A Resource Estimation and Call Admission Algorithm for Wireless Multimedia Networks using the Shadow

1
A Resource Estimation and Call Admission
Algorithm for Wireless Multimedia Networks using
the Shadow Cluster Concept
David A. Levine, Ian F. Akyildiz, Mahmoud
Naghshineh
Presented by Naresh Verma (naverma_at_vt.edu) Ap
ril 16, 2002
IBM T. J. Watson Research Center Yorktown
Heights, NY 10598
Broadband and Wireless Networking Lab School of
Electrical and Computer Engineering Georgia
Institute of Technology Atlanta
2
AGENDA

• Abstract
• Background
• Introduction
• Wireless Architecture
• Shadow Cluster Concept
• Methods
• Performance Evaluation
• Conclusion
• Questions / Comments

3
ABSTRACT
The idea - Shadow Cluster Concepts - Summary

• Estimate future resource requirements
• Decide if a new call can be admitted to a
  wireless network based on its QoS requirements
  and local traffic conditions
• Message system between mobile terminals and base
  station exchanging information about it position
  and movement parameters

4
BACKGROUND
Issues in Wireless Networks

• Connectivity
• Power
• Bandwidth
• Precious but scarce resource
• Need to efficiently use it

5
BACKGROUND
Issue

• Bandwidth Management
• a base station may need to reserve resources to
  support active users expected to arrive in its
  cell
• at the cost of denying admission to new calls to
  the network

6
BACKGROUND
Issue

• Base Station Bandwidth Management Conflicting
  Requirements
• Maintain maximum resource (bandwidth) utilization
• Reserve enough bandwidth resources ( so handoffs
  are smooth)

The probability of unsuccessful hand-offs - a
QoS metric , e.g., call dropping probability,
that network agrees to maintain
7
BACKGROUND
Challenges

• Accurate determination of resources
• Provide customized QoS metrics to the network
  beforehand

Wireless networks will support wide-range of
applications with diverse bandwidth
requirements Demand on wireless networks
expected to change abruptly On a per call
and/or on a service basis, enabling users to
select a level of service according to pricing
plan
8
BACKGROUND
Understand the Problem

• One needs to understand issues and relationship
  between
• Resource reservation,
• Channel assignment,
• Call admission, and
• Traffic intensity

• Several schemes have been proposed.
• Consider limited information from neighboring
  cells and do not consider admission control
  policies as means to prevent congestion

9
INTRODUCTION
The Paper

• Shadow Cluster Scheme
• A predictive resource estimation scheme
• Dynamically reserves resources that are needed
• Maintains call dropping probabilities requested
  by network
• Provides high wireless network utilization
• Call admission decisions, based on call
  requirement and local traffic conditions

10
INTRODUCTION
Shadow Cluster Framework

• Requirements
• Position
• Movement Parameters

11
INTRODUCTION
Shadow Cluster Framework

• Resource Requirements

12
INTRODUCTION
The Paper

• Proposes
• Shadow Cluster Concepts

A predictive resource allocation scheme that
provides high wireless network utilization by
dynamically reserving only those resources that
are required to maintain the call dropping
probability requested by the wireless connection
13
INTRODUCTION
Shadow Cluster Scheme

• Dynamic,
• Pro-active
• Amount of resource to be reserved is determined
  on-the-fly
• Call admission are aimed at preventing congestion
  conditions
• Designed to operate in future ATM-bases wireless
  networks
• small cell sizes,
• High cell hand-offs

14
INTRODUCTION
Shadow Cluster Scheme

• Handle traffic with very diverse characteristics
  and requirements
• Handle traffic produced by mobile applications

15
INTRODUCTION
Shadow Cluster Scheme

• Increases QoS
• By decreasing dropped calls during hand-offs

16
Wireless Architecture
Internet Gateway
ATM Network Architecture
PSTN Gateway
ATM Switch
ATM Mux
BS
BS
BS
cell
17
Wireless Architecture
Internet Gateway
ATM Network Architecture
PSTN Gateway
ATM Switch
root
ATM Mux
leaf
BS
BS
BS
cell
18
Wireless Architecture
Internet Gateway
ATM Network Architecture
PSTN Gateway
ATM Switch
root
ATM Mux
leaf
BS
BS
BS
cell

Call setup
19
Wireless Architecture
Internet Gateway
ATM Network Architecture
PSTN Gateway
ATM Switch
VCN
root
ATM Mux
leaf
BS
BS
BS
cell

VCN
20
SYSTEM MODEL
Shadow Cluster Concept

• Consider
• A micro/nano cell wireless network system

supporting mobile
terminals running applications that demand
bandwidth resources

• Mobile users can roam freely within the network
  coverage area, and experience a large number of
  hand-offs

21
SYSTEM MODEL
Shadow Cluster Concept

• Users
• Expect good quality of service
• Small connection setup time
• Low delays
• Small call dropping probability
• Small packet loss probability

• System
• Ideally should satisfy the requested service
• Ideally should deny network access to new calls
  only when strictly necessary

22
SYSTEM MODEL
Shadow Cluster Concept
Problem would have been solved if
System could the future movements
and its related information about a node
predict
Use knowledge of the past to predict future
behavior
23
SYSTEM MODEL
Shadow Cluster Concept Fundamental Principal
Every mobile with an active wireless connection
establishes an influence upon the cells (and
their base stations) in the vicinity of its
current location and its direction of travel.
24
SYSTEM MODEL
Shadow Cluster Concept

25
SYSTEM MODEL
Shadow Cluster Concept


26
SYSTEM MODEL
Shadow Cluster Concept
Shadow Cluster Center
Bordering Neighbor
Non Bordering Neighbor
27
SYSTEM MODEL
Shadow Cluster Concept
The amount of darkness of the shadows that cover
a cell reflect the amount of resources that the
cells base station needs to support the actives
mobiles currently in its own neighboring cells.
Philosophically and Conceptually darkness
reflects resources needed by base station to
support mobile terminals
28
SYSTEM MODEL
Shadow Cluster Concept

• Decision to accept new call request involves
• Shadow clusters and
• The establishment of one or more virtual
  connection (inh case of ATM)

29
SYSTEM MODEL
Shadow Cluster Concept

• Base station
• Exchange information among each other
• Limited Info ( not everything)
• Overburdening with control messages
• Decide which call to accept
• Decide which request to deny

30
INTRODUCTION
Shadow Cluster Framework

• Resource Requirements

31
SYSTEM MODEL
Shadow Cluster How it Works

• Active Mobile handed off to another cell
• Shadow cluster moves with it
• All of the base station in the old shadow cluster
  is notified of the movement
• Mobiles new base station assumes the
  responsibility of supplying the appropriate
  information to the new base station in the new
  shadow cluster

32
SYSTEM MODEL
Shadow Cluster How it Works

• New base station added to the shadow clusters are
  provided appropriate information
• Base stations not in new shadow clusters delete
  their corresponding entries thus freeing resource
  requirements
• If needed new virtual connection trees are
  established, old trees are torn down (if not
  needed)

33
METHODS
Implementation
of an active
mobile node to be found in a certain cell at a
particular time
Projected probability
34
METHODS
Implementation

• Each base station controls a set of active
  mobiles
• Each base station informs its neighbor about the
  future location probabilities of each of its
  active mobiles currently in its control
• Only the probability information is shared among
  base stations

35
SYSTEM MODEL
Shadow Cluster Concept
Exchanges probability information

36
METHODS
Active Mobile Probability

• The probability that an active terminal will be
  found in a future cell at a particular time
• Every base station must inform its neighbors
  about the future location probabilities of the
  active mobile terminals currently under its
  control
• For each active mobile terminal, only those base
  stations which belong to that particular mobile
  terminals shadow cluster are informed about these
  probabilities

37
METHODS
Calculating Active Mobile Probability
j
k
z





x
J is the set of all base station, where j e J
X is the set of all active mobile at present
time, x e X
Set of base station that form the shadow cluster
of active mobile x is K(x), where k e K(x) (
Excluding Current base station)
38
METHODS
Calculating Active Mobile Probability
j
k
z



If x is in control of j, then j should find the
projected active mobile probability
Pam( x, j, k, T) Pam( x, j, k, t1), Pam( x, j,
k, t2), ...Pam(x, j, k, tn)
39
METHODS
Calculating Active Mobile Probability

• Thus the active probabilities Pam are the
  projected probabilities that a mobile will remain
  active in the future and at a particular location
  ( within the shadow cluster)
• The probabilities Pam( x, j, k, T) can be
  interpreted as the percentage of the total amount
  of resources currently being used by mobile x
  that the base station j recommends base station k
  to have available at times t1, t2, .. , tn

Pam( x, j, k, T) Pam( x, j, k, t1), Pam( x, j,
k, t2), ..,Pam( x, j, k, tn)
40
METHODS
Calculating Active Mobile Probability

• Estimates for the immediate future more accurate
  than distant future

41
METHODS
Calculating Active Mobile Probability

• The more knowledge about the dynamics of a mobile
  past and present and call holding patterns, the
  more complex and accurate the probability
  function is likely to become

42
METHODS
Calculating Active Mobile Probability

• Examples
• Current dynamics of a mobile is available, the
  AMP becomes a function of the position, velocity,
  and acceleration vectors of the mobile
• APM can also be defined and refined with mobiles
  past information , geographical features of the
  region where the mobile is currently located

43
METHODS
Calculating Active Mobile Probability in Wireless
Env
For one dimension
x
x
x
right
left
For two dimension
x
x
x
x
x
x
x
44
METHODS
Calculating Active Mobile Probability one
dimension
x
Highway
w
2
1
v
j
J-1
J1
right
left
Assume that mobiles past information is available
and we have its resident time in a cell
45
METHODS
Calculating Active Mobile Probability one
dimension

• We define three parameters
• residence time probability density function fx, j
  (t),
• initial handoff probability vector matrix ?x, j
  (t), and
• initial residence time pdf vector gx, j (t),

46
METHODS
Calculating Active Mobile Probability one
dimension
Residence time probability density function
matrix fx, j (t),
fx, j (t)
where fx, j vw (t), for v, w 1 , 2 is the
pdf of the residence time of mobile terminal x in
cell j, given the fact that mobile enters through
the side v and leaves through w, and mobile does
not turn around
Thus there are pdfs for each direction of travel
from v ? w and w ? v
47
METHODS
Calculating Active Mobile Probability one
dimension
initial handoff probability vector matrix ?x, j
(t)
?x, j0 (t)
?x, j2 (t)
?x, j1 (t)
?x, j (t)
Where ?x, j0 (t) is the probability that mobile
terminal will remain in cell j, given call
initiated in cell j
?x, j w (t) for w 1, 2 is the probability
that mobile terminal will have left cell j by
time t through side w, given call initiated in j
The handoff probability for this vector are
required because unless a mobile crosses a cell
boundary it is not possible to determine
direction of travel of a mobile that is
requesting admission to cell j
48
METHODS
Calculating Active Mobile Probability one
dimension
initial residence time pdf vector gx, j(t),
gx, j(t)
gx, j 2 (t)
gx, j 1 (t)
Where gx, j w (t), for w 1, 2 is the
residence time distribution of mobile terminal x
in cell j, given the call initiated in cell j,
and that the mobile terminal exits the cell
through w
49
METHODS
Calculating Active Mobile Probability one
dimension

• Remember, Active Mobile Probabilities are
• projected probabilities
• that a mobile terminal will remain active and
• will be at a specific location

Thus we consider active pdfs of the form
hx, M(x)(t)
represents distribution of call lengths for
mobile x using a service ( video, audio, voice,
fax ) with class descriptor M(x)
M(x) (t) affects QoS
50
METHODS
Calculating Active Mobile Probability one
dimension

• We assume hx, M(x) (t)
• is independent of the dynamics of mobile x
• constructed by measuring the connection times of
  mobile x when using the service with class
  descriptor

51
METHODS
Calculating Active Mobile Probability one
dimension
Finally, compute active mobile probabilities P x,
j, j , (t) for this cell
For a mobile x initiating a call of class M(x)
while in cell j at time t, P x, j, j , (t) for
this cell is
1 - hx, M(x) (t)

P x, j, j , (t)


?x, j w (t)
1 - gx, j w (t)
?x, j0 (t)


P x, j, j (t) probability that the call will
not end by time t and the probability that the
mobile will still be in cell j at time t
52
METHODS
Calculating Active Mobile Probability one
dimension
Next, compute active mobile probabilities P x, j,
k , (t) for neighboring cells
x
Assume mobile x traveling to the right
Determine residence time pdfs q x, j, k , (t) ,
for k gt j

• Assuming mobile is in cell j, q x, j, k , (t) ,
  describes the probable residence times of mobile
  x in cell j to k and
• q(t) available for other cells
• pdfs are independent

53
METHODS
Calculating Active Mobile Probability one
dimension
Thus residence time pdfs q x, j, k , (t) for
neighboring cells is determined by convolving the
pdfs which correspond to the residence times of
the cells that mobile x will visit enroute to
cell k
..
gx, j 2 (t)
fx, j 1 1 2 (t)
q x, j, k , (t)
fx, k 1 2 (t)
fx, k - 1 1 2 (t)

• initial residence time pdf vector gx, j (t)

• residence time probability density function fx, j
  (t)

54
METHODS
Calculating Active Mobile Probability one
dimension
If Q x, j, k , (t) cumulative residence time
pdfs then

• Q x, j, k , (t) is the probability that the
  mobile x will be beyond cell k at time t, and
• Q x, j, k-1 , (t) - Q x, j, k , (t) is the
  probability that the mobile x will be in cell k
  at time t

55
METHODS
Calculating Active Mobile Probability one
dimension
Thus active mobile probabilities P x, j, k (t)
that a mobile x in cell j will be in cell k at
time t is
1 - hx, M(x) (t)

P x, j, k, (t)


?x, j 2 (t)
Q x, j, k-1 , (t) - Q x, j, k , (t)

56
METHODS
Calculating Active Mobile Probability two
dimensions

• Assume
• Exact position and dynamics of mobile terminals
  are unknown
• Mobile terminals last position is limited to area
  where last handoff occurred

4
3
5
j
2
6
1
57
METHODS
Calculating Active Mobile Probability two
dimensions

• Note infinite number of routes can be taken when
  traveling two cells.
• We need to limit possible number of routes in
  calculations ( those with short travel times)

58
METHODS
Calculating Active Mobile Probability two
dimensions

• Assume
• Probabilistic information of mobile is available
  in the form of
• pdfs which describe the time spend in different
  cells and
• Handoff probabilities that mobile will move from
  a given cell to neighboring cell
• Users are well behaved ( moves through well
  defined paths and not at random)

59
METHODS
Calculating Active Mobile Probability two
dimensions
The handoff probability matrix Fx, j (t) for
mobile x describes the probability that a mobile
x given that it enters cell j through a
particular side, will remain in the cell, or will
exit the cell through a specific side
Fx, j (t) depends on t
Because the mobile handoff behavior is different
for different time of the day
60
METHODS
Calculating Active Mobile Probability two
dimensions
Probability that mobile x will remain in cell j
given that it enters the cell through side v
Fx, j \ 14 (t)
4
Fx, j \ 13 (t)
Fx, j \ 15 (t)
3
5
j
Fx, j \ v0 (t)
2
6
Fx, j \ 12 (t)
Fx, j \ 16 (t)
Probability that mobile x will leave cell j
through side w, given that it enters the cell
through side v
1
Fx, j \ 11 (t)
61
METHODS
Calculating Active Mobile Probability two
dimensions

• We also define these parameters
• residence time probability density function fx, j
  (t),
• initial handoff probability vector matrix ?x, j
  (t), and
• initial residence time pdf vector gx, j (t),

62
METHODS
Calculating Active Mobile Probability two
dimensions
P x, j, j (t), probability that the call will
not end by time t and the probability that the
mobile will still be in cell j at time t
1 - hx, M(x) (t)

P x, j, j , (t)


?x, j w (t)
1 - gx, j w (t)
?x, j0 (t)


63
METHODS
Calculating Active Mobile Probability two
dimensions
Calculating Routes/Path
Mobile x in cell j can enter cell k through any
sides of the cell v 1, 2, ,6. For a
particular side there can be r 1, 2,
different routes that can be taken.
64
METHODS
4
3
5
j
x
2
6
1
k
l
65
METHODS
All possible sides v where mobile x can enter
cell k
All possible routes r v from cell j to cell k
4
The probability that the mobile terminal will
stay in cell k at time t given that the terminal
later exits the cell through side w
3
5
j
x
2
6
1
k
x
l
66
METHODS
Calculating Active Mobile Probability two
dimensions
Calculating Routes/Path
The probability Rx, r v of taking a particular
route is
Rx, r v

?x, jw ( r v , j ) (t)
Fx, l \ v ( r v, l ) w ( r v, l) (t)

• Where
• l is a cell along the route rv, and
• v (r v, l) and w(r v, l) are the entrance and
  exit sides (in cell l) , respectively.

67
METHODS
Calculating Active Mobile Probability two
dimensions
Thus q x, j, m (r v) (t), the residence time
pdf for the cells along the route up to cell k and
Q x, j, k (r v) w (t) the probability that
the mobile terminal x will be in cell k at time
t, given that it takes route rv from cell j to
cell k , and that it will leave cell k through
side w
68
METHODS
Calculating Active Mobile Probability two
dimensions
P x, j, k (t) probability that mobile x,
currently in cell j will be active and in cell k
at time t
69
METHODS
Calculating Active Mobile Probability two
dimensions

• P x, j, k (t) should be reasonable in practice
• For a given mobile x in cell j
• Handoff probabilities and residence time pdfs
  should be zero or will be taken as an average of
  several xs
• Introduction of navigation systems should
  increase accuracy

70
METHODS
Calculating Resource Demands
Principle responsibilities of a base station?

• Provide incoming hand-offs from active mobiles
  with different bandwidth requirement
• Provide uninterrupted service for new connection
  requests accepted by the base station

71
METHODS
Calculating Resource Demands

• Assign every base station a C Bandwidth Units (
  BU)

• BU is the minimum quota of (uplink) bandwidth
  resources assigned to a mobile terminal

• Example
• Voice call may require a single BU
• Video session may need several BUs

72
METHODS
Calculating Resource Demands

• Assign every base station a C Bandwidth Units (
  BU)

• Every base station has
• used up BUs

Cu(t)
Cf(t)

• free BUs

C Cu(t) Cf(t)
73
METHODS
Calculating Resource Demands
C Cu(t) Cf(t)

• C is constant over time
• For TDMA/FDMA fixed channel allocation

Shadow clusters approach can be extended to
include values of C that vary over time
74
METHODS
Calculating Resource Demands
Cuj in a cell j varies over time
When a base station receives handoff requests ,
it makes a best effort to honor them
Dropped call due to insufficient resources is a
no no..

• Dropped calls are due to
• cell overloading
• excessive active users
• insufficent free resources
• And should be avoided to maintain a good QoS

Shadow clusters approach can be extended to
include values of C that vary over time
75
METHODS
Calculating Resource Demands
Shadow clusters provides information about BUs
estimates to be used in the future

• Based on these estimates base station can decide
• Accept or
• Reject a connection

76
METHODS
Calculating Resource Demands
Calculating BUs
x
x
x
An active mobile terminal in a cell has three
possible outcomes

• It can remain active and stay in the cell

• It can terminate the connection without leaving
  the cell

• It can remain active and move to a neighboring
  cell

77
METHODS
Calculating Resource Demands
Calculating BUs
Thus Cuj(t) estimate on the number of BU to be
used by base station j at times t t1, t2, .,
tm
Cuj(t) Cuj(t0) Cuj(t) Cuj(t)

• Cuj(t0) initial of busy BU

• Cuj(t) estimate of of BU s which will
  become free by active users which end their calls
  to move to other cells by time t

• Cuj(t) estimate of of BU s which will
  become busy due to handoffs by external mobile
  terminals moving from neighbor cells within the
  shadow cluster to cell j

78
METHODS
Calculating Resource Demands
How to calculate Cuj(t) , Cuj(t)
Let C(x) denote the number of BUs being used
by active mobile x, and let Xj e X be the set of
all active mobile terminals within cell j. Then
Cuj(t) and Cuj(t)
1 - P x, j, j , (t) C(x)
Cuj(t)
P x, j, k (t) C(x)
Cuj(t)
79
METHODS
Calculating Resource Demands
gt
lt
Cuj(t)
Cuj(t0) Cuj(t) Cuj(t)
Calls need to be dropped
C j
number of BU to be used by base station j at
times t t1, t2, ., tm
Total number of BU available in cell j
80
METHODS
Calculating Resource Demands

• Base station exchange information with its
  bordering and non bordering neighbors
  periodically
• Base station only transmits the estimates of the
  number of BUs to be used in other cells
• Ideally base station recomputes active mobile
  probabilities estimates at each time step, and
  determines estimates the next time t1, t2,
  t3,,tm steps
• Calculates free BUs estimates Cfj(t)
• Cfj(t) Cj - Cuj(t)

81
METHODS
Call Admission Algorithm

• Assume time is quantized in slots of length t
• New call request are received at the beginning of
  each time slot
• Decision regarding an admission request is made
  sometime before the end of the same time slot
  where the request was received

82
METHODS
Call Admission Algorithm

• Steps for admitting a call
• Base station j gather call connection request
  from mobile terminals within its cell
• Base station j defines a shadow cluster K(x) for
  each mobile terminal
• Base station j receives preliminary estimates on
  active mobile probabilities and number of
  requested BUs from neighboring base station
• Base station then computes survivability
  estimates for each mobile terminal
• Preliminary decision is made whether to accept or
  reject a call

83
PERFORMANCE EVALUATION
Simulation

• 4 hour of real time highway traffic
• The highway is covered by 10 cells, laid at 1-km
  intervals
• Each cell has a capacity of 40 BUs
• Constant cell load 360 new calls/hour/cell
• Mobile terminal speed ranging from 70 to 105
• Mobil terminals can transmit voice, audio and
  video with probability 0.7, 0.2 and 0.1
• BUs required by each call type is voice 1,
  audio 5 and video 10

84
PERFORMANCE EVALUATION
Simulation

• Up to one new mobile terminal is generated in
  each cell during each time step t
• If a base station j receives an admission request
  from newly generated mobile x
• It first checks whether it has enough BUs to
  support this call c(x) Cfj(t0)
• If it has enough resources it computes active
  mobile probabilities P x, j, k (t) for its own
  and other cells
• Base station j then shares this probability
  information with other cells which falls within
  the shadow clusters of its current mobile
  terminals
• It receives mobile probabilities of its current
  mobile terminals with call admission requests in
  the present time slots

85
PERFORMANCE EVALUATION
Simulation
Average of BUs that are used in all cells in
the network
Average Bandwidth Utilization
Average Bandwidth Utilization, Percentage of
Dropped Calls
Percentage of Dropped Calls
Rejection Threshold
86
PERFORMANCE EVALUATION
Simulation
Percentage of Accepted Calls
Percentage of Accepted Calls
Rejection Threshold
87
CONCLUSION

• Shadow cluster concepts useful in
• Wireless networks with small cells (nano, micro,
  mini)
• Irregular and time varying traffic loads
• High number of cell handoffs per call

88
CONCLUSION

• Algorithms proposed
• Presume knowledge about probability that a mobile
  terminal will be active in a given cell at a
  particular time
• The accuracy of these assumption is important and
  depends on the amount of knowledge available on
  the behavioral patterns of the mobile terminal
• The effectiveness is closely tied to accuracy of
  active mobile probabilities and variances in
  pdfs

89
CONCLUSION

• Algorithms proposed
• Are scalable for resource prediction and call
  admission

90
QUESTIONS COMMENTS
Thank You Naresh Verma naverma_at_vt.edu