Spectrum Sharing in OFDM-Based Cognitive Radio Networks

1
Spectrum Sharing in OFDM-Based Cognitive Radio
Networks

• C. Rosenberg

This work was done in collaboration with Dr. L.
Le, Profs. P. Mitran A. Girard.
2
Outline

• Introduction to Dynamic Spectrum Sharing
• Our 3 Resource Allocation Problems
• Models
• Formulations
• Results
• Heuristics
• Description
• Results
• Des
• Conclusions

3
The Spectrum and Its Management

• Most governments consider the electromagnetic
  spectrum to be a public resource.
• It is usually allocated by a governmental
  organization (FCC, CRTC, ETSI, ARIB, etc.) that
  defines the spectrum management policy.
• Most of the spectrum is currently licensed to
  users to further the public good, e.g., radio,
  television, etc.
• Examples of licensing
• TV channels, radio,
• Cellular service,
• Unlicensed free for all, subject to some
  constraints (e.g., 900 Mhz cordless phones, 2.4
  Ghz wireless WiFi).
• Common belief we are running out of usable radio
  frequencies. Is that true?

4
Current Spectrum Management Policy

• Fixed allocation
• Rigid requirements on how to use
• Little sharing

5
Spectrum Usage in Space, Time, Frequency
Actual measurements by the FCC have shown that
many licensed spectrum bands are unused most of
the time. In NYC, spectrum occupancy is only 13
between 30 MHZ 3.0 GHz.
6
Spectrum Usage

• Good quality spectrum is under-utilized.
• Hence the problem is more a spectrum management
  policy issue than a physical scarcity.
• The problem is begging for a solution based on
  dynamic spectrum management or access. There are
  many possibilities.
• Cognitive Radio is a (BAD but CATCHY) synonym of
  dynamic spectrum access.

7
Dynamic Spectrum Sharing

• There are 3 ways to share the spectrum
  dynamically
• Dynamic Exclusive Access extension to the
  current licensing policy. Flexible licensing. An
  improvement but not fast enough.
• Open Sharing Model horizontal sharing, a
  generalization of the unlicensed band policy. All
  users/nodes have equal regulatory status. Based
  on the huge success of WiFi and other
  technologies working in the ISM band.
• Hierarchical Access Model vertical sharing. All
  users do not have equal regulatory status (i.e.,
  primary users and secondary users). Secondary
  users can opportunistically access the spectrum
  as long as it does not affect the primary users
  performance. Allows for prioritized spectrum
  sharing provided no harmful interference caused
  to primary users.

8
Harmful Interference

• What is harmful interference?
• Ultimately depends on the application.
• There are generally two broad approaches to avoid
  harmful interference
• Interference avoidance (spectrum overlay)?
• Interference control (spectrum underlay)
• Of course they can be combined
  (overlay)
  (underlay)?

9
Spectrum Overlay Interference Avoidance

• Spectrum overlay approach impose restrictions on
  when and where the secondary users may transmit.
  Secondary users have to identify and exploit the
  spectrum holes defined in space, time, and
  frequency.
• Compatible with the existing spectrum allocation
  legacy systems can continue to operate without
  being affected by the secondary users.
• Regulatory policies define basic etiquettes for
  secondary users to ensure compatibility with
  legacy systems.
• In principle, interference avoidance involves
  only two steps
• Look for holes in spectrum/time.
• Transmit only in those bands at those times.
• Sounds a lot easier than it is.
• Detection of spectral holes is difficult due to
  the large range of potential modulation/coding
  schemes careful measurements based on actual
  primary signal statistics and signatures is
  needed.
• Hidden terminal problem we have to protect the
  primary receivers (but where are they?).
• Fast detection time needed.

10
How to Use Holes?

• Suppose that after some sophisticated signal
  processing, we determine that spectrum occupancy
  is
• How do we use these (non-contiguous) holes?
• OFDM based approach solves the problem naturally.
• OFDM has the advantages that
• It is low complexity (FFT and IFFT based)
• Can be naturally adjusted to fit almost any
  configuration of spectral holes.
• Is growing in popularity (802.11a, 802.16,
  802.22)

11
Spectrum Underlay Interference Control

• Interference avoidance is worst-case design
• In practice, this may be too soft and overly
  limit throughput of secondary users.
• Spectrum underlay approach constraints the
  transmission power of secondary users so that
  they operate below the interference temperature
  limit of primary users (i.e., the receivers).
• Interference temperature introduces new
  opportunities at a cost
• Additional difficulties
• Secondary user needs to measure/know temp. at
  primary receivers.
• Secondary measurements
• Feedback from primary
• Treats interference as noise.

12
Spectrum Opportunity

• Channel is available at A (tx) if no primary rx
  nearby.
• Channel is available at B (rx) if no primary tx
  nearby.
• Channel is an opportunity if available at both A
  and B.

13
A Definition of Cognitive Radio (CR)

• A cognitive radio is an unlicensed communication
  system
• that is aware of its environment
• learns from its environment
• adapts to the statistical variations of its
  environment
• and uses these to
• achieve reliable communication and spectral
  efficiency by employing spectral holes or
  opportunities and does not generate harmful
  interference to the incumbents.

? Cognitive Radios will be complex devices.
14
Resource Allocation for the Secondary Network

• The most common network configuration in practice
  has a star topology.
• Because users have different channel gains and
  bandwidth demands, resources must be allocated
  carefully (this is always true)
• Power
• Rate Modulation/Coding scheme
• We will assume OFDM ? Not all sub-channels are
  feasible for all secondary users
• There are challenging trade-offs between
  sub-channel allocation, power allocation and
  rate.
• Since primary users can be mobile, re-allocation
  must be done in real-time to protect the primary.

15
Some Examples

• Two examples of star networks with cognitive
  features
• IEEE 802.16h (WiMAX) provides extensions to
  support unlicensed co-existence
• IEEE 802.22 is an explicit cognitive WRAN that
  will exploit vacant TV broadcast bands

TV Transmitter
WRAN Base Station
Typical 33km Max. 100km
WRAN Base Station
CPE
16
A little more about IEEE 802.22

• IEEE 802.22 has the following interesting
  characteristics
• Has a complex architecture to detect primary
  users.
• Follows the spectrum overlay approach (avoids
  interfering with primary users altogether)
• Is OFDM based

17
Our Class of Problems

• The class of problems we are interested in is
  resource allocation for star topology cognitive
  networks.
• Our problem is similar to IEEE 802.22, except
  that we follow the spectrum underlay approach
• Our assumptions
• Star based network, downlink only, OFDM, limited
  instantaneous power budget at the base-station,
  max-min fair.

18
Distributed Sensing

• We assume N secondary users, M sub-channels, z
  modulations schemes (rates R1,,Rz and SNR
  threshold ?1,?z).
• The BS is the master of distributed sensing and
  resource allocation, etc.
• As a result of distributed sensing, a table T is
  created, which provides the BS with constraints
  on its transmit power on any given sub-channel to
  avoid harmful interference to primary users.
• T decouples the problem of sensing from that of
  resource allocation.
• Given T, find the best joint sub-channel, rate,
  and power allocation. This allocation has to be
  computed fast (and often).

19
Assumptions (the channel dimension)

• The bandwidth is divided into M subchannels.
• Each subchannel may or may not be used by primary
  users.
• We assume that as a result of channel sensing,
  transmission power at the base station has a
  known constraint on each subchannel j
  (depends on the location of the primary receiver
  using that subchannel).

20
Assumptions (the time dimension)

• The time is slotted. Each user i sends
  periodically information on its perception of the
  primary activity on each channel (mi) its
  channel gains (gi).
• The BS compute the table T and then a resource
  allocation (RA) map that is valid for the
  duration of a frame.
• The BS has a power budget on a per time-slot
  basis to share among all its channels/users.

21
Assumptions (the time dimension)

• If the frame is made of L time-slots (TS), one
  can consider 3 cases
• A RA problem computed on a one-TS basis. The
  resulting allocation is then repeated for the F
  TS of the frame. The RA map then looks like (A).
• A RA problem computed on a frame-basis. The RA
  map looks like (B).
• A RA problem on a F TS-basis and then repeated
  kL/F times.
• These 3 cases can be summarized by taking k in
  1,,L. The larger k, the better the flexibility
  and the higher the complexity.

TS 1 TS 2 TS L
1 i (P11) k l (PL1)
2 j (P12) i l (PL2)
3 k (P12) l m (PL3)

M-1 i (P1M-1) m n (PLM-1)
M i (P1M) n i (PLM)
Channel Users (power)
1 i (P1)
2 j (P2)
3 k (P3)

M-1 i (PM-1)
M i (PM)
Joint sub-channel, rate, and power allocation
22
Our 3 Resource Allocation Problems

• First problem k1, table T, no queues. Very
  similar to a traditional OFDM scheduling problem.
  The only difference is T.
• Second problem kgt1, table T, no queues.
  Surprisingly, nobody seems to have studied this
  case even in a traditional OFDM system.
• Third problem kgt1, table T, with queues. Clearly
  introducing queues, will allow us to be more
  efficient in the way we share the resources. The
  question is does that make the scheduler more
  complicated?
• These three problems are NP hard. NP hard does
  not mean that we should try to solve the problem
  exactly for reasonable size network! It will blow
  up but how fast is not clear.

23
First Optimization Problem

• Parameters Number of subchannels
  Number of secondary users Number of coding
  and modulation schemes Rate of modulation
  and coding scheme .
• Formal optimization problem

max-min rate sijz 1 if channel j is allocated
to transmission between the BS and i with
modulation z
A channel can only be allocated once
Min power to tx from BS to i on subchannel j with
mod. z.
From sensing
Total power constraint
24
Remarks on Optimization Problem

• This is an integer linear program in
• There are variables.
• Example N 40 users, M 120 channels, z 5
  modulation/coding schemes
• 24,000 variables, only 120 of which are not
  zero!
• Problem can be solved using a commercial
  integer programing tool such as CPLEX.
• Takes seconds to minutes, sometimes only yields
  bounds
• Useful for evaluating fast online heuristics.

25
Second Optimization Problem

• New parameter
• F the number of TS over which the RA is done
  (kL/F). We will refer to it as a subframe.
• Formal optimization problem
• Let
• Then

Straightforward generalization. The number of
variables is now multiplied by F.
26
Test Cases

• Primary and secondary users are distributed at
  random inside disks of radius km
  and km respectively.

• Each primary user (receiver) assigned a random
  primary channel.
• Channel gains are mix of Ricean fading and path
  loss

27
Test cases

• The cognitive constraints are determined
  by
• Limiting received power from the secondary
  base-station at any primary receiver on its
  primary channel to at most .
• The system is multirate with rates and SINR
  thresholds Rate SINR
  (dB)1 102 14.773 18.454 21.765 24.9
  1

By default ? 0 dB (we double the noise level)
28
Results (Impact of F and Pmax, Np0)
Average max-min rate for (MNNp) (120 40 0),
infinite queue backlogs (20 realizations per
point)
29
Results (Impact of F and Pmax, Np30)
Average max-min rate for (MNNp) (120 40
30), infinite queue backlogs
30
Results (Impact of F and Pmax, Np60)

Average max-min rate for (MNNp) (120 40
60), infinite queue backlogs
31
Results (Impact of ?)
Average max-min rate for (MNNp) (120 40
50), infinite queue backlogs, F1
32
Third Optimization Problem (1)

• This RA problem takes into account the values of
  the queues at the BS.
• Assumption the BS has one queue per user i and
  uses the number qi of packets in the queue when
  computing the RA at the beginning of a frame.
• We want to ensure that we do not give more
  resources than needed to users.
• Formulating an optimization problem that includes
  the queues is not trivial.
• We will say that a user i has its queue fully
  satisfied if
• Let S sijzt be a feasible resource allocation
  over a subframe (and S be the set of all such
  feasible RA), i.e., one that satisfies all the
  constraints in the previous problem.
• Let O(S) be the set of users whose queues are
  fully satisfied when performing the feasible
  resource allocation S and Oc(S) be its
  complement.
• Then for each feasible RA, S we can compute the
  minimum rate received by a CPE in Oc(S) (i.e.,
  whose queue is not entirely satisfied). Our
  objective is to maximize this minimum over all
  feasible S

33
Third Optimization Problem (2)

• To remove the dependence of the min operation
  over the set of non-bottleneck users Oc(S) we can
  write the objective function in an equivalent
  form as follows
• With µ(x,q) is a function which is defined as
• where ? is a sufficiently large number. This
  transformation can be interpreted as follows. For
  a user i such that is satisfied the objective
  function is large enough that this user will not
  be a bottleneck for the min operation. Therefore,
  the min in the objective function is only applied
  to users with queue backlogs that are not met.
• The problem formulated above is a very large
  non-linear problem with integer variables. It is
  very general and captures several important
  resource allocation problems.

34
Solution Using An Integer Program Solver

• The objective function of the optimal allocation
  problem is not linear in its optimization
  variables. Hence its solution cannot be readily
  obtained by an Integer Program (IP) solver.
• We develop an iterative procedure to obtain its
  solution using a IP solver so that we could
  compute benchmark results for our heuristics.

35
Results (Finite Queues)
Average max-min rate for (MNNp) (120 40 0),
finite queue backlogs
36
Results (Finite Queues)
Average max-min rate for (MNNp) (120 40
30), finite queue backlogs
37
Results (Finite Queues)
Average max-min rate for (MNNp) (120 40
60), finite queue backlogs
38
Need for Heuristics

• There is much literature on downlink resource
  allocation in OFDM.
• Need to develop fast (fast enough to adapt to
  changing primary behaviour) and efficient
  heuristics.
• There are clearly different approaches to develop
  heuristics.
• A common one is to use the following three
  steps1. Power Allocation Distribute power to
  subchannels first.2. Channel and Rate
  Allocation Allocate subchannels and rate to
  users given the power allocated to each
  subchannel.3. Rate and Power Allocation Perform
  rate and power allocation given the channel
  allocation obtained in step 2.
• We have adapted these 3 steps to our cognitive
  framework and added a 4th step that makes the
  heuristic more accurate. We have also improved
  step 2 by reallocating power not being used as we
  go along.
• We have also adapted the 4 steps to take queue
  backlogs into account.
• ? We have created a versatile class of heuristics
  with different trade-offs between accuracy and
  speed.

39
Results (Infinite Queues)
Average max-min rate for (MNNp) (120 40 0),
infinite queue backlogs (20 realizations per
point)
40
Results (Infinite Queues)
Average max-min rate for (MNNp) (120 40 0),
infinite queue backlogs (20 realizations per
point)
41
Results (Infinite Queues)
Average max-min rate for (MNNp) (120 40
30), infinite queue backlogs (20 realizations per
point)
42
Results (Infinite Queues)
Average max-min rate for (MNNp) (120 40
30), infinite queue backlogs (20 realizations per
point)
43
Results (Infinite Queues)
Average max-min rate for (MNNp) (120 40
60), infinite queue backlogs (20 realizations per
point)
44
Results (Infinite Queues)
Average max-min rate for (MNNp) (120 40
60), infinite queue backlogs (20 realizations per
point)
45
Results (Finite Queues)
Average max-min rate for (MNNp) (120 40 0),
finite queue backlogs (20 realizations per point)
46
Results (Finite Queues)
Average max-min rate for (MNNp) (120 40 0),
finite queue backlogs (20 realizations per point)
47
Results (Finite Queues)
Average max-min rate for (MNNp) (120 40
30), finite queue backlogs (20 realizations per
point)
48
Results (Finite Queues)
Average max-min rate for (MNNp) (120 40
30), finite queue backlogs (20 realizations per
point)
49
Results (Finite Queues)
Average max-min rate for (MNNp) (120 40
60), finite queue backlogs (20 realizations per
point)
50
Results (Finite Queues)
Average max-min rate for (MNNp) (120 40
60), finite queue backlogs (20 realizations per
point)
51
Importance of Using F3 with the Heuristics
With queues Np30
Np60
52
Contributions

• On the modelling front Introduce table T to
  represent the cognitive aspect of the system. T
  decouples the distributed sensing from the RA
  problem. Introduce F and w.
• On the benchmark front We show that IP solver
  (i.e., CPLEX) can be used for benchmarking even
  for relatively large systems. This is of course
  true also for pure OFDM system. Nobody seems to
  have done it even in this context, hence limiting
  themselves to small problems.
• On the optimization front Introduce Qs in the
  picture to allow better usage of resources.
  Needed a careful problem formulation. Trick to
  solve the problem with Qs using CPLEX to
  benchmark.
• On the heuristics front
• We have adapted the3 steps to our cognitive
  framework and added a 4th step that makes the
  heuristic more accurate. We have also improved
  step 2 by reallocating power not being used as we
  go along.
• Adapt the heuristics to the case with Qs.
• On the engineering front
• Importance of Pmax.
• Importance of F especially when Pmax is large
  and high Np.
• Importance of taking Q's into account.
• Importance of ?.
• Very good heuristics and importance of step 4.

53
Back-up
54
Results (F1, Impact of Np)
0 30 60
5W 3.5 3.5 3
30W 9 9 8.5
60W 12 10.5 10
90W 13 10.5 10
Np
Pmax
55
Results (F3, Impact of Np)
0 30 60
5W 3.8 3.7 3.2
30W 10 9.2 8.6
60W 12.5 11.5 10.7
90W 13 13 11.5
Np
Pmax