ROBUST QOS CONTROL

FOR IP-BASED CELLULAR NETWORKS

Akihito Okura, Hiroshi Kawakami, Takeshi Ihara, Akira Miura

Network Laboratories, NTT DoCoMo, Yokosuka, Japan

Keywords: proactive control, multipath control, anomaly detection, attractor, linear optimization

Abstract: This paper proposes "Proactive Control and Multipath Control" to realize a robust QoS control system for

IP-based cellular networks. In

these networks, all kinds of traffic will share the same backbone network.

This requires a QoS system that differentiates services according to the required quality. Though DiffServ is

thought to be a promising technique for achieving QoS, a technique that is proof against rapid traffic

changes and an effective path control scheme are not yet available. Our solution is proactive control using

traffic anomaly detection and multipath control using linear optimization. Simulation results show that

proactive control and multipath control improve system performance in terms of throughput and packet loss

when rapid traffic change takes place.

1 INTRODUCTION

It is expected that the ALL-IP network will become

the de-facto platform for the future cellular network,

and so all kinds of real time and non-real time traffic

will share the same backbone network. In cellular

networks, the traffic pattern drastically changes

when certain events occur, such as natural disasters

or fireworks display. This requires a QoS system

that differentiates services according to the required

quality. DiffServ has been proposed by IETF, the

standardization group for Internet technology. When

a packet enters the DiffServ network, the edge router

writes DSCP (DiffServ Code Point) in the IP header.

DSCP is an identifier of the traffic class. In the

router, PHB (Per Hop Behavior), which decides

queue assignment and scheduling, is set as per

DSCP. DiffServ is a scalable and promising

technique because it divides traffic into several

classes as units of priority control. The problem with

DiffServ is that since it only deals with router

internals, it can’t guarantee QoS. There are two

problems that prevent DiffServ from functioning

effectively.

Problem 1 is how to detect rapid traffic change.

entionally in cellular networks, a traffic

anomaly is judged according to a threshold, and

congestion control is begun only after the traffic

exceeds the threshold. This requires a router metric

that mirrors the current traffic situation. After that,

router parameters must be appropriately set. If the

control procedure is begun after the threshold is

exceeded, these procedures might not finish in time

leading to a deterioration in communication quality.

Problem 2 involves path control between routers.

out

ing protocols such as Open Shortest Path First

(OSPF) can be used to implement path control. In

OSPF, traffic concentrates on a specific path

because they forward all traffic across the shortest

path (Moy, 1998). The use of label switching like

Multi Protocol Label Switching (MPLS) is also

available, but the initial Label Switching Path (LSP)

of MPLS is the same as that of the other routing

protocols (Davie, 1998). Therefore, an effective

multipath algorithm is needed to realize QoS control

regardless of the forwarding protocol used.

With regard to these two problems, one approach

is to

repare in advance router parameters for each

type of congestion. For example, by customizing the

threshold, or preparing paths groups for each type of

congestion, calculations for each traffic situation

become unnecessary. However, because in the

ubiquitous networks of the future, various kinds of

traffic in addition to voice traffic will coexist, it will

be impossible to prepare for every possible traffic

situation. For this reason we must calculate the

router parameters individually for each traffic

situation.

This paper tackles Problem 1 by proposing

oact

ive control which can deal with various kinds

199

Okura A., Kawakami H., Ihara T. and Miura A. (2004).

ROBUST QOS CONTROL FOR IP-BASED CELLULAR NETWORKS.

In Proceedings of the First International Conference on E-Business and Telecommunication Networks, pages 199-206

DOI: 10.5220/0001385601990206

 SciTePress

of traffic situations. For proactive control, we first

describe the details of a traffic anomaly detection

approach that isn’t based on thresholding. Next, we

introduce a control switching method that makes

effective use of traffic anomaly detection. To handle

Problem 2, we propose multipath control which

enables the path control to be based on traffic

classes. As the multipath algorithm, we apply a

linear optimization algorithm to guarantee

optimality.

From the QoS viewpoint, it is said that

distributed control system is not suitable because it

fails to offer traffic quality guarantees. Accordingly,

we chose the centralized control model. We use a

QoS manager that can control the entire network

(Figure 1). Proactive control decides whether control

is necessary or not. When control is necessary,

proactive control sends a message to multipath

control which calculates the optimal path setting to

realize QoS control.

Proactive Control

QoS Manager

Network Devices

Multipath Control

Measure for Traffic Change New Parameter Setting

Figure 1: Architecture

The remainder of the paper is organized as

follows. Section 2 describes proactive control. First

we define the meaning of proactive, and then explain

traffic anomaly detection, which is based on the use

of attractors. In Section3 we propose a QoS aware

multipath algorithm that uses linear optimization.

Computer simulations and result are discussed in

Section4. Section 5 concludes our paper.

2 PROACTIVE CONTROL

2.1 Definition of Proactive Control

In conventional cellular network operation, a certain

control procedure is triggered after the threshold is

exceeded. This operation can be described as

reactive control. The problem of reactivity is the

delay in triggering system responses such as

congestion control. Using a threshold means that the

control procedure is not begun until the traffic

exceeds the threshold. If traffic increases rapidly, the

control procedure may not be completed in time. To

solve this problem, it is natural to lower the

threshold to detect traffic anomalies earlier, but this

causes control overhead because low thresholds are

exceeded far more often.

This problem is caused by the lack of an

approach to cope with traffic anomalies. Thus, our

solution is to define proactive control as a

combination of both local and global traffic anomaly

detection. First we describe traffic anomaly

detection.

2.2 Traffic Anomaly Detection

There are several approaches to detect traffic

anomalies. The methods described are rule-based

approaches, finite state machine models, pattern

matching, and statistical analysis (Thottan, 2003).

The rule-based approach uses an exhaustive

database containing the rules of system behavior to

determine if an anomaly has occurred (Ndousse,

1996). Rule approaches are too slow for real-time

detection and are dependent on prior knowledge

about the anomalous conditions on the network

(Lewis, 1993). Moreover, rule approaches rely

heavily on the expertise of the network manager, and

do not adapt well to an evolving network

environment (Franceshi, 1996).

Anomaly detection using finite state machines

model alarm sequences that occur during and prior

to fault events. A review of such state machine

techniques can be found in (Lazar, 1992) and

(Jakobson, 1992). The difficulty encountered in

using the finite state machine method is that not all

faults can be captured by a finite sequence of alarms

of reasonable length. This may cause the number of

states required to explode as a function of the

number and complexity of faults modeled (Thottan

2003).

Statistical analysis uses the standard sequential

change point detection approach. The source of such

analysis is SNMP MIB data. (Thottan, 1998)

proposed duration filter heuristics to obtain real-time

alarms using MIB variables.

In pattern matching approach, online learning is

used to build a feature map for a given network.

These maps are categorized by time of day, day of

week, and special days, such as weekends and

holidays. The simplest way of making the feature

map is to reproduce the traffic pattern. This map,

however, has a time-axis which means that the

memory capacity increases when the monitoring

interval shortens. As for the change in IP traffic

volume, changes over periods of 1 second or less are

important. In this time scale, it is impractical to

make the map mirror the real traffic.

All four approaches are complementary. So we

can combine these approaches to realize a better

detection system. In this paper, we propose a pattern

matching method which utilizes an attractor (Takens,

1981). An Attractor map is constructed from the

ICETE 2004 - SECURITY AND RELIABILITY IN INFORMATION SYSTEMS AND NETWORKS

200

traffic volume. The traffic volume range is not

influenced by the time scale. Moreover, the

granularity of the traffic-axis is less significant than

that of the time-axis. This means that using only

traffic volume makes the map scalable.

2.3 Attractor Map

The attractor is known as a way of extracting a

feature map from time series data, which is based on

deterministic chaos (Takens, 1981). The attractor

can also be used to map certain linear/non-linear

time series data. If the traffic has some patterns, such

as periodic features, the attractor is created by just

coordinate transformation. For example, to extract

the attractor of traffic data, all that is necessary is to

convert the coordinates of the observation time

series system into the delay coordinate system

equivalents (Packard, 1980).

Attractor space can be represented as a multiple-

dimensional space. We explain attractor

transformation for a two-dimensional space. We

select two points from traffic data. These data are

separated by d. These two points yield one point in

the attractor field (Figure 2). The vertical axis of

attractor represents the traffic volume at time n, and

the horizontal axis is the traffic volume at time n-d.

The whole periodic traffic trace is projected to form

a closed attractor orbit. We utilize this useful

characteristic in proposing the attractor map method.

V(1)

V(2)

Y(t-d)

Y(t)

V(1)

V(2)

Traffic Data Attractor

Coodinate

Transformation

Figure 2: Attractor

First we create an attractor from usual traffic

data and prepare an array of bits that has the same

dimensions as the attractor space. We set those bits

that lie within the attractor orbit as "Normal" and

those bits which do not lie within the attractor as

"Anomaly”. We can detect traffic anomalies by

comparing the attractor map for usual traffic with

that for the current traffic (Figure 3).

Current Traffic Data

Attractor Map

Normal traffic is projected to

"Normal" element

Anomaly traffic is projected to

"Anomaly" element

Figure 3: Usage of Attractor

2.4 Soft Standby

Conventional cellular network control is based on

thresholding. A threshold represents to two things:

anomaly detection and response triggering. The

traffic is recognized as anomalous if it exceeds the

threshold. When the traffic exceeds this threshold,

certain control actions such as congestion control are

initiated. If the threshold is set too low, unnecessary

control actions are triggered. If the threshold is set

too high, the network has insufficient time to react

properly if the traffic increase is rapid. Though the

attractor map enables the detection of anomalous

traffic at an early stage, if congestion control is

triggered at each detection point, the overheads

would be excessive. In this paper, we propose the

control switching method; Soft Standby, which

combines the attractor map with threshold-based

triggering; this is shown inFigure 4. The QoS

manager collects the traffic information and verifies

it using the attractor map. Figure 4 (A) indicates the

state in which the network is normal and each router

is using its default setting. If traffic begins to

increase abnormally and the attractor map detects an

anomaly, the system enters the proactive state Figure

4 (B). In this state, the QoS manager forecasts the

future traffic using linear prediction and calculates

new multipath parameters for each router. Details of

multipath setting are given later. Routers store the

settings but continue to use the default setting. If the

network returns to normal, the system returns to

state Figure 4 (A) and default settings continue to be

used. On the other hand, if the traffic continues to

increase and exceeds the threshold, the system enters

the anomalous state Figure 4 (C). The routers then

activate the new settings; this prevents any quality

deterioration such as packet loss or delay increase.

The conventional method delays the

determination of new parameter settings until the

threshold is exceeded. In our proposal, the new

settings are calculated and loaded when the anomaly

is detected. When the traffic reaches the network

limit, the only control function needed is to activate

the new router parameters and this change can be

ROBUST QOS CONTROL FOR IP-BASED CELLULAR NETWORKS

201

done in a short time. It follows that our proposal

yields faster control response than the conventional

method.

3 MULTIPATH CONTROL

Using proactive control, the QoS manager can deal

with rapid traffic change. The next problem is

calculating the parameter settings. As mentioned

before, current path control is based on singlepath

routing. This causes the traffic to concentrate on the

shortest path. To solve this problem, this chapter

details the multipath control approach.

In conventional networks, the traffic

concentrates on a specific path because existing

routing protocols forward the traffic using shortest

path information. Label switching techniques such

as MPLS do offer a form of multipath control, but

the initial path of MPLS is the same as that of

routing protocols. In an MPLS network, when a

node sends priority traffic, the sender should collect

network information such as topology/path

utilization and calculate the best path using a

constraint-based routing protocol such as CSPF

(Jamoussi, 2002). This calculation is independent in

each node, so optimality is not guaranteed. For

instance, when the low priority traffic dominates the

shortest path, high priority traffic has to make a

detour because the sender of high priority traffic

can't make the sender of low priority traffic accept a

detour route. This sender-based approach can't

realize complete QoS control. Moreover, network

information such as topology and path utilization

can not be sent to end terminals, so sender-based

approaches are not practical. Another approach to

realize QoS label switching is to set up LSPs

beforehand. High priority traffic is allocated to the

shortest paths, and low priority traffic is allocated to

round-about routes. This approach makes the lower

priority traffic take a detour route even if the shortest

path is empty. This degrades communication and

consumes network resources. Therefore, a truly

effective multipath algorithm is needed.

In this paper, we propose a centralize method

which utilizes a linear programming to find optimal

QoS paths group.

QoS

Manager

R R

Default Setting

New Setting

Default Setting

New Setting

Default Setting

New Setting

Threshold

Anomaly

Detection

Exceed

Threshold

Back to

Usual

Anomaly

Detection

Below

Threshold

Back to

Usual

QoS

Manager

R R

Default Setting

QoS

Manager

R R

Default Setting

New Setting

Default Setting

New Setting

Default Setting

New Setting

AB C

AAB

A: Normal State B: Proactive State C: Anomalous State

Figure 4: Outline of Soft Standby

3.1 Linear Optimization

The problems that QoS Manager should solve are, a)

can the current path can satisfy the demands of the

current traffic? b) if the current path is sufficient,

how to make the best correspondence between the

paths and traffic classes?, c) if the current path is not

sufficient, how to find the additional path? Problem

a) is equal to a linear problem that verifies whether

the answer that satisfies constraints exists or not.

Problem b) is equal to a linear optimizing problem

which minimizes a certain objective function. In

short, linear programming can be applied to problem

a) and b). Refer to (Press, 2002) for details of linear

programming. In linear programming, the objective

function that should be minimized or maximized is

represented by a linear expression. The constraints

are also composed of linear expressions. The

objective of linear programming is 1) to verify

whether the answer that meets the all constraints

exist, 2) when the answer exists, to find the optimal

variable that minimizes/maximizes the objective

function.

ICETE 2004 - SECURITY AND RELIABILITY IN INFORMATION SYSTEMS AND NETWORKS

202

For N independent variables, minimize the

function

xaxaxaf

0202101

+++= L

(1)

subject to the primary constraints

0,,0,0

≥≥≥

xxx L

(2)

(

2211

≥≤+++

iiNiNii

bbxaxaxa L

)

(3)

2211

≥≥+++

jNjNjj

bxaxaxa L

(4)

2211

≥=+++

kNkNkk

bxaxaxa L

(5)

can have either sign or be zero. A set of

values that satisfies the constraints is called a

feasible vector. The function that we are trying to

minimize is called the objective function. The

feasible vector that minimizes the objective function

is called the optimal feasible vector.

Characteristic 1) is applicable to problem a),

while characteristic 2) is applicable to problem b).

The unsolved problem is problem c). In order to find

the necessity additional path, an understanding of

graph theory is essential.

3.2 Bottleneck Link Detection

We set the traffic amount of each path as the

variable of linear programming. Constraints are

composed of equalities/inequalities concerning the

link bandwidth, traffic amount, delay etc. When a

present path group doesn’t offer enough bandwidth,

there exists a bottleneck link. In this case, at least

one path around the bottleneck is necessary. By

repeating the addition of these necessary paths, we

can get the new path group that satisfies all

constraints. This addition is composed of only

necessary paths, so we can define this as optimal

path addition.

To realize optimal path addition, we first have to

find the bottleneck link first. We take advantage of

the Simplex method, commonly used on linear

programming, to do this.

Imagine that we start with a full N-dimensional

space of candidate vectors. We then eliminate the

regions that are indicated by each constraint. The

area that remains after the elimination is called the

feasible region. If the feasible region doesn’t exist,

there is no feasible vector that satisfies all

constraints. If the feasible vector exists, the optimal

feasible vector should be on the boundary because

the objective function is linear. The simplex method

takes advantage of this characteristic and searches

for the optimal feasible vector along the boundary.

The simplex method is composed of two steps.

First step is the judgment of whether a feasible

vector exists; the second step is to find the optimal

feasible vector. In step 1 we replace our objective

function by a so-called auxiliary objective function

which becomes 0 if the feasible vector exists.

To solve the linear programming problem, we

need to get rid of the inequalities that have form (3)

or (4). We do this by adding to the problem so-called

slack variables;

which, when their

nonnegativity is required, convert the inequalities to

equalities. There is another trick to the auxiliary

objective function. That is the introduction of

artificial variables; we denote them by

The introduction of slack variables and artificial

variables turns (3) (4) (5) into

iNiiNNiNiii

xcxaxabz

−

(6)

Form (6) is not the same as (3) (4) (5). Only

when all

are zero do these forms become the

same. Thus we set (7) as the auxiliary objective

function.

∑

−=

′

(7)

M is the number of constraints.

In step 1, we try to maximize the auxiliary

objective function. If the auxiliary objective function

becomes 0, the feasible vector exists and the next

step begins. If the auxiliary objective function

doesn’t become 0, the simplex method stops and

returns the result that there is no feasible vector.

As previously mentioned, this maximization

process advances along the wall of the constraints.

When the process reaches the vertex, and the

auxiliary objective function has not become 0, there

is no feasible vector that satisfies all constraints. The

coordinate of the vertex represents the answer of the

simultaneous equations used by the process. In other

words, this coordinate means the feasible vector that

satisfies the constraints followed by the simplex

procedure, so this coordinate is the current

maximum traffic volume of each path within the

network limitation. This allows us to find the

bottleneck link by comparing the vector at which the

simplex method stopped with the link bandwidth.

After detecting the bottleneck link, we use Dijkstra's

algorithm, well known as the shortest path algorithm,

to find an additional roundabout path. We

recursively repeat this procedure until we get the

sufficient paths to satisfy the constraints.

3.3 Minimization of Network

Resource Consumption

If the paths are sufficient and the feasible vector

exists, step 2 begins to find the optimal feasible

vector. In our proposal, the objective function

realizes optimality in terms of network resource

consumption.

ROBUST QOS CONTROL FOR IP-BASED CELLULAR NETWORKS

203

is the cost of each link

per traffic unit. The

cost of path

is set as follows

∑

∈

(8)

where variable

represents the traffic load of

path

∑

≤

xCf

(9)

If we can find a feasible vector that satisfies the

constraints and minimizes form (9), it can be said

that the feasible vector is the optimal answer that

minimizes the network resource consumption.

3.4 Scalability Consideration

The practicality of the linear programming method is

not clear. In fact, an attempt to solve for an entire

network may yield a large amount of variables and

constraints. However, the linear programming

formulation can be applied to a uniform topology,

such as a cellular network. Furthermore, the linear

programming method can generate optimal solutions

for small areas of a network, and these local

solutions can be combined to create a global

approximation. And with suitable simplification,

some variables and constraints can be removed.

These considerations above show the applicability of

linear programming to the cellular networks.

4 SIMULATION

We combined our proposals, Proactive Control and

Multipath Control, and evaluated its performance by

computer simulation. To this end we modified

Network Simulator2 (NS2). Figure 5 shows the

simulated topology, which is tree-based structure

typical of cellular and ISP (Internet Service

Provider) networks.

AR1

QoS

Manager

R R RR

AR3

AR4

AR5

AR6

AR2

350Mbps

700Mbps

300Mbps

50Mbps

5msec

30msec

5msec

Figure 5: Simulation Topology

The simulation parameters are shown in Table 1.

The communication time was assumed to follow an

exponential distribution; Destination AR was

randomly set. Figure 6 (a) is a graph of usual traffic

for one day. This traffic data was collected from the

statistics database of Ministry of Public

Management, Home Affairs, Posts and

Telecommunications in Japan (Soumu). Figure 6 (b)

image is a sample of the resulting attractor. These

attractors constructed from normal traffic data are

projected to two-dimensional attractor maps (50x50

elements).

We evaluated our proposal in 2 scenarios. For

each scenario, additional traffic is added to usual

traffic. The additional traffic followed a normal

distribution. The parameters were variance

V[minutes] and terminal number N[terminals].

Table 1: Simulation Parameter

Attractor Map

Time Delay [d] 15 minutes

Soft Standby

Threshold 150M

Multipath

Link Cost of all Paths 1

Traffic

Traffic Class Bandwidth Delay

Class1 150Kbps 50msec

Class2 150Kbps -

Class3 Best Effort

Number of Terminals per AR 3000 (1000 per Class)

Traffic of All Classes 150Kbps UDP

QoS Manager

Preparation time of New Param. 30sec

DiffServ Router

Number of Queues 3

Queuing Method

Weighted Fair Queuing

Q1:Q2:Q3=10:5:1

Queue Assignment

Class1/Signaling:Q1

Class2:Q2 Class3:Q3

ICETE 2004 - SECURITY AND RELIABILITY IN INFORMATION SYSTEMS AND NETWORKS

204

(a) Usual Traffic

Throughput (Mbps)

Class1

Class2

Class3

0 30 60 90 120 150

120

150

0:00

Throughput (Mbps)

Time

Class1

Class2

Class3

0 4:00 8:00 12:00 16:00 20:00 24:00

120

150

(b) Attractor

Figure 6: Usual Traffic and Attractor

Scenario 1 is the fireworks congestion model.

The feature of the fireworks congestion model is that

the additional traffic follows a wide distribution;

V=135, N=9000. The fireworks event takes place at

17:00. In Scenario 2, the earthquake congestion

model, the additional traffic follows a narrow

disrtibution; V=6, N=6000. The earthquake takes

place at 8:00. Figure 7 plots the ideal throughput of

each scenario.

The events in both scenarios were generated

using AR1. The especially high traces in Figure 7,

which represent the theoretical throughput of AR1,

indicate the traffic concentration. We compared 3

methods. Method 1 is a conventional singlepath

control that doesn’t change the path even when

traffic becomes congested. Method 2 is a

combination of our multipath routing and

conventional reactive control; threshold based

control triggering was used. The threshold is set to

trigger multipath control when link utilization arises

to 80%. Method 3 is our proposal, multipath and

proactive control; anomalous traffic detection was

used. In this method, QoS manager detects a traffic

anomaly using attractor maps and trigger multipath

control when link utilization exceeds 80%.

0:00

Throughput (Mbps)

Time

Class1

Class2

Class3

4:00 8:00 12:00 16:00 20:00 24:00

100

150

200

250

(b) Earthquake Traffic

0:00

Throughput (Mbps)

Time

Class1

Class2

Class3

4:00 8:00 12:00 16:00 20:00 24:00

100

150

200

250

(a) Fireworks Traffic

Figure 7: Theoretical Throughput

4.1 Fireworks Scenario

The resulting throughput and packet loss in the

fireworks scenario are shown in Figure 8. The

difference between singlepath and multipath is seen

in the throughput at the traffic peak. The throughput

of Class2 in AR1 is lower than that of Class1. On

the other hand, with multipath control, both Class1

and Class2 keep their guaranteed throughputs and

Class3 throughput is also improved. Moreover

Class3 throughput at AR2 is decreased. This is

because Class2 traffic was detoured and influenced

Class3 traffic in multipath control. In short,

multipath control improved the whole network

throughput and realized QoS control. On the

contrary, there is little difference between reactive

control and proactive control. This is because the

preparation procedure can be finished in time since

the traffic increase is not rapid, unlike earthquake

congestion. In short, in the fireworks scenario,

multipath is better than singlepath but there is little

difference between proactive and reactive.

4.2 Earthquake Scenario

The resulting throughput and packet loss in the

earthquake congestion model are shown in Figure 9.

As in the fireworks scenario, multipath control

offers improved throughput with AR1. In this

scenario, there is a slight difference between reactive

and proactive control when the earthquake occurs.

Throughput (Mbps)

Time

Class1

Class2

Class3

0:00 4:00 8:00 12:00 16:00 20:00 24:00

100

150

200

250

Throughput (Mbps)

Time

Class1

Class2

Class3

0:00 4:00 8:00 12:00 16:00 20:00 24:00

100

150

200

250

Throughput (Mbps)

Time

Class1

Class2

Class3

0:00 4:00 8:00 12:00 16:00 20:00 24:00

100

150

200

250

AR1 Class1

AR1 Class2

AR1 Class3

AR1 Class1

AR1 Class2

AR1 Class3

AR2 Class3

AR1 Class1

AR1 Class2

AR1 Class3

AR2 Class3

(a) Singlepath (b) Multipath & Reactive

Class1

Class2

Class3

Number of Lost Packets

Singlepath

Multipath

Reactive

Multipath

Proactive

2000

4000

6000

8000

10000

12000

14000

16000

18000

20000

(d) Packet Loss

Figure 8: Results of Fireworks Scenario

ROBUST QOS CONTROL FOR IP-BASED CELLULAR NETWORKS

205

Throughput (Mbps)

Time

Class1

Class2

Class3

0:00 4:00 8:00 12:00 16:00 20:00 24:00

100

150

200

250

Throughput (Mbps)

Time

Class1

Class2

Class3

0:00 4:00 8:00 12:00 16:00 20:00 24:00

100

150

200

250

Throughput (Mbps)

Time

Class1

Class2

Class3

0:00 4:00 8:00 12:00 16:00 20:00 24:00

100

150

200

250

AR1 Class1

AR1 Class2

AR1 Class3

AR1 Class1

AR1 Class2

AR1 Class3

AR1 Class1

AR1 Class2

AR1 Class3

Class1

Class2

Class3

Number of Lost Packets

Singlepath Multipath

Reactive

Multipath

Proactive

2000

4000

6000

8000

10000

(a) Singlepath (b) Multipath & Reactive

Figure 9: Results of Earthquake Scenario

The difference is seen in the number of Class2

packets lost. The reason for the difference is the

difference in preparation time. In the proactive

method, the procedure began when the QoS manager

detected the traffic anomaly using agent-initiated

traffic reporting and the attractor map so the

calculation and preparations could be finished before

the traffic exceeded the network limit. On the other

hand, in the reactive method, the QoS manager

didn't detect the rapid traffic change until the next

monitoring period. Since the procedure began after

the traffic exceeded the threshold, it could not be

completed in time, so Class2 packets were lost until

the new settings were established. In short, in the

earthquake scenario, both proactive and multipath

control improved communication quality.

5 CONCLUSION

In this paper, we proposed proactive control and

multipath control to cope with the rapid traffic

change in IP-based cellular networks. Proactive

control is based on traffic anomaly detection via

attractor maps. Proactive control includes a control

triggering method that efficiently switches the

network states. Linear programming is applied to

determine the multipath setting and guarantee

optimality in terms of network resource

consumption.

NS2simulations showed that our multipath

control approach achieved better performance in

terms of throughput and packet loss than

conventional singlepath control. The simulations

also showed that proactive control is especially

effective when the traffic increase is rapid. Proactive

control prevented packet loss in priority traffic class

unlike the conventional reactive control.

REFERENCES

J.T. .Moy, "Anatomy of an Internet Routing Protocol",

Addison-Wesley, January, 1998.

B.S. Davie, Y. Rekhter, "MPLS: Technology and

Applications", Morgan Kaufmann Pub., May, 2000.

M. Thottan, C. Ji, "Anomaly Detection in IP Networks",

IEEE Transactions of signal processing, vol.51, no.8,

August 2003.

T.D. Ndousse and T. Okuda "Computational intelligence

for distributed fault management in networks using

fuzzy cognitive maps", in Proc. IEEE ICC, Dallas, TX,

pp.1558-1562, Jun. 1996.

L. Lewis, "A case based reasoning approach to the

management of faults in communication networks", in

Proc IEEE INFOCOM, vol.3, San Francisco, CA, pp.

1422-1429, March 1993.

A.S. Franceschi, L.F. Kormann, and C.B. Westphall,

"Pefrormance evaluation for proactive network

management", in Proc. IEEE ICC, Dallas, TX, pp. 22-

26, June 1996.

A. Lazar, W. Wang, and R. Deng, "Models and algorithms

for network fault detection and identification: A

review," in Proc. IEEE Int. Contr. Conf., 1992

G. Jakobson and M. D.Weissman, “Alarm correlation,”

IEEE Network, vol. 7, pp. 52–59, Nov. 1993.

M. Thottan and C. Ji, “Proactive anomaly detection using

distributed intelligent agents,” IEEE Network, vol. 12,

pp. 21–27, Sept./Oct. 1998.

Floris Takens. "Detecting strange attractors in turbulence,"

in "Dynamical Systems of Turbulence", Vol.898, of

Lecture Notes in Mathematics, pp.366-381, Berlin,

1981.

N.N. Packard, J.P. Crutchfield, J.D. Farmar, and R.S.

Shaw. "Geometry from a Time Series", Physical

Review Letters, Vol.45, No.9, pp.712-716, September,

1980.

B. Jamoussi, et al, "Constraint-Based LSP Setup using

LDP", RFC3212, January, 2002.

W. H. Press, et al "Numerical Recipes in C++",

Cambridge univ. press, 2002.

NS2, http://www.isi.edu/nsnam/ns/

Soumu, http://www.soumu.go.jp/index.htm

ICETE 2004 - SECURITY AND RELIABILITY IN INFORMATION SYSTEMS AND NETWORKS

206