error
  we shouldnt constantly change it
• This creates a big problem, because the error
  sequence now varies depending on the packet
  sequence ordering, which isnt right

20
The solution

• To fix this, we establish a baseline error rate
  (the maximum 5 slot length, without FEC), which
  is fixed
• Scaling is done based on the actual distance
  and packet type
• Instead of the prob. of error in a bad state
  being 1, we modify it based on the packet type
  and distance

21
Back to simulating

• Now that we have fixed the GE model, we can apply
  it to some simulations
• The main study involves an FTP connection between
  one master, and one or two slaves

22
One channel results
23
Two channels C1
24
Two channels - C2
25
Correlation Sharing

• The previous two graphs show the throughput as a
  function of C2 correlation
• C1 correlation is fixed at 0 (i.e. the errors
  occur randomly)
• Note that the C2 performance doesnt change very
  much
• However, the C1 performance looks much like the
  single-channel results

26
Correlation Sharing 2

• What happens here is that because the two
  channels are correlated (the alpha parameter),
  the state of C2 affects C1
• Even though C1 has randomly occurring errors, the
  spatial correlation causes some portion of the C2
  error stream to be superimposed on C1
• This means that the effective correlation value
  for C1 isnt actually zero
• C2 is not affected by correlation sharing,
  because C1 has zero correlation

27
So what does this mean?

• In some circumstances, the performance predicted
  by a simple model may be completely different
  than the actual performance
• The use of a more realistic channel model (with
  non-uniform distribution of error) provides a
  more accurate model of performance
• TCP/IP performance over a Bluetooth link could be
  really bad, depending on the physical situation

28
Acknowledgements

• Thanks to Hanan and Serguei for acting as
  supervisors!
• C. Snow, and S. Primak, Performance Evaluation
  of TCP/IP in Bluetooth Based Systems, Proc. IEEE
  VTC Spring 02, in press.

29
Questions?

• You must have some questions to ask. ?????