Performance Analysis of Bayesian Networks-based Distributed Call Admission Control for NGN
Abul Bashar, [email protected]
College of Computer Engineering and SciencesPrince Mohammad Bin Fahd University
Al-Khobar, KSA 31952
Detlef Nauck, [email protected]
Research and Technology British Telecom, Adastral Park
Ipswich, UK IP5 3RE
DANMS 2012: 5th Workshop on Distributed Autonomous Network Management Systems
Gerard Parr, [email protected] Sally McClean, [email protected] Bryan Scotney, [email protected]
School of Computing and Info. Engg.University of Ulster
Coleraine, UK BT52 1SA
Outline
DANMS 2012, 16th April 2012
Introduction & Motivation Related Work Proposed Approach Implementation Details Results and Discussion Future Work and Conclusion
Motivation : NGN and its Challenges
IP-based, over WDM
NGN: ITU-T recommendation, Guaranteed QoS, Converged services
Reduces: CAPEX and OPEX Challenges: Complex, heterogeneous, unpredictable Qos Provisioning: Call Admission Control (CAC) at network
edges Problems with existing CAC: analytically intractable, non-
scalable Machine Learning for CAC: Autonomic, Scalable and Predictive
solutions Our contribution: Distributed CAC for NGN
Fixed, wireless & mobile
Call Admission Control function for QoS
DANMS 2012, 16th April 2012
Related Work and Research Objectives
Neural Networks (in CDMA Cellular networks) Reinforcement Learning (in Wireless Cellular networks) Support Vector Machines (in UMTS networks) Genetic Algorithms (in Wireless Mesh Networks) Bayesian Networks (in NGN)
Existing Approaches : ML-based CAC for various networks
Our proposed objectives Study pros and cons of centralised and distributed solutions To compare ML-based Centralized and Distributed CAC
approaches Performance Analysis : Prediction Accuracy, Complexity,
Speed, Call Blocking Probability and QoS provisioning
Drawbacks of Existing Approaches Implemented on single network element : Stand-alone
solutions Centralised solutions : Multiple element solutions are not
distributed No solution concerning ML-based distributed CAC
DANMS 2012, 16th April 2012
Bayesian Network Representation
BN is a probabilistic graphical model, a mapping of physical system variables into a visual and intuitive model
Directed Acylic Graph structure : using nodes and arcs Encodes conditional independence relation among system
random variables Defined mathematically using joint probability distribution
formulation Inference feature : Repeated use of Baye’s rule to estimate
unobserved nodes based on evidence of observed nodes
PHYSI CAL SYSTEM
IP CORE NETWORK
EDGE NETWORK
ACCESS NETWORK
ACCESS DEVICES
SERVICE USERS
APPLICATIONS / SERVICES
GATEWAY
B
S
A
X
ED L
T
BAYESI AN NETWORK MODEL
DANMS 2012, 16th April 2012
Basic theory of BN-based CAC CAC is generally implemented
at network edges Input
Traffic Descriptors (Peak rate, Average rate, Burst duration, Service Class)
Qos Metrics (Packet Loss, Delay, Jitter)
System State (Link Bandwidth, Buffer occupancy)
Output Admission Decision (Admit or
Reject) Estimation of Qos Metrics
(Packet Loss, Delay, Jitter) Operation
Trained offline and then used for online decision-making
Key Performance measure: Prediction accuracy, Model complexity, Speed, Blocking Prob. & QoS metrics
BN-based CAC Framework on a Single Link
DANMS 2012, 16th April 2012
Distributed Bayesian Network Formulation
Multiple edge router topology for distributed CAC study Three edge router pairs (IR0-ER0, IR1-ER1 and IR2-ER2) Three BN models for each pair (BN0, BN1 and BN2)
DANMS 2012, 16th April 2012
BNDAC Framework for Multiple Routers BN Models
Experimental Setup Details
Parameter Value
Sources S0, S1, S2
Destinations D0, D1, D2
Ingress Routers IR0, IR1, IR2
Egress Routers ER0, ER1, ER2
Core Routers CR0, CR1, CR2
Parameter Value
Flow generation rate (flows/sec) 5
Average flow duration (sec) 2.0
Packet generation rate (packets/sec) Exponential (4)
Packet size (bits) Exponential (1024)
Type of serviceExpedited
Forwarding
Topology definition
Source Traffic definition
Network Topology in OPNET
BN Node Description
Traffic Incoming Traffic
Queue Queue Size
Delay E2E Packet Delay
Loss Lost Packets
BN Nodes Definition
DANMS 2012, 16th April 2012
Offline Simulation Results : Prediction Accuracy
Delay Prediction Accuracy Comparison
Centralised_CAC has about 11% more prediction accuracy as compared to the Distributed_CAC
Reason: Centralised model has global system knowledge & hence provides accurate decisions. Distributed models provide local optimal solution.
DANMS 2012, 16th April 2012
0 500 1000 1500 2000 2500 3000 350070
75
80
85
90
95
Distributed_CAC Centralised_CAC
Number of Training Cases
Pre
dic
tio
n A
ccu
racy
(%
)
Simulation Results : Implementation Complexity (1)
Structure Learning Time Comparison
Centralised_CAC takes about 75% more time (3000 cases) to learn the structure as compared to the Distributed_CAC
Reason: Centralised model has to learn more BN nodes and their relationships (i.e more data)
DANMS 2012, 16th April 2012
0 500 1000 1500 2000 2500 3000 35000
20
40
60
80
Distributed_CAC Centralised_CAC
Number of Training Cases
Str
uct
ure
Lea
rnin
g T
ime
(ms)
Simulation Results : Implementation Complexity (2)
Parameter Learning Time Comparison
Centralised_CAC takes about 92% more time (3000 cases) to learn the parameters as compared to the Distributed_CAC
Reason: Centralised model has to learn the parameter for more BN nodes (i.e more data)
DANMS 2012, 16th April 2012
0 500 1000 1500 2000 2500 3000 35000
50
100
150
200
250
300
Distributed_CAC Centralised_CAC
Number of Training Cases
Par
amet
er L
earn
ing
TIm
e (m
s)
Online Simulation Results : Decision-Making Time
Decision-Making Time Comparison
Centralised_CAC has similar performance as compared to the Distributed_CAC
Reason: Once the models are learnt the online decision-making time is fairly low and does not vary much with the number of training cases.
DANMS 2012, 16th April 2012
0 500 1000 1500 2000 2500 3000 350015
20
25
30
Distributed_CAC Centralised_CAC
Number of Training Cases
Dec
isio
n M
akin
g T
ime
(ms)
Online Simulation Results : Blocking Probability
Blocking Probability Comparison
Centralised_CAC has higher blocking probability as compared to the Distributed_CAC
Reason: In centralised all call request comes to a centralised model and hence takes more time to decide. In distributed model, they make independent decisions
DANMS 2012, 16th April 2012
0 100 200 300 400 500 600 700 800 900 10000
0.2
0.4
0.6
0.8
1
No_CAC Distributed_CAC Centralised_CAC
Simulation Time (sec)
Blo
ckim
g P
rob
abil
ity
Online Simulation Results : Delay Metric
Delay Metric Comparison
Centralised_CAC has lesser average packet delays as compared to the Distributed_CAC
Reason: In centralised CAC it admits lesser calls and hence lesser packets in the queues. The tradeoff between blocked calls and QoS, Distributed scenario is still better.
DANMS 2012, 16th April 2012
0 100 200 300 400 500 600 700 800 900 10000
100
200
300
400
500
600
No_CAC Distributed_CAC Centralised_CAC
Simulation TIme (sec)
Ave
rag
e P
acke
t D
elay
(m
s)
Summary
FEATURE CENTRALISED DISTRIBUTED
PREDICTION ACCURACY HIGH LOW
TRAINING TIME HIGH LOW
ONLINE SPEED SAME SAME
CALL BLOCKING HIGH LOW
QOS HIGH LOW
DANMS 2012, 16th April 2012
Acknowledgement
The authors would like to acknowledge the support of Prince Mohammad Bin Fahd University, University of Ulster, IU-ATC and British Telecom for performing this research work.
DANMS 2012, 16th April 2012