9 771949 242004 02 · assume the binary phase shift keying (BPSK) modulation. The result however...

Communications and Network, 2010, 2, 145-206 Published Online August 2010 in SciRes (http://www.SciRP.org/journal/cn/)

Copyright © 2010 SciRes. CN

TABLE OF CONTENTS

Volume 2 Number 3 August 2010 Downlink MBER Transmit Beamforming Design Based on Uplink MBER Receive

Beamforming for TDD-SDMA Induced MIMO Systems

S. Chen, L. L. Yang………………………………………………………………………………………………………………145

Compromise in CDMA Network Planning

S. Hurley, L. Hodge………………………………………………………………………………………………………152

A Novel Black Box Based Behavioral Model of Power Amplifier for WCDMA Applications

A. S. Sappal, M. S. Patterh, S. Sharma……………………………………………………………………………162

Comparison of 4 Multi-User Passive Network Topologies for 3 Different Quantum Key Distribution

F. Garzia, R. Cusani……………………………………………………………………………………………………166

An Introduction to RFID Technology

S. Ahuja, P. Potti………………………………………………………………………………………………………183

A Security Transfer Model Based on Active Defense Strategy

Z. Ying……………………………………………………………………………………………………………………………187

Optimization of UMTS Network Planning Using Genetic Algorithms

F. Garzia, C. Perna, R. Cusani……………………………………………………………………………………………………193

Performance Evaluation of Java Web Services: A Developer’s Perspective

S. Ahuja, J. L. Yang…………………………………………………………………………………………………………200

Communications and Network (CN)

Journal Information SUBSCRIPTIONS

The Communications and Network (Online at Scientific Research Publishing, www.SciRP.org) is published quarterly by Scientific

Research Publishing, Inc., USA.

Subscription rates:

Print: $50 per issue.

To subscribe, please contact Journals Subscriptions Department, E-mail: [email protected]

SERVICES

Advertisements

Advertisement Sales Department, E-mail: [email protected]

Reprints (minimum quantity 100 copies)

Reprints Co-ordinator, Scientific Research Publishing, Inc., USA.

E-mail: [email protected]

COPYRIGHT

Copyright©2010 Scientific Research Publishing, Inc.

All Rights Reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by

any means, electronic, mechanical, photocopying, recording, scanning or otherwise, except as described below, without the

permission in writing of the Publisher.

Copying of articles is not permitted except for personal and internal use, to the extent permitted by national copyright law, or under

the terms of a license issued by the national Reproduction Rights Organization.

Requests for permission for other kinds of copying, such as copying for general distribution, for advertising or promotional purposes,

for creating new collective works or for resale, and other enquiries should be addressed to the Publisher.

Statements and opinions expressed in the articles and communications are those of the individual contributors and not the statements

and opinion of Scientific Research Publishing, Inc. We assumes no responsibility or liability for any damage or injury to persons or

property arising out of the use of any materials, instructions, methods or ideas contained herein. We expressly disclaim any implied

warranties of merchantability or fitness for a particular purpose. If expert assistance is required, the services of a competent

professional person should be sought.

PRODUCTION INFORMATION

For manuscripts that have been accepted for publication, please contact:

E-mail: [email protected]

Communications and Network, 2010, 2, 145-151 doi:10.4236/cn.2010.23022 Published Online August 2010 (http://www.SciRP.org/journal/cn)


Downlink MBER Transmit Beamforming Design Based on Uplink MBER Receive Beamforming for TDD-SDMA

Induced MIMO Systems

Sheng Chen, Lie-Liang Yang School of Electronics and Computer Sciences, University of Southampton, Southampton, UK

E-mail: sqc, [email protected] Received January 10, 2010; revised April 26, 2010; accepted May 30, 2010

Abstract The downlink minimum bit error rate (MBER) transmit beamforming is directly designed based on the uplink MBER receive beamforming solution for time division duplex (TDD) space-division multiple-access (SDMA) induced multiple-input multiple-output (MIMO) systems, where the base station (BS) is equipped with multiple antennas to support multiple single-antenna mobile terminals (MTs). It is shown that the dual relationship between multiuser detection and multiuser transmission can be extended to the rank-deficient system where the number of users supported is more than the number of transmit antennas available at the BS, if the MBER design is adopted. The proposed MBER transmit beamforming scheme is capable of achieving better performance over the standard minimum mean square error transmit beamforming solution with the support of low-complexity and high power-efficient MTs, particularly for rank-deficient TDD-SDMA MIMO systems. The robustness of the proposed MBER transmit beamforming design to the downlink and uplink noise or channel mismatch is investigated using simulation. Keywords: Minimum Bit Error Rate, Time Division Duplex, Multiple-Input Multiple-Output, Transmit

Beamforming, Receive Beamforming, Space-Division Multiple-Access

1. Introduction Motivated by the demand for increasing throughput in wireless communication, antenna array assisted spatial processing techniques [1-7] have been developed in order to further improve the achievable spectral efficiency. In the uplink, the base station (BS) has the capacity to implement sophisticated receive (Rx) beamforming schemes to separate multiple user signals transmitted by mobile terminals (MTs). This provides a practical means of realising multiuser detection (MUD) for space-division multiple-access (SDMA) induced multiple-input multiple-output (MIMO) systems. Tradi- tionally, adaptive Rx beamforming is based on the minimum mean square error (MMSE) design [2,5,6,8], which requires that the number of users supported is no more than the number of receive antenna elements. If this condition is not met, the system becomes rank- deficient. Recently, adaptive minimum bit error rate (MBER) Rx beamforming design has been proposed [9-12], which outperforms the adaptive MMSE Rx beamforming, particularly in hostile rank-deficient systems.

In the downlink with non-cooperative MTs at the receive end, the mobile users are unable to perform sophisticated cooperative MUD. If the downlink's channel state information (CSI) is known at the BS, the BS can carry out transmit (Tx) preprocessing, leading to multiuser transmission (MUT). The assumption that the downlink channel impulse response (CIR) is known at the BS is valid for time division duplex (TDD) systems due to the channel reciprocity. However, for frequency division duplex systems, where the uplink and downlink channels are expected to be different, CIR feedback from the MT's receivers to the BS transmitter is necessary [13]. Many research efforts have been made to discover the equivalent relationship between the MUD and MUT [14-19]. Notably, Yang [20] has derived the exact equivalency between the MUD and MUT for TDD systems under the condition that the number of antennas at the BS is no less than the number of MTs supported1. According to the results of [20], the MUT can be

1All the existing works, such as [14-20], consider the designs of MUD and/or MUT using second-order statistics based criteria, which implies that the MIMO system must have full rank.

S. CHEN ET AL. 146 obtained directly from MUD. Since the BS has to implement MUD, it can readily implement MUT based on its uplink MUD solution with no extra computational complexity cost. This is very attractive as this strategy enables the employment of low-complexity and high power-efficient MTs to achieve good downlink performance.

In general, the BS can design MUT when the downlink CSI is available. The Tx beamforming design based on the MMSE criterion is popular owing to its appealing simplicity [13,21]. Since the bit error rate (BER) is the ultimate system performance indicator, research interests in MBER based Tx beamforming techniques have intensified recently [22-25]. This MBER based MUT design invokes a constrained nonlinear optimisation [22-25], which is typically solved using the iterative gradient-based optimisation algorithm known as the sequential quadratic programming (SQP) [26]. However, the computational complexity of the SQP based MBER MUT solution can be excessive and may become impractical for high-rate systems [23]. This contribution adopts a very different approach to design the MBER Tx beamforming for TDD-SDMA induced MIMO systems, which does not suffer the above mentioned difficulty of high complexity. We prove that Yang's results [20] are more general and the exact equivalency between the MUD and MUT is not restricted only to second-order statistics based designs. Therefore, we can apply the results of [20] to implement the MBER Tx beamforming scheme directly based on the MBER Rx beamforming solution already available at the BS with no computational cost at all, even for rank-deficient systems where the number of antennas at the BS is less than the number of MTs supported. The robustness of the proposed scheme is also investigated when the downlink and uplink noise powers or channels mismatch.

2. Multiuser Beamforming System

The TDD-SDMA induced MIMO system considered is depicted in Figure 1, where the BS employs antennas to support single-antenna MTs. When the uplink is considered, the received signal vector

at the BS is given by

LK

TU L,1 ,2 ,= [ ]U U Ux x xx

=1

= n = nK

U U k kk

s x Hs h U (1)

where the channel matrix is given by L K1 2= [ ]KH h h h , =k k kAh g with kA and kg deno-

ting the channel coefficient and the steering vector for

the kth user, respectively, 1 2= [ ]TKs s ss contains the

K data symbols transmitted by the K MTs to the BS, and denotes the uplink channel nU

Figure 1. Schematic diagram of the TDD-SDMA induced MIMO system employing transmit and receive beamfor- mings at the BS. The system employs L antennas at the BS to support single-antenna MTs. K

additive white Gaussian noise (AWGN) vector with

, and 2[n n ] = 2HU U U LE I LI represents the L L

identity matrix. Without the loss of generality, we assume the binary phase shift keying (BPSK) modulation. The result however can be extended to modulation schemes with multiple bits per symbol by adopting the minimum symbol-error-rate design [9]. The MUD at the BS consists of a bank of Rx beamformers

, = u , 1 ,HU k k Uy x k K (2)

where is the Rx beamformer's weight vector for

user and

uk

k H denotes the Hermitian operator. The

decision variable vector for

the ,1 ,2 ,= [ ]T

U U U U Ky y yy

K transmitted symbols can be expressed as

= =H H HU U y U x U Hs U nU (3)

with the L K Rx beamforming coefficient matrix expressed by

1 2= [u u u ].KU (4)

The real part of is a sufficient statistics for detecting

. Uy

sThe BS employs the Tx beamforming for the

downlink transmission to the K MTs, with the L K transmission preprocessing matrix

1 2= [d d d ],KD (5)

where is the precoder's coefficient vector for

preprocessing the symbol kd

ks to be transmitted to the

th MT. Note that we use the same notation s to represent the downlink symbol vector, without distinction from the uplink symbol vector for the purpose of notational simplification. Due to the reciprocity of the downlink and uplink channels, the received signal vector

, received by the

k

y ,1D Dy y ,2 ,= [ ]TD D Ky K MTs, is

expressed as

= T n ,D Dy H Ds (6)

where nD is the downlink AWGN vector with 2[n n 2] =H

D D DE IK and denotes the transpose T


S. CHEN ET AL. 147 operator. The real part of the decision variable ,D ky is

used by the th MT for detecting the symbol k ks

transmitted from the BS to the th MT. kKUnder the condition , there exists an exact

equivalency between the Tx beamforming preprocessing matrix and the Rx beamforming weight matrix expressed by [20]

L

D U

*= ,D U (7)

where * denotes the conjugate operator, Λ=diag

1 2 , , , K is for achieving the transmit power

constraint, and or is assumed to have been designed based on some second-order statistic criterion.

The exact relationship (7) is valid for

U D

2 = 2D U . A

simple scheme to implement the transmit power constraint is to set = 1k

HU

/ k u for [20].

The relationship (7) is easy to understand. Under the

condition of , of (3) and of (6) have the same full rank and the same second-order statistic properties, given (7).

1 k

TH D

K

HU

L K H

For rank-deficient systems where ,

and no longer have the full rank. Indeed, both the MMSE Rx beamforming and MMSE Tx beamforming turn out to be deficient in this case, exhibiting a high BER floor. However, the MBER Rx beamforming scheme [10,12] has been shown to consistently outperform the MMSE design and is capable of operating in rank-deficient systems where the number of MTs is more than the number of receive antennas at the BS. We will show in the next section that the relationship (7) is not restricted only to second-order statistics based designs. Specifically, with the MBER MUD and MUT designs, (7) is valid and, moreover, the restriction

is no longer required. This enables us to use (7) directly for designing the MBER Tx beamforming based on the available MBER Rx beamforming solution even for rank-deficient systems.

<L K HTH D

KL

3. MBER Receive and Transmit

Beamforming

The BER of detecting ks using the uplink Rx beamfor-

ming with the weight vector can be shown to be

[10,12] ku

( ) ( )

,=1

( ) [ ],

[

1=

N q H qsk k

Rx k kqs k U

sgn sP Q

N

u Hs

uu

( ) (8)

where is the usual Gaussian error function, ( )Q ]

denotes the real part, = 2KsN is the number of all the

equiprobable legitimate transmit symbol vectors,

for

( )qs

1 sq N , and ( )qks denotes the th element of

. The MBER solution for is then defined as

k

(Ru u

( )qs

ku

ku

g minkku

MB ,= ar ).x kP kER, (9)

The optimisation (9) can be solved using a gradient-based numerical optimisation algorithm [12]. Note that the BER is invariant to a positive scaling of

, and one can always normalise the beamforming

weight vector to a unit length, yielding = 1ku , which

significantly reduces the computational complexity of optimisation. This is also useful for directly implementing the Tx beamforming design from the Rx beamforming design using the relationship (7), as the scaling matrix

can be chosen as the identity matrix in this case.

Adaptive MBER Rx beamforming can be achieved using the stochastic gradient-based algorithm known as the least bit error rate [10, 12]. It is clear from (8) and (9) that the optimal MBER Rx beamforming solution to

is self-centred without regarding the effect on the

other users. This type of optimisation is referred to as the egocentric-optimisation (E-optimisation) and the resul

Λ

ku

ting solution is known to be egocentric-optomum (E-opti -mum) [20]. The average BER over all the K users for the Rx beamforming with the beamforming weight matrix is obviously given by U

,

( ) ( )

=1 =1

1( ) = ( ) =

( ) [ ]1,

Rx Rx k k

N q H qK sk k

k qs k U

K

s

KN

U u

u Hs

u

U

=1

K

k

P P

sgnQ

(10)

and the MBER Rx beamforming solution

MBER = arg ( )min RxPU

U (11)

is simply given by MBER MBER,1 MBER,2 MBER,= [ ]KU u u u

MBER,ku

.

Since all the Rx beamforming vectors for

1 k K are optimum in some sense (E-optimum), the Rx beamforming matrix is overall optimum

(O-optimum) [20]. MBERU

With the precoder coefficient matrix , the BER of

detecting by the th MT is

D

ks k( ) ( )

,=1

( ) [ ]1( ) =

N q T qsk k

Tx kqs D

sgn sP Q

N

h Ds

D (12)

for 1 k K . The average BER over all the K users for the Tx beamforming with the precoder’s weight matrix is then given by D

,=1

( ) ( )

=1 =1

1( ) = ( ) =

( ) [ ]1.

K

Tx Tx kk

N q T qK sk k

k qs D

P PK

sgn sQ

KN

D D

h Ds (13)


S. CHEN ET AL. 148

D

The MBER Tx beamforming solution for can be obtained by minimising subject to the given

transmit power constraint

D( )TxP D

MBER = arg ( )min TxPD

D (14)

s.t.transmit power constraint is met

This constrained nonlinear optimisation problem for example can be solved using the SQP algorithm [22,23, 25], which is however computationally expensive. Unlike the E-optimum of the MBER Rx beamforming, the optimal MBER soultion to a precoder’s column vector is not self-centred, as it is clear from (12) to

(14) that an optimal solution to of user not only

maximises the th user’s performance but also pays attention on mitigating its effect on the other

kd

kd k

k1K

users. This type of optimisation is referred to as the altruistic-optimisation (A-optimisation) and the resulting solution is known to be altruistic-optimum (A-optimum) [20]. Denote . Since MBER MBER,1 MBER,2 MBER,= [ ]kD d d d

all the columns for 1 are optimum in

some sense (A-optimum), the precoding matrix

is also O-optimum [20].

MBER,kd k K

MBERD

Instread of applying the computationally expensive SQP algorithm to find the MBER Tx beamforming solution, we can directly derive it as

*MBER MBER=D U , (15)

with no computational cost at all, provided that the relation (7) is not restricted to second-order statistics based designs and it is also valid for the MBER design. We now show that (7) is indeed much more general. We start by examining the probability density functions (PDFs) of [ ]D y and , the real part of [ ]

u y Dy in

(6) and the real part of in (3), respectively. Without

loss of generality, we assume that Uy

= 1Hk ku u , 1 k K ,

as the BER is invariant to the length of [12]. Denote ku

= R IjH

I

H H

=

, with the real part , the

imaginary part and . Similarly , let

= [ ]HRH2 = 1j = [H ]H

R IjUU U and =U U UR Injn n . Then

[ ] = ( ) ( ).T T T TU R R I I R U I UR I

y U H U H s U n U n (16)

The PDF of , denoted as , is

obviously Gaussian with the mean

[ ]U y 1 2( , , , )Up w w w K

s

HK

[ [ ]] = ( )T TU R R I IE y U H U H (17)

and the covariance matrix

2 2 21 2 1

2 2 22 1 2

2 2 21 2

[ [ ]] =

[ ] [ ]

[ ] [ ].

[ ] [ ]

U

H HU U U K

HU U U

H HU K U K U

Cov

y

u u u u

u u u u

u u u u

(18)

On the other hand, with the notation =D D DR Ijn n n

and given , *=D U

[ ] = ( )T TD R R I I D .

R y H U H U s n (19)

The PDF of [ ]D y , denoted as 1 2( , , , )D Kp w w w , is

Gaussian with the mean

[ [ ]] = ( )T TD R R I IE y H U H U s (20)

and the covariance matrix 2[ [ ]] = .D D KCov y I (21)

The BER (10) is determined by the K marginal PDFs of [ ]U y , for 1 , which are

Gaussian distributed and are specified by the mean vector (17) and the diagonal elements of the covariance matrix (18). The marginal PDFs of

, ( )U k kp w

K

k K

[ ]D y ,

, ( )D k kp w for 1 k K , are also Gaussian distributed

and are specified by the mean vector (20) and the diagonal elements of the covariance matrix (21). These two sets of the marginal PDFs are almost ``identical'',

given 2 2=U D

)

. The difference is that only

depends on the th column vector of while ,U kp ( )kw

Uk

, (D k kp w depends on the entire matrix , leading to

the BER expressions (10) and (13). We quote the following result from [20].

*U

Proposition 1 An E-optimum solution in a MUD is equivalent to an A-optimum solution in the corresponding MUT.

Now let be the MBER Rx Beamforming

solution of (11). That is, the column vectors of

are E-optimum in the context of the MBER Rx

Beamforming. Then is the MBER Tx Beamfor-

MBERU

MBERU

*MBERU

ming solution of (14), and the column vectors of

are A-optimum in the context of the MBER Tx Beamforming. Note that we do not require .

*MBERU

L K 4. Simulation Study Full-rank system. The BS employed a four-element linear antenna array with half-wavelength element spacing to support four single-antenna BPSK users. The

angles of arrival (departure) for the four users were 2 ,

, and , respectively, and the uplink channel coefficients for the four users were

16

=k

15

.7071

30

710 0.70A j , 1 k 4 . The full uplink CSI

was assumed to be known at the BS and was used by the BS to design the uplink Rx beamforming. Figure 2 compares the average BER performance, defined in (10), of the uplink MMSE and MBER Rx beamforming schemes. The BS then directly implemented the down-


S. CHEN ET AL. 149

Figure 2. Average BER performance comparison of the uplink Rx and downlink Tx beamforming schemes with both the MMSE and MBER designs for the TDD-SDMA induced MIMO system which consists of a four-antenna BS to support four single-antenna BPSK MTs. link Tx beamforming from the resulting uplink Rx beamforming according to (7). Assuming the exact reciprocity of the uplink and downlink channels as well

as 2 =U2D , the average BERs, as defined in (13), of

both the MMSE and MBER Tx beamforming schemes so designed are also plotted in Figure 2, in comparison with the average BERs of the MBER and MMSE Rx beamforming designs. As expected, the performance of the proposed downlink Tx beamforming design agreed with that of the uplink Rx beamforming design.

The robustness of the proposed Tx beamforming design was next investigated when the downlink and uplink noise powers or channels were mismatched. In the case of noise mismatching, the downlink noise power was 3 dB more than the uplink noise power. The average BERs of the MBER and MMSE Tx beamforming designs under this uplink and downlink noise power mismatching are plotted in Figure 3, in comparison with the case of equal uplink and downlink noise powers. It can be seen that the 3 dB noise-power mismatching had little influence on the performance of the MBER Tx beamforming scheme but it had some influence on the performance of the MMSE Tx beamforming design. This was expected as the MMSE design is explicitly influen- ced by the noise power while the BER calculation is relatively insensitive to the noise variance estimate used. In the case of channel mismatching, the uplink channel coefficients were = 0.7071 0.7071kA j for 1 4k ,

but the downlink channel coefficients were = 0.6 0.8kA j for 1 . The average BER 4k performance of the MBER and MMSE Tx beamforming designs under this uplink and downlink channel mismatching are plotted in Figure 4, in comparison with the

Figure 3. Average BER performance of the downlink Tx beamforming schemes with both the MMSE and MBER designs for the TDD-SDMA induced MIMO system which consists of a four-antenna BS to support four single-antenna BPSK MTs, when the downlink and uplink noise powers mismatch.

Figure 4. Average BER performance of the downlink Tx beamforming schemes with both the MMSE and MBER designs for the TDD-SDMA induced MIMO system which consists of a four-antenna BS to support four single-antenna BPSK MTs, when the downlink and uplink channels mismatch. case of the exact reciprocity of uplink and downlink channels. From Figure 4, it can be seen that this imperfect CSI had little influence on the MMSE Tx beamforming scheme. It is also seen that the MBER Tx beamforming design was not overly sensitive to the imperfect CSI.

Rank-deficient system. The BS again employed a four-element linear antenna array with half-wavelength element spacing but the number of single-antenna BPSK users was increased to six. The angles of arrival (depar-

ture) for the six users were 2 , , , , 15 10 30 25


S. CHEN ET AL. 150

Figure 5. Average BER performance comparison of the uplink Rx and downlink Tx beamforming schemes with both the MMSE and MBER designs for the TDD-SDMA induced MIMO system which consists of a four-antenna BS to support six single-antenna BPSK MTs.

and , respectively, and the uplink channel co-effi- 36

cients for the six users were ,

. Given the exact reciprocity of the uplink and

downlink channels as well as

= 0.7071 0.7071kA j

2 2=U D

1 k 6

, the average

BERs of both the MBER and MMSE Tx beamforming designs, implemented directly from the corresponding Rx beamforming schemes according to (7), are plotted in Figure 5, in comparison with the average BERs of the MBER and MMSE Rx beamforming designs. For this rank-deficient system, both the MMSE Rx and Tx beamforming solutions exhibited similar high BER floors while both the MBER Rx and Tx beamforming solutions achieved similarly adequate performance.

The robustness of the proposed Tx beamforming design was then investigated again under the scenarios of mismatched downlink and uplink noise powers or channels. In the case of the downlink channel having 3 dB more noise power than the uplink channel, the average BERs of the MBER and MMSE Tx beamforming designs, implemented directly from the corresponding Rx beamforming solutions according to the relation (7), are plotted in Figure 6, in comparison

with the case of 2 =U

Figure 6. Average BER performance of the downlink Tx beamforming schemes with both the MMSE and MBER designs for the TDD-SDMA induced MIMO system which consists of a four-antenna BS to support six single-antenna BPSK MTs, when the downlink and uplink noise powers mismatch.

Figure 7. Average BER performance of the downlink Tx beamforming schemes with both the MMSE and MBER designs for the TDD-SDMA induced MIMO system which consists of a four-antenna BS to support six single-antenna BPSK MTs, when the downlink and uplink channels mismatch.

2D

0.8

. It can be seen that the 3 dB

noise-power mismatching had little influence on performance. In the case of channel mismatching, the uplink channel coefficients were

for , but the downlink channel coefficients were

= 0.7071 0.7071kA j1 6k

= 0.k 6A j for 1 . The average BER

performance of the MBER and MMSE Tx beamforming designs under this uplink and downlink channel mismatching are plotted in Figure 7, in comparison with the case of an identical uplink and downlink channel.

6k

From Figure 7, it can be seen that the MBER Tx beamforming design was not overly sensitive to the imperfect CSI.

5. Conclusions

The downlink MBER transmit beamforming solution has been derived directly based on the uplink MBER receive beamforming design for TDD-SDMA induced MIMO systems. It has been shown that even for rank-deficient TDD systems, where the number of MTs supported is


S. CHEN ET AL.


151

more than the number of transmit antennas available at the BS, the equivalent relationship between the MUD and the corresponding MUT is still valid, if the MBER design is adopted. The proposed MBER transmit beamforming design imposes no computational cost at all at the BS and is capable of achieving good downlink BER performance with the support of low-complexity and high power-efficient MTs. The robustness of this transmit beamforming scheme to the downlink and uplink noise or channel mismatching has been investigated by simulation. 6. References [1] A. J. Paulraj, D. A. Gore, R. U. Nabar and H. Bölcskei,

“An Overview of MIMO Communications—A Key to Gigabit Wireless,” Proceedings of IEEE, Vol. 92, No. 2, 2004, pp. 198-218.

[2] A. Paulraj, R. Nabar and D. Gore, “Introduction to Space-Time Wireless Communications,” Cambridge University Press, Cambridge, 2003.

[3] D. Tse and P. Viswanath, “Fundamentals of Wireless Communication,” Cambridge University Press, Cam- bridge, 2005.

[4] J. H. Winters, “Smart Antennas for Wireless Systems,” IEEE Personal Communications, Vol. 5, 1998, pp. 23-27.

[5] J. Litva and T. K. Y. Lo, “Digital Beamforming in Wire- less Communications,” Artech House, London, 1996.

[6] J. S. Blogh and L. Hanzo, “Third Generation Systems and Intelligent Wireless Networking—Smart Antennas and Adaptive Modulation,” Wiley, Chichester, 2002.

[7] P. Vandenameele, L. van Der Perre and M. Engels, “Space Division Multiple Access for Wireless Local Area Networks,” Kluwer Academic Publishers, Boston, 2001.

[8] D. N. C. Tse and S. V. Hanly, “Linear Multiuser Re- ceivers: Effective Interference, Effective Bandwidth and User Capacity,” IEEE Tranactions on Information Theory, Vol. 45, 1999, pp. 641-657.

[9] S. Chen, A. Livingstone, H.-Q. Du and L. Hanzo, “Adap- tive minimum symbol error rate beamforming assisted detection for quadrature amplitude modulation,” IEEE Transactions on Wireless Communications, Vol. 7, No. 4, 2008, pp. 1140-1145.

[10] S. Chen, L. Hanzo and N. N. Ahmad, “Adaptive Mini- mum Bit Error Rate Beamforming Assisted Receiver for Wireless Communications,” Proceedings of International Conference on Acoustics, Speech, and Signal Processing, Hong Kong, 2003, pp. 640-643.

[11] S. Chen, L. Hanzo, N. N. Ahmad and A. Wolfgang, “Adap- tive MinimumBit Error Rate Beamforming Assisted Receiver for QPSK Wireless Communication,” Digital Signal Processing, Vol. 15, No. 6, 2005, pp. 545- 567.

[12] S. Chen, N. N. Ahmad and L. Hanzo, “Adaptive Mini- mum Bit Error Rate Beamforming,” IEEE Transactions

on Wireless Communications, Vol. 4, No. 2, 2005, pp. 341-348.

[13] D. Yang, L. -L. Yang and L. Hanzo, “Performance of SDMA systems using transmitter preprocessing based on noisy feedback of vector-quantized channel impulse responses,” Proceedings of Vehicular Technology Confe- rence, Dublin, 2007, pp. 2119-2123.

[14] A. M. Khachan, A. J. Tenenbaum and R. S. Adve, “Linear Processing for the Downlink in Multiuser MIMO Systems with Multiple Data Streams,” Proceedings of International Conference on Communications, Istanbul, 2006, pp. 4113-4118.

[15] D. N. C. Tse and P. Viswanath, “Downlink-uplink Duality and Effective Bandwidths,” Proceedings of IEEE 2002 International Symposium Information Theory, Palais de Beaulieu, Switzerland, 2002, p. 52.

[16] M. Schubert and H. Boche, “Iterative Multiuser Uplink and Downlink Beamforming under SINR Constraints,” IEEE Transactions on Signal Processing, Vol. 53, No. 7, 2005, pp. 2324-2334.

[17] M. Schubert and H. Boche, “Solution of the Multiuser Downlink Beamforming Problem with Individual SINR Constraints,” IEEE Transactions on Vehicular Techno- logy, Vol. 53, No. 1, 2004, pp. 18-28.

[18] S. Shi, M. Schubert and H. Boche, “Downlink MMSE Transceiver Optimization for Multiuser MIMO Systems: Duality and Sum-MSE Minimization,” IEEE Transactions on Signal Processing, Vol. 55, No.11, 2007, pp. 5436-5446.

[19] W. Yu, “Uplink-downlink Duality via Minimax Duality,” IEEE Transactions on Information Theory, Vol. 52, No. 2, 2006, pp. 361-374.

[20] L.-L. Yang, “Design Linear Multiuser Transmitters from Linear Multiuser Receivers,” Proceedings of International Conference on Communications, Glasgow, 2007, pp. 5258- 5263.

[21] B. R. VojĈi c Ć and W. M. Jang, “Transmitter Proceed- ings in Synchronous Multiuser Communications,” IEEE Transactions on Communications, Vol. 46, No. 10, 1998, pp. 1346-1355.

[22] A. Hjørungnes and P. S. R. Diniz, “Minimum BER Prefilter Transform for Communications Systems with Binary Signaling and Known FIR MIMO Channel,” IEEE Signal Processing Letters, Vol. 12, No. 3, 2005, pp. 234-237.

[23] R. Irmer, “Multiuser Transmission in Code Division Multiple Access Mobile Communication Systems,” PhD thesis, Technique University of Dresden, Dresden, 2005.

[24] R. Irmer, R. Habendorf, W. Rave and G. Fettweis, “Non- linear Multiuser Transmission Using Multiple Antennas for TDD-CDMA,” Proceedinds of the 6th International. Symposium. Wireless Personal Multimedia Communica- tions, Yokosuka, 2003, pp. 251-255.

[25] S. Tan, “Minimum Error Rate Beamforming Transcei- vers,” PhD thesis, School of Electronics and Computer Science, University of Southampton, Southampton, 2008.

[26] J. Nocedal and S. J. Wright, “Numerical Optimization,” Springer, New York, 1999.

http://ieeexplore.ieee.org/xpl/mostRecentIssue.jsp?punumber=8363





Compromise in CDMA Network Planning

Stephen Hurley, Leigh Hodge School of Computer Science, Cardiff University, Cardiff, UK

Email: [email protected] Received June 7, 2010; revised August 16, 2010; accepted August 20, 2010

Abstract CDMA network planning, for example in 3G UMTS networks, is an important task whether for upgrading existing networks or planning new networks. It is a time consuming, computationally hard, task and generally requires the consideration of both downlink and uplink requirements. Simulation experiments presented here suggest that if time is a major consideration in the planning process then as a compromise only uplink needs to be considered. Keywords: CDMA, Network Planning, Optimization, Simulation

1. Introduction The past decade has seen the emergence of many compu-tational approaches for cellular network site selection, configuration and dimensioning. Many of these contribu-tions have paid attention to planning wide-area FTDMA systems such as second generation GSM where planning is generally carried out using a downlink transmission model and independent criteria for coverage and capacity, e.g [1].

A number of researchers have considered the rather more complex problem of network planning for UMTS networks. Amaldi et al ([2-5]) propose a mathematical programming model which accounts for both the uplink and downlink directions as well as for base station con-figuration issues including location, height, tilt and azimuth. To allow solutions to be sought in reasonable time, approximate solutions are sought via application of the tabu search meta-heuristic. In [6] and [7], Zhang, Yang, et al. propose a mathematical framework for UMTS network planning that considers fast power control, soft handover and pilot signal power in the uplink and down-link directions. Again, solutions are sought via the ap-plication of meta-heuristics (SA and evolutionary SA). Ben Jamaa et al. ([8,9]] propose an approach which em-ploys a multi-objective GA (MOGA) to simultaneously optimize capacity and coverage by adjusting antenna parameters and common channel transmitted powers (antenna locations fixed). A multi-objective fitness func-tion is employed which can consider objectives such as coverage, capacity and cost. The result of the MOGA is a Pareto set of non-dominated solutions.

For third generation systems such as UMTS, the planning

problem is significantly more complex than for FTDMA systems due to the dependency between capacity and coverage. The underlying CDMA protocol requires that on each link, a target signal to noise ratio (SNIR) is maintained, and consequently per-link power allocation is required before user service coverage and cell load can be accurately assessed. However, determining this is non-trivial as one user transmission is seen as interference by all other users, making coverage/capacity evaluation sensitive to other users. Transmission power minimization is important and the real-time UMTS system achieves this by fast power control. However for modeling pur-poses, this is costly to repeatedly simulate because all links are required to frequently re-evaluate their SNIR and adjust power accordingly.

Whitaker R. M. et al describe two efficient heuristic al-gorithms that enable the evaluation of service coverage and cell loading in both the uplink and downlink direc-tions. In this paper, we investigate the application of these heuristics to the problem of cell planning for UMTS networks. The cell planning problem (CPP) is concerned with the selection of antennae from a set of candidate antennae, and the configuration of these an-tennae, such that an optimal configuration is achieved. As for the frequency assignment problem (FAP), the CPP has NP-complete computational complexity. This dictates that exact solutions to the CPP cannot be at-tained in practice. Hence we consider a meta-heuristic optimization approach.

The remainder of this paper is organized as follows. Section 2 describes the model and Section 3 provides a brief overview of the uplink and downlink service cov-erage/load evaluation heuristics. Section 5 outlines the

S. HURLEY ET AL. 153 optimization problem and the meta-heuristic employed. A number of test problems are defined in Section 5 and the results and analysis of applying our optimization ap-proaches to these problems can be seen in Section 6. 2. Model

The uplink (UL) and downlink (DL) dedicated channels and the pilot signal is included in our model. Parameters are described in Table 1, Table 2 and Table 3 and are

defined relative to the link direction under consideration. The terms cell, antenna and transmitter are used

interchangeably when describing aspects of coverage. A number of candidate antenna locations are defined for a given region. A planning/optimization process is em-ployed to select and configure antennae based on defined objectives. Discrete test points from the region are used to sample service coverage. Each test point is a physical position (expressed in two dimensional Cartesian co-ordinates). Two types of test point are defined in our

Table 1. Global parameters.

Symbol Description W CDMA chip rate.

Ri Data rate for service i.

SA

Set of all antennas in the working region.

Sptp

The set of covered pilot test points.

Sstp

The set of service test points.

Ostp

An ordering of the stp.

Table 2. Uplink parameters.

Symbol Description

pULxy Received power from stp x at a cell y.

Iown

Total received power from stp active in cell y. Ioth

Total received power from stp active in cells other than y. Iy Total received power from all active stp.

N Noise power seen at the antennas receiver in an empty cell.

(Eb

/No

)*UL Target threshold for E

b/N

o ratio at an stp for the dedicated UL channel (service dependent).

ηUL,y

Uplink load at cell y.

Table 3. Downlink parameters.

Symbol Description Iown Total power received from serving cell (all links and pilot). Ioth Total power received from all cells other than the serving cell. α Orthogonality Factor.

Pn

Noise power (thermal and equipment) seen at a test point.

pDLxy Power allocated by cell y for stp x as received at stp x.

ppilotxy Pilot power from cell y as received at stp x.

(Ec/I

o)pilot

Target threshold for pilot Ec/I

o ratio.

(Eb

/No

)*UL Target threshold for E

b/N

o ratio at an stp for a dedicated DL channel (service dependent).

ηDL,y

Downlink load at cell y.

PtxTotaly

Total of allocated transmit powers in cell y.

Ptxmaxy Maximum transmit capability of cell y.

ηpilot,y

Proportion of Ptxmaxy allocated for pilot signal at cell y.


S. HURLEY ET AL.154 model: service test points (stp) and pilot test points (ptp). The ptp are used to assess pilot signal quality. At an stp, quality of both UL and DL dedicated channels are as sessed for a particular service, which is defined prior to evaluation.

2.1. Test Point Coverage and Cell Load

The pilot signal is transmitted at a proportion pilot,y of the maximum cell power. A ptp x is served by antenna y when the received energy per chip relative to the total spectral density Ec/Io at least meets the target Ec/Iopilot. Letting Iy = Iown + Ioth, then x is served if and only if:

pilotxy

c o piloty

pE /I

N+I (1)

An stp is covered in a particular link direction if energy per bit relative to spectral noise density (Eb/No) at least meets the required target threshold. For an stp x connected to antenna y, x is UL covered if and only if:

UL

*xy

ULi y xy

pWEb/No

R I p N

(2)

In the downlink, for an stp x and serving antenna y, x is DL covered if and only if:

DL

*xy

DLi own oth n

pWEb/No

R I (1-α)+I +p (3)

There are various ways in which cell loading can be assessed. Wideband power-based measurement is used in this model because it directly identifies the resources being allocated. The downlink load at cell y is estimated by:

yDL,y

y

PtxTotalη =

Ptxmax (4)

while the uplink load at cell y is estimated by:

yUL,y

y

Iη =

I +N (5)

Note that a ptp’s ability to be served depends on downlink cell load. Consequently a ptp is covered if and only if it is served when all cells y are operating at max-

imum permitted downlink load . Covered ptp can

see the pilot signal independent of traffic and are collec-tively denoted Sptp.

maxDL,y

To ensure that an stp can see the pilot signal, it is required that Sstp Sptp. A list Ostp of the set Sstp is also required to specify the order in which stp are prioritized for admission. The ordering is defined based on the received signal strength from the best serving antenna with those with the strongest signal given priority.

3. Evaluation Heuristics

Calculating off-line transmission power for target Eb/No

attainment on a link requires knowledge of interference levels or equivalently cell loads. However, interference/cell loads depend on per-link transmission powers. This dependency has led to the analytical characterization of the problem [11]. We employ an algorithmic approach which initially over-estimates interference/cell loading and then uses a feedback mechanism to iteratively update and reduce the conservative error. When this feedback mechanism is applied, the heuristic can converge to a state where inaccuracy in power allocation and cell loading is negligible. From this, stp coverage and cell loads can be directly obtained.

Detailed discussion of the uplink and downlink evalu-ation heuristics used here can be found in [10]. 4. Optimization Problem It is assumed that for optimization, the objective is to select/activate and configure (where appropriate) a subset of antennae from the set of candidate antennae such that coverage is maximized for a specified number of active transmitters. After experimentation with a number of meta-heuristics it was determined that tabu search (TS) was the most effective approach for this optimization problem. The TS algorithm employed is summarized in Figure 1. A detailed description of the TS meta-heuristic can be found in [12].

The operation of our tabu search approach can be cha-racterized by the following components: starting con-figuration; moves; evaluation type and cost function. 4.1. Starting Configuration The starting configuration can impact on the final con-figuration achieved by the TS. Having investigated a number of starting configurations (i.e., all transmitters inactive, all transmitters active, random transmitters active and Halton configuration - approximately random uniformly distributed) it has been shown that whilst the Generate starting configuration. Evaluate cost of starting configuration. Set best_cost_so_far = 0. Initialise memory structures. FOR i = 0 to max_iteratations DO Evaluate all possible moves. Sort moves on cost (prioritization). Accept first move where move is non-tabu or is tabu but meets aspiration criteria. Update memory structures. IF cost of configuration after move improves upon best_cost_so_far THEN Set best_configuration = current configuration. Set best_cost_so_far = cost of current configuration. END IF END FOR

Figure 1. Generic tabu search algorithm


S. HURLEY ET AL. 155 effectiveness of each starting configuration is dependent on the problem scenario and other parameter settings, starting with all transmitters inactive leads to the best solutions in general.

4.2. Moves

A range of different moves are employed by the TS. At each iteration of the TS, the impact of each of the available moves is evaluated by applying the move to each candidate antenna in order to determine the best possible move at that instance. The quality of each move is determined by the cost function. A range of moves have been imple-mented from which a subset of moves can be selected for evaluation: • Activate an inactive transmitter i.e. make operational. • Deactivate an active transmitter i.e. shut down. • Swap transmitters: • Deactivate an active transmitter and activate a randomly selected inactive transmitter . • Activate an inactive transmitter and deactivate a randomly selected active transmitter . • Determine the best azimuth for a transmitter - the azimuth for a given transmitter is varied (controlled by azimuth_increment) such that all available azimuth configurations in the sector are evaluated. The configuration with the best cost can then be determined. • Determine the best tilt for a transmitter - the tilt for a given transmitter is varied (controlled by tilt_increment) such that all available tilt configurations in the range tilt_max to tilt_min are are evaluated. The configuration with the best cost can then be determined.

4.3. Evaluation Type

When evaluating moves, cost functions are employed in conjunction with an evaluation type. The evaluation type defines the evaluation heuristic used to determine service (i.e. uplink or downlink evaluation heuristic) and any constraints on the feedback mechanism employed to determine total load. In its least constrained form the iterative feedback mechanism repeats until the variation in the load is less than a predefined threshold. Iterating to convergence may require much iteration. This is time consuming and additional iterations achieve decreasing returns. As a result, in order to achieve an acceptable runtime for the TS (which need to perform large numbers of evaluations) the number of feedback iterations is con-strained when employed for evaluating moves, only run-ning to convergence for the final TS solution. 4.4. Cost Function TS requires that a cost is associated with problem con-figurations such that an optimal or near optimal con-figuration can be sought. A weighted cost function is

employed to enable the following objectives to be con-sidered:

1) Meet the constraint on the number of transmitters. 2) Maximise coverage. 3) Favour configurations with lower total loads. The tabu search seeks to minimize the total cost of a

configuration, defined as:

total_cost = (covg_cost *wc ) + (actv_cost * wa) +

（ld_cost * wl) (6)

where wc, wa and wl are weightings for coverage cost,

active cost and load cost respectively. Coverage cost is defined as

covg_cost = 100 – coverage (7) where coverage is the percentage of stp that are covered in the downlink or uplink direction.

Active cost is defined as actv_cost = trans_thrs – active_trans (8)

where trans_thrs is the desired number of active trans-mitters and active_trans is the number of active trans-mitters.

Load cost is defined as

ld_cost= ηDL (9)

for all active transmitters in the downlink direction and

ld_cost= ηUL (10)

for all active transmitters in the uplink direction. It should be noted that as some objectives are competing

(e.g. coverage and total load) it may not be possible to determine the configuration which exhibits the optimal trade-off between objectives. This would require a more complex (and time consuming) multi-objective approach. 5. Experimentation The purpose of experimentation is to compare the per-formance of network configurations generated using up-link and downlink evaluation heuristics. Performing optimization for both link directions is time consuming. Consequently, it would be useful if we could identify an optimization configuration that provides a good trade-off between optimizing for uplink and for downlink, i.e., a single approach that produces network configurations that perform well in both link directions. Whilst this trade-off may not be acceptable when producing final configurations, it could be beneficial in preliminary stages of network planning where some accuracy can be traded for the decreased evaluation time associated with optimizing for a single link direction.

5.1. Test Problems

All experiments consider a 3 km × 3 km transmission region containing 36 directional candidate antennae lo-


S. HURLEY ET AL.


156

• DL optimize and evaluate configuration for DL and UL percentage coverage/service, and load.

cated at 12 uniformly distributed sites. Eight different problem scenarios have been considered as summarized in Table 4. Each scenario consists of a number of uni-formly distributed stp with defined service requirements. Different scenarios have been generated by varying the number of stp considered and the distribution of services over these stp. Scenarios 5a, 5b and 5c have the same number of stp and number of stp with each service re-quirement, but a different distribution of these services over the stp. Signal attenuation is defined by the Hata path loss model.

• UL optimize and evaluate configuration for DL and UL percentage coverage/service, and load.

This enables us to determine how well configurations optimized in one link direction perform with respect to the opposite link direction. Further analysis is undertaken to determine:

1) Coverage Difference (’Covg Diff’ column in the tables) - the difference between uplink and downlink coverage for each instance of a problem scenario. This gives an indication of how optimizing for one link direc-tion impacts on the other.

5.2. TS Configuration

2) Maximum DL Coverage Difference (’Max DL Covg Diff’ column) - for each instance of a problem scenario the maximum percentage DL coverage (’Max % DL Covg’ column) for all optimization approaches is identified i.e. the maximum percentage DL coverage obtained from the DL or UL optimized AS or ASBTBA method. From this, the Maximum DL Coverage Difference is deter-mined (by subtracting the coverage obtained for a problem instance from the maximum percentage DL coverage value), i.e. this indicates how well an optimization ap-proach performs with respect to the best result. Conse-quently, this value gives an indication of which optimi-zation approach performs best for each problem instance, and over all problem instances (based on the mean val-ue).

The tabu search was constrained to run for a maximum of 400 iterations and terminate after 50 iterations in which there is no improving move.

For each scenario and evaluation heuristic, two sets of moves are employed: 1) Activate/deactivate transmitter and swap transmitter activity (AS). 2) Activate/deactivate transmitter, swap transmitter activity, best tilt and best azimuth (ASBTBA)1.

These sets of moves were selected in order to investigate the impact of tuning the configuration of the antennae. 6. Results

3) Maximum UL Coverage Difference (’Max UL Covg Diff’ column) - as Maximum DL Coverage Difference, but in the uplink direction.

In this section we present the results of optimization for each problem scenario. For each problem scenario, a number of problem instances are considered, each with different constraints on the maximum number of antennae allowed. Four optimization approaches are considered:

6.1. Sample Results

1) DL optimisation heuristics with AS Due to the volume of results generated from experimen-tation, only a subset of results is presented here2. For Problem 1, the results of all optimization approaches i.e. AS and ASBTBA are included (see Tables A1 to A4 in Appendix A). For other problems, the results for optimiza-tion approach ASBTBA are presented only (see Tables

2) DL optimisation heuristics with ASBTBA 3) UL optimisation heuristics with AS 4) UL optimisation heuristic with ASBTBA On completion of the TS the resulting configuration is

evaluated for the opposite link direction, i.e.:

Table 4. Problem scenarios.

No. stp assigned per service type

Scenario No. stp Pilot 12.2 kbps 64kbps 144 kbps 384 kbps Total Capacity Req (kbps)

1 441 100 220 44 44 23 20,668

2 441 147 0 294 0 0 18,816

3 441 392 0 0 49 0 18,816

4 961 630 220 44 44 23 20,668

5a, 5b, 5c 961 299 440 88 88 46 41,336

6 3721 1116 1675 372 372 186 169,235

1For best tilt and best azimuth moves, an increment of 1 degree is ap-plied.

2A complete set of results can be found in Appendix B at: www.cs.cf.ac.uk/bounds/documentation.htm

S. HURLEY ET AL. 157

Table 5. Analysis summary.

Link Direction/ Optimisation

Highest Mean DL Covg

Highest Mean UL Covg

Lowest Mean Covg

Lowest Mean Max DL Covg

Lowest Mean Max UL Covg

Lowest Combined Mean Covg

Method Difference Differnece Difference Difference DL AS 1 3 1 DL ASBTBA 2, 3, 4, 5a, 2, 3, 4, 5a, 5b, 5c, 6 5b, 5c, 6 UL AS 3 3 2 UL ASBTBA 1,2,3,4,5a, 1, 2, 4, 5a, 1, 2, 3, 4, 5a, 1, 3, 4, 5a, 5b, 5c, 6 5b, 5c, 6 5b, 5c, 6 5b, 5c, 6

A5 to A18 in Appendix A) as these performs best in gen-eral (see Section 6.2). 6.2. Analysis of Results and Conclusions

In order to compare the effectiveness of the different optimization approaches, we examine for each scenario which approach gives the best performance with respect to a number of metrics:

1) highest mean DL % coverage (’Highest Mean DL Covg’ column);

2) highest mean UL % coverage (’Highest Mean UL Covg’ column);

3) lowest mean coverage difference (’Lowest Mean Covg Difference’ column), i.e. for each scenario which method gives the lowest Coverage Difference value;

4) lowest mean maximum DL coverage difference (Lowest Mean Max DL Covg Difference’ column) i.e. for each scenario which method gives the lowest mean Maximum DL Coverage Difference value;

5) lowest mean maximum UL coverage difference (Lowest Mean Max UL Covg Difference’ column), i.e. as above but defined for UL, and

6) lowest combined (i.e., DL and UL) mean coverage difference (’Lowest Combined Mean Covg Difference’ column) i.e. for each scenario which method gives the lowest value when adding the maximum DL and UL coverage difference means.

The results of this summary analysis can be seen in Table 5 which indicates the optimization method/ sce-nario combination that gives the best result for each of the above six metrics. As expected, the results show that including antenna configuration moves during optimiza-tion leads to increased coverage3.The results show that the DL ASBTBA optimization approach generally leads to the best downlink coverage and therefore the best i.e. lowest mean maximum DL coverage difference. Simi-larly the UL ASBTBA approach is consistently the best in terms of uplink.

Furthermore UL ASBTBA leads to the best (lowest) combined mean coverage difference i.e. for scenarios 1,3,4,5a,5b,5c and 6. This indicates that although opti-mization using the UL ASBTBA approach does not lead

to the best levels of downlink coverage (though it is competitive in many places as illustrated by low mean maximum DL coverage difference values) it does result in the best overall combined mean coverage difference indicating that it performs better in terms of DL coverage than DL optimization does in terms of UL coverage. As a result, where time for planning a network is limited, the experimental results presented here suggest that a com-promise in many cases is to optimize for uplink only rather than optimizing in both the uplink and downlink directions. 7. Acknowledgements

This work was funded by the UK’s Engineering and Physical Science Research Council. 8. References

[1] S. Hurley, “Planning Effective Cellular Mobile Radio Networks,” IEEE Transactions on Vehicular Technology, Cardiff, Vol. 51, No. 2, March 2002, pp 243-253.

[2] E. Amaldi, A. Capone, F. Malucelli and F. Signori, “Optimization Models and Algorithms for Downlink UMTS Radio Planning,” Proceedings of Wireless Communications and Networking, Milan, Vol. 2, March 2003, pp. 827-831.

[3] E. Amaldi, A. Capone, F. Malucelli and F. Signori, “A Mathematical Programming Approach for W-CDMA Radio Planning with Uplink and Downlink Constraints,” Proceedings of the Vehicular Technology Conference, 2003, Vol. 2, October 2003, pp. 806-810.

[4] E. Amaldi, A. Capone, F. Malucelli and F. Signori, “Radio Planning and Optimization of W-CDMA Sys- tems,” Lecture Notes in Computer Science - Personal Wireless Communications, Vol. 2775, 2003, pp. 437-447.

[5] E. Amaldi, A. Capone and F. Malucelli, “Radio Planning and Coverage Optimization of 3G Cellular Networks,” Wireless Networks, Vol. 14, No. 4, August 2008, pp. 435-447.

[6] J. Zhang, J. Yang, M. E. Aydin, and J. Y. Wu, “Mathematical Modelling and Comparisons of Four Heuristic Optimization Algorithms for WCDMA Radio Network Planning,” Proceedings of the International Conference on Transparent Optical Networks, 2006, Vol. 3, June 2006, pp. 253-257.

3This is also confirmed by Amaldi et al in [2] who have also shown that it is preferable to simultaneously optimize antenna location and con-figuration than to do so separately.


S. HURLEY ET AL. 158 [7] J. Yang, M. E. Aydin, J. Zhang and C. Maple, “UMTS

Base Station Location Planning: a Mathematical Model and Heuristic Optimisation Algorithms,” Communications, Vol. 1, No. 5, October 2007, pp. 1007-1014.

[8] S. B. Jamaa, Z. Altman, J. M. Picard and B. Fourestie, “Multi-objective Strategies for Automatic Cell Planning of UMTS Networks,” Proceedings of the Vehicular Technology Conference, Vol. 4, May 2004, pp. 2420-2424.

[9] S. B. Jamaa, Z. Altman, J. M. Picard and B. Fourestie, “Combined Coverage and Capacity Optimisation for UMTS Networks,” Telecommunications Network Strategy and Planning Symposium, Moulineaux, June 2004, pp.

175-178.

[10] R. M. Whitaker, S. Allen and S. Hurley, “Efficient Offline Coverage and Load Evaluation for CDMA Network Modeling,” IEEE Transactions on Vehicular Technology, Vol. 58, No. 7, August 2009, pp 3704-3712.

[11] L. Mendo and J. M. Hernando, “On Dimension Reduc- tion for Next Generation Microcellular Networks,” IEEE Transactions on Communications, Vol. 49, No. 2, 2001, pp. 243-248

[12] F. W. Glover and M. Laguna, “Tabu Search,” Springer, German, 1997.


S. HURLEY ET AL. 159 Appendix A

Table A1. Scenario 1 - DL Optimised (AS).

Num Num DL % DL UL % UL Covg Max % Max DL Max % Max UL Tx Sites Covg Load Covg Load Diff DL Covg Covg Diff UL Covg Covg Diff 15 12 97.7324 4.91592 93.424 6.92558 4.3084 97.9529 0.2205 97.5057 4.0817 13 12 96.5986 3.85497 92.9705 6.77027 3.6281 97.5057 0.9071 96.6916 3.7211 11 11 95.9184 4.51131 88.8889 5.65139 7.0295 96.6508 0.7324 93.424 4.5351 9 9 93.424 3.67729 80.7256 4.75283 12.6984 93.424 0 89.5692 8.8436 7 7 90.0227 3.61884 73.2426 4.2 16.7801 90.0227 0 85.6208 12.3782 Mean 94.73922 85.85032 8.8889 0.372 6.71194

Table A2. Scenario 1 - DL Optimised (ASBTBA).

Num Num DL % DL UL % UL Covg Max % Max DL Max % Max UL Tx Sites Covg Load Covg Load Diff DL Covg Covg Diff UL Covg Covg Diff 15 12 97.9529 5.45044 93.8776 6.90332 4.0753 97.9529 0 97.5057 3.6281 13 12 97.5057 5.54612 90.7029 5.94145 6.8028 97.5057 0 96.6916 5.9887 11 11 95.2381 4.56179 89.5692 5.69875 5.6689 96.6508 1.4127 93.424 3.8548 9 9 92.7438 4.30398 85.7143 4.98792 7.0295 93.424 0.6802 89.5692 3.8549 7 7 90.0227 3.69683 79.8186 3.9371 10.2041 90.0227 0 85.6208 5.8022 Mean 94.69264 87.93652 6.75612 0.41858 4.62574

Table A3. Scenario 1 - UL Optimised (AS).

Num Num DL % DL UL % UL Covg Max % Max DL Max % Max UL Tx Sites Covg Load Covg Load Diff DL Covg Covg Diff UL Covg Covg Diff 15 11 97.2789 4.23953 97.5057 8.27713 0.2268 97.9529 0.674 97.5057 0 13 10 95.9184 3.75661 95.6916 7.2764 0.2268 97.5057 1.5873 96.6916 1 11 11 94.7846 3.8032 93.424 6.22613 1.3606 96.6508 1.8662 93.424 0 9 9 91.61 3.14089 89.5692 5.12837 2.0408 93.424 1.814 89.5692 0 7 7 87.9819 2.88475 81.1791 4.09059 6.8028 90.0227 2.0408 85.6208 4.4417 Mean 93.51476 91.47392 2.13156 1.59646 1.08834

Table A4. Scenario 1 - UL Optimised (ASBTBA).

Num Num DL % DL UL % UL Covg Max % Max DL Max % Max UL Tx Sites Covg Load Covg Load Diff DL Covg Covg Diff ULCovg Covg Diff 15 12 97.5057 4.39257 97.5075 8.29146 0.0018 97.9529 0.4472 97.5075 0 13 11 96.3719 4.33431 96.6916 7.24909 0.3197 97.5057 1.1338 96.6916 0 11 9 96.6508 3.28007 93.1973 6.17624 3.4535 96.6508 0 93.424 0.2267 9 9 91.1565 2.98996 89.1156 5.4 2.0409 93.424 2.2675 89.5692 0.4536 7 7 89.1156 3.37209 85.2608 4.08703 3.8548 90.0227 0.9071 85.6208 0.36 Mean 94.1601 92.35456 1.93414 0.95112 0.20806


Num Num DL % DL UL % UL Covg Max % Max DL Max % Max UL Tx Sites Covg Load Covg Load Diff DL Covg Covg Diff UL Covg Covg Diff 15 12 92.0635 7.76091 91.3832 8.12563 0.6803 92.0635 0 93.424 2.0408 13 12 89.7959 6.94523 86.6213 7.26294 3.1746 89.7959 0 89.3424 2.7211 11 11 85.7143 6.02062 82.7664 6.6 2.9479 85.7143 0 83.22 0.4536 9 9 79.3651 4.92076 75.737 5.10716 3.6281 79.3651 0 77.3243 1.5873 7 7 71.4286 3.97083 68.4807 4.1063 2.9479 71.4286 0 68.9342 0.4535 Mean 83.67348 80.99772 2.67576 0 1.45126


Num Num DL % DL UL % UL Covg Max % Max DL Max % Max UL Tx Sites Covg Load Covg Load Diff DL Covg Covg Diff UL Covg Covg Diff 15 12 90.4762 6.80249 93.424 8.68051 2.9478 92.0635 1.5873 93.424 0 13 12 87.9819 5.60654 89.3424 7.8 1.3605 89.7959 1.814 89.3424 0 11 11 83.22 5.42435 83.22 6.48371 0 85.7143 2.4943 83.22 0 9 9 78.2313 4.23473 77.3243 5.3544 0.907 79.3651 1.1338 77.3243 0 7 7 70.9751 3.65601 68.9342 4.16268 2.0409 71.4286 0.4535 68.9342 0 Mean 82.1769 82.44898 1.45124 1.49658 0


S. HURLEY ET AL. 160


Num Num DL % DL UL % UL Covg Max % Max DL Max % Max UL Tx Sites Covg Load Covg Load Diff DL Covg Covg Diff UL Covg Covg Diff 15 12 98.8662 2.15203 99.5465 7.59356 0.6803 99.5465 0.6803 100 0.4535 13 12 99.093 2.55291 99.093 5.97961 0 99.093 0 99.7732 0.6802 11 10 98.6395 2.67339 98.1859 5.05613 0.4536 98.6395 0 99.093 0.9071 9 9 97.9592 3.08148 97.2789 4.58512 0.6803 97.9592 0 97.9592 0.6803 7 6 96.3719 2.44903 95.9184 3.82552 0.4535 96.5986 0.2267 96.5986 0.6802 Mean 98.18596 98.00454 0.45354 0.1814 0.68026


Num Num DL % DL UL % UL Covg Max % Max DL Max % Max UL Tx Sites Covg Load Covg Load Diff DL Covg Covg Diff UL Covg Covg Diff 15 12 99.3197 2.45275 100 6.94763 0.6803 99.5465 0.2268 100 0 13 11 99.093 2.50739 99.7732 6.52997 0.6802 99.093 0 99.7732 0 11 10 98.1859 2.21099 99.093 5.88246 0.9071 98.6395 0.4536 99.093 0 9 9 97.5057 2.25701 97.9592 5.01653 0.4535 97.9592 0.4535 97.9592 0 7 7 96.5986 2.44812 96.5986 4.02627 0 96.5986 0 96.5986 0 Mean 98.14058 98.6848 0.54422 0.22678 0


Num Num DL % DL UL % UL Covg Max % Max DL Max % Max UL Tx Sites Covg Load Covg Load Diff DL Covg Covg Diff UL Covg Covg Diff 15 12 98.231 3.48392 95.0052 5.71615 3.2258 98.3351 0.1041 98.231 3.2258 13 12 98.4391 4.22513 96.0458 5.90235 2.3933 98.4391 0 97.9188 1.873 11 10 97.8148 4.23715 92.82 4.91646 4.9948 97.8148 0 96.7742 3.9542 9 9 97.0864 4.48578 92.5078 5.04474 4.5786 97.0864 0 95.3174 2.8096 7 7 95.3174 3.85111 89.4901 4.2 5.8273 95.3174 0 93.3403 3.8502 Mean 97.37774 93.17378 4.20396 0.02082 3.14256


Num Num DL % DL UL % UL Covg Max % Max DL Max % Max UL Tx Sites Covg Load Covg Load Diff DL Covg Covg Diff UL Covg Covg Diff 15 11 98.0229 3.93876 98.231 7.83294 0.2081 98.3351 0.3122 98.231 0 13 11 98.0229 4.08886 97.9188 7.36299 0.1041 98.4391 0.4162 97.9188 0 11 10 97.1904 3.7737 96.7742 6.17889 0.4162 97.8148 0.6244 96.7742 0 9 9 95.6296 3.40389 95.1093 5.27193 0.5203 97.0864 1.4568 95.3174 0.2081 7 7 94.589 3.03997 93.3403 4.16951 1.2487 95.3174 0.7284 93.3403 0 Mean 96.69096 96.27472 0.49948 0.7076 0.04162

Table A11. Scenario 5a - DL Optimised (ASBTBA).


Table A12. Scenario 5a - UL Optimised (ASBTBA).



S. HURLEY ET AL.


161

Table A13. Scenario 5b - DL Optimised (ASBTBA).


Table A14. Scenario 5b - UL Optimised (ASBTBA).

Num Num DL % DL UL % UL Covg Max % Max DL Max % Max UL Tx Sites Covg Load Covg Load Diff DL Covg Covg Diff UL Covg Covg Diff 30 12 93.7565 14.4344 94.589 15.9992 0.8325 94.693 0.9365 94.7971 0.2081 28 12 93.6524 7.0646 94.1727 16.0077 0.5203 95.2133 1.5609 94.1727 0 26 12 93.8606 6.80798 94.849 14.5496 0.9884 94.7971 0.9365 94.849 0 24 12 93.1322 7.31611 92.7159 13.1161 0.4163 94.1727 1.0405 92.7159 0 22 11 91.3632 11.3738 90.9469 12.3443 0.4163 93.7565 2.3933 90.9469 0 Mean 93.15298 93.45470 0.63476 1.37354 0.04162

Table A15. Scenario 5c - DL Optimised (ASBTBA).

Num Num DL % DL UL % UL Covg Max % Max DL Max % Max UL Tx Sites Covg Load Covg Load Diff DL Covg Covg Diff UL Covg Covg Diff 30 12 95.2133 8.6136 91.155 18 4.0583 95.2133 0 94.7971 3.6421 28 12 94.7971 8.33628 92.5078 13.8027 2.2893 94.7971 0 94.9011 2.3933 26 12 94.4849 8.44079 91.155 12.5475 3.3299 94.4849 0 93.6524 2.4974 24 12 94.693 9.255303 89.8023 11.7329 4.8907 94.693 0 92.4037 2.6014 22 12 94.0687 9.59018 87.0968 10.4895 6.9719 94.0687 0 92.6119 5.5151 Mean 94.65140 90.34338 4.30802 0 3.32986

Table A16. Scenario 5c - UL Optimised (ASBTBA).



Num Num DL % DL UL % UL Covg Max % Max DL Max % Max UL Tx Sites Covg Load Covg Load Diff DL Covg Covg Diff UL Covg Covg Diff 30 12 76.5117 16.5578 53.5609 14.0817 22.9508 76.5117 0 60.602 7.0411 28 12 75.6248 15.7692 52.3246 13.2735 23.3002 75.6248 0 58.8283 6.5037 26 12 75.6517 14.9694 53.9371 13.9292 21.7146 75.6517 0 57.4308 3.4937 24 12 78.8723 13.9065 51.5453 12.5168 27.327 78.8723 0 57.4577 5.9124 22 12 73.4211 12.9731 50.1478 11.526 23.2733 73.4211 0 54.9583 4.8105 Mean 76.01632 52.30314 23.71318 0 5.55228


Num Num DL % DL UL % UL Covg Max % Max DL Max % Max UL Tx Sites Covg Load Covg Load Diff DL Covg Covg Diff UL Covg Covg Diff 30 12 74.308 15.7713 60.602 17.7095 13.706 76.5117 2.2037 60.602 0 28 12 73.3405 14.5942 58.8283 16.0287 14.5122 75.6248 2.2843 58.8283 0 26 12 71.6743 14.0077 57.4308 15.2955 14.2435 75.6517 3.9774 57.4308 0 24 12 71.5937 13.1524 57.4577 14.4 14.136 78.8723 7.2786 57.4577 0 22 12 69.9543 12.3729 54.9583 13.0276 14.996 73.4211 3.4668 54.9583 0 Mean 72.17416 57.85542 14.31874 3.84216 0



A Novel Black Box Based Behavioral Model of Power Amplifier for WCDMA Applications

Amandeep Singh Sappal, Manjeet Singh Patterh, Sanjay Sharma2 1University College of Engineering, Punjabi University Patiala, Punjab, India,

2Dept. of ECE, Thapar University, Punjab, India E-mail: [email protected]

Received June 28, 2010; revised August 10, 2010; accepted August 15, 2010

Abstract In this paper, Black Box approach is presented for behavioral modeling of a non linear power amplifier with memory effects. Large signal parameters of a Motorola LDMOS power amplifier driven by a WCDMA signal were extracted while taking into considerations the power amplifier’s bandwidth. The proposed model was validated based on the simulated data. Some validation results are presented both in the time and frequency domains, using WCDMA signal. Keywords: Behavioral Model, Black Box, Power Amplifiers

1. Introduction The introduction of the third generation UMTS, based on WCDMA technology, is a further step towards satisfying the ever increasing demand for data/internet services. 3G is quickly moving on to 3.5G, 3.9G, and 4G and is changing the way the world communicates. The evolu-tion of wireless technologies including CDMA2000, GPRS, EGPRS, WCDMA, HSDPA and 1xEV, allow development of new wireless devices that combine voice, internet, and multimedia services. In the future GSM and other parallel 2G systems are likely to be replaced with 3G and beyond, and the bands that today are used for GSM will then be used for WCDMA and other standards. WCDMA in the 900 MHz band is a cost effective way to deliver nationwide high-speed wireless coverage .This evolution has brought new requirements on the RF parts of the transceivers, especially the Power Amplifier (PA). Thus the simulation of PA circuits is becoming a very important issue in nowadays communication scenarios.

Due to broadband nature of signals, frequency-dependent behavior of PA is encountered, i.e., memory effects. To accurately model a PA, we have to take into account both nonlinearities and memory effects. Several works have recently been published proposing behavioral models and extraction procedures for envelope behavioral model simulation [1-4]. The Volterra series has been used by several researchers to describe the relationship between the input and the output of a power amplifier with memory effects [1]. However, high computational com-

plexity makes methods of this kind impractical in some real cases, e.g., modeling a PA with strong nonlinearities and/or with long-term memory effects. This is because the number of coefficients to be estimated in the model increases exponentially with the degree of nonlinearity and with the memory length of the system. To overcome the modeling complexity, various model-order reduction approaches have been proposed to simplify the Volterra model structure [5-15]. Although these simplified models have been employed to characterize PAs with reasonable accuracy under certain conditions, there is no systematic way to verify if the model structure chosen is truly appropriate to the PA under study. In this paper the principle of a novel approach, called ‘Black Box’ has been presented. The Black Box model is directly derived from the topology of the amplifier.

The variables used to describe the signals at both ports are the classical incident and scattered voltage waves [16], typically defined in a characteristic impedance of 50 Ω, together with the dc current and voltage biasing parameters.

Input Signal

Figure 1. Parameter estimation of Black Box model parameters.

Device under Test (DUT)

Bias Voltages

Parameter Extraction

A. S. SAPPAL ET AL. 163

2. Description of Black Box Model

The Black Box model is derived directly from the circuit topology of the PA. The transistors can be considered to be two port non-linear networks which can be modeled in terms of nonlinear scattering parameters. If

1, 2 and 1, 2 represents the incident and reflected waves respectively. Using first order Taylor series expansion, the scattering parameter model of PA can be written as [16]

ijc a a b b

12

22

*11 1 12 1 11 1 1

*2 2 221 1 22 1 1

0

0

S a S a S ab a a

b a aS a S a S a

(1)

ij iS a

12

22

1

2

11 1 11 12 1 12 1

221 1 21 22 1 22

*1 12 1

*21 22

. .

. .

0 .

0

b

b

S a c S a c a

aS a c S a c

S a c a

aS a c

(2)

ijc represents the normalized frequency function.

3. Valediction of the Proposed Model

The proposed model is implemented in Agilent ADS at a frequency of 1950 MHz. The input signal power is varied from 0 to 20 dB. The model is implemented for Motorola LDMOS PA circuit available in Agilent library as shown in Figure 2 and the measurement setup is shown in Figure 3. The data file has been extracted for

represents non-linear scattering parameters as a

function of input waves. But in real practice scattering pa-rameters are also found to be function of PA band width. So in order to consider the effect of PA bandwidth also (1) is modified as [17]

Figure 2. Motorola PA circuit topology used to validate the proposed model.

CIRCUIT BIAS SETUP

Motorola_PAX2

Ldrainline2=1161 milLdrainline1=540 milLgateline2=610 milLgateline1=1425 mil

AmplifierS2D_SetupX1

SSFreq_Step=-1.0 HzSSFreq_Stop=-1.0 HzSSFreq_Start=-1.0 HzPin_Step=2 _dBPin_Stop=20 _dBmPin_Start=0 _dBmFreq_Step=0.1 GHzFreq_Stop=1950 MHzFreq_Start=1950 MHzOrder=11Filename="Motorola_PA.s2d"

S2DSetup

Agi le nt Tec hnologies

outin

ParamSweepSweep_BiasL

Step=0.1Stop=2.1Start=1.9SimInstanceName[6]=SimInstanceName[5]=SimInstanceName[4]=SimInstanceName[3]=SimInstanceName[2]=SimInstanceName[1]="Sweep_BiasU"SweepVar="BiasL"

PARAMETER SWEEP

ParamSweepSweep_BiasU

Step=0.5Stop=6.3Start=5.3SimInstanceName[6]=SimInstanceName[5]=SimInstanceName[4]=SimInstanceName[3]=SimInstanceName[2]=SimInstanceName[1]="X1.HB1"SweepVar="BiasU"

PARAMETER SWEEP

V_DCSRC4Vdc=BiasL

V_DCSRC3Vdc=BiasU

VARVAR1

BiasL=2.0BiasU=5.8

EqnVar

Figure 3. The setup for measurement of parameters.


A. S. SAPPAL ET AL. 164

Figure 4. Comparison of gain compression (AM/AM characteristics).

Figure 5. Comparison of phase compression (AM/PM characteristics).

-6 -4 -2 0 2 4 6-8 8

-110

-100

-90

-80

-70

-60

-50

-40

-30

-20

-120

-10

Frequency (MHz)

Pow

er (d

Bm)

Reference Signal Spectrum

Figure 6. Input spectrum of the signal.

the proposed model and for the valediction of the proposed model; the results are compared with the results of PA circuit topology. Gain in dB and phase in degree is plotted against the input power in dBm as shown in Figures 4 and 5. Results validate the proposed model at the applied frequency.

-6 -4 -2 0 2 4 6-8 8

-100

-90

-80

-70

-60

-50

-40

-30

-20

-110

-10

Frequency (MHz)

Pow

er (d

Bm)

Distorted Signal Spectrum

Figure7. Output spectrum of the signal.

-1.0 -0.5 0.0 0.5 1.0-1.5 1.5

-1.0

-0.5

0.0

0.5

1.0

-1.5

1.5

In-Phase

Qua

drat

ure

Reference Constellation

Figure 8. Constellations of the reference signal.

Figure 9. Constellations of the reference signal (distorted).


A. S. SAPPAL ET AL.


165

4. Measurement of Parameters on A WCDMA Signal

The proposed model was also tested for measurements on a WCDMA signal centered on 1950 MHz. Input and output spectrum of the input signal and the output signal were measured.

Upper channel Adjacent Channel Leakage Ratio (ACLR) for the reference signal is -52.476 and for the distorted signal is -34.525. Also Lower channel Adjacent Channel Leakage Ratio (ACLR) for the reference signal is -52.717 and for the distorted signal is -34.608. Con-stellations of the reference signal and the distorted signal are also plotted as show in Figures 8 and 9 respectively.

The peak value of Error vector magnitude (EVM) of the reference signal was calculated as 35.13% and 53.21% respectively.

5. Conclusions

A novel behavioral model based on a Black Box modeling is presented. The model has been validated using Motorola LDMOS power amplifier. The results have been validated both in time and in frequency domain. This new enables a good prediction of the PA’s behavior. Some measurements of important parameters (like ACLR and EVM) used to describe the nonlinear behavior of the power amplifier driven by WCDMA signal has been also carried out. 6. References

[1] A, Zhu and T. J. Brazil, “Behavioral Modeling of RF Power Amplifiers Based on Pruned Volterra Series,” IEEE Microwave and Wireless Components Letters, Vol. 14, No. 12, December 2004, pp. 563- 565.

[2] N. Le Galou, E. Ngoya, H. Buret, D. Barataud and J.M. Nebus, “An Improved Behavioral Modeling Technique for High Power Amplifiers with Memory,” IEEE MTT-S International Microwave Symposium, Vol. 2, May 2001, pp. 983-986.

[3] T. Wang and T. J. Brazil, “Volterra-Mapping-Based Be-havioral Modeling of Nonlinear Circuits and Systems for High Frequencies,” IEEE Transactions on Microwave Theory and Technology, Vol. MTT-51, May 2003, pp. 1433-1440.

[4] M. Ibnkahla, N. J. Bershad, J. Sombrin and F. Castanié, “Neural Network Modeling and Identification of Nonlinear Channels with Memory: Algorithms, Applications, and Analytic Models,” IEEE Transactions on Signal Processing, Vol. SP- 46, May 1998, pp.1208-1220.

[5] J. C. Pedro and S. A. Maas, “A comparative Overview of Microwave and Wireless Power-amplifier Behavioral Modeling Approaches,” IEEE Transactions on Micro-

wave Theory Technology, Vol. 53, No. 4, April 2005, pp. 1150-1163.

[6] H. Ku, M. Mckinley and J. S. Kenney, “Quantifying Memory Effects in RF Power Amplifiers,” IEEE Trans-actions on Microwave Theory Technology, Vol. 50, No. 12, December 2002, pp. 2843-2849.

[7] J. Kim and K. Konstantinou, “Digital Predistortion of Wideband Signals Based on Power Amplifier Model with Memory,” Electronics Letters, Vol. 37, No. 23, Novem-ber 2001, pp. 1417-1418.

[8] A. Zhu and T. J. Brazil, “Behavioral Modeling of RF Power Amplifiers Based on Pruned Volterra Series,” IEEE Microwave and Wireless Components Letter, Vol. 14, December 2004, pp. 563-565.

[9] C. Silva, A. Moulthrop, and M. Muha, “Introduction to Poly-spectral Modeling and Compensation Techniques for Wideband Communications Systems,” 58th ARFTG Conference Digest, San Diego, November 2001, pp. 1-15.

[10] D. Mirri et al., “A nonlinear Dynamic Model for Per-formance Analysis of Large-signal Amplifiers in Com-munication Systems,” IEEE Transactions on Instrumen-tation Measurement, Vol. 53, No. 2, April 2004. pp. 341-350.

[11] E. Ngoya et al., “Accurate RF and Microwave System Level Modeling of Wideband Nonlinear Circuits,” IEEE MTT-S International Microwave Symposium Digest, Bos-ton, Vol. 1, June 2000, pp. 79-82.

[12] A. Zhu, J. Dooley and T. J. Brazil, “Simplified Volterra Series Based Behavioral Modeling of RF Power Amplifi-ers Using Deviation Reduction,” IEEE MTT-S Interna-tional Microwave Symposium Digest, San Francisco, 2006, pp. 1113-1116.

[13] A. Zhu, J. C. Pedro and T. J. Brazil, “Dynamic Deviation Reduction Based Volterra Behavioral Modeling of RF Power Amplifiers,” IEEE Transaction on Microwave Theory Technology, Vol. 54, No. 12, December 2006, pp. 4323-4332.

[14] A. Zhu and T. J. Brazil, “RF Power Amplifiers Behav-ioral Modeling Using Volterra Expansion with Laguerre Functions,” IEEE MTT-S International Microwave Sym-posium Digest, 2005, pp. 963-966.

[15] M. Isaksson and D. Rönnow, “A Kautz–Volterra Behav-ioral Model for RF Power Amplifiers,” IEEE MTT-S In-ternational Microwave Symposium Digest, 2006, pp. 485-488.

[16] Verspecht, J, “Scattering Functions for Nonlinear Behav-ioral Modeling in the Frequency Domain,” “Fundamen-tals of Nonlinear Behavioral Modeling: Foundations and Applications Workshop,” IEEE MTT-S International Mi-crowave Symposium, June 2003, pp. 1-26.

[17] F. X. Estagerie, T. Reveyrand, S. Mons, R. Quéré, L. Constancias and P. Le Helleye, “From Circuit Topology to Behavioral Model of Power Amplifier Dedicated to Radar Applications,” Electronics Letters , Vol. 43, No. 8, April 2007, pp. 477-479.



Comparison of 4 Multi-User Passive Network Topologies for 3 Different Quantum Key Distribution

Fabio Garzia, Roberto Cusani INFOCOM Department, SAPIENZA – University of Rome, Rome, Italy

E-mail: [email protected] Received March 30, 2010; revised May 7, 2010; accepted June 1, 2010

Abstract The purpose of this paper is to compare the performance of four passive optical network topologies in im-plementing multi-user quantum key distribution, using 3 protocols proposed by quantum cryptography (B92, EPR, and SSP). The considered networks are the passive-star network, the optical-ring network based on the Signac interferometer, the wavelength-routed network, and the wavelength-addressed bus network. The quantum bit-error rate and sifted key rate for each of these topologies are analysed to determine their suit-ability for providing quantum key distribution-service to networks of various sizes. The efficiency of the three considered protocols is also determined. Keywords: Quantum Cryptography, Quantum Key Distribution

1. Introduction The quantum cryptography term represents the set of the techniques which allow two entities, Alice and Bob, to exchange reserved information by means of a quantum channel. A quantum channel is an optical channel governed by the quantum mechanics. The job in cryptographic field of the quantum mechanics allows results impossible to be obtained with the only mathematics. More precisely, talking about quantum key distribution (QKD) is oppor-tune: the quantum channel is used to transmit a sequence of bits, well known only to Alice and Bob and then able to constitute the secret key of a cryptographic system. Therefore the next communications which are ciphered with such key can be made on a conventional channel (not quantum). Quantum key distribution is a method for securely distributing one-time-use encryption keys that are used for secure communications. These quantum systems are based on the theorem of Heisenberg [1], according to which the measurement of a quantum sys-tem generally perturbs it and gives an incomplete piece of information on his state preceding the measurement, and on the quantum no-cloning theorem [2], which for-bids the perfect copying of two non-orthogonal quantum states.

Therefore the quantum nature of a channel makes sure that any interception is noticed. Hence an eavesdropper, Eve, cannot get any information about the communica-tion without introducing perturbations which would reveal

her presence. To share a secret key, Alice and Bob must follow a

protocol (BB84, B92, EPR, SSP). Once developed the procedure requested by the protocol, if any eavesdropper were not noticed, Alice and Bob share a secret key, which exchanged themselves without having to turn to a third reliable part and initially sharing no information, except that the one necessary to authenticate their com-munications part. The frequency used by Alice and Bob to share the sifted secret key is denominated sifted key rate (RSIFT). To reveal the presence of an eavesdropper, Eve, Alice and Bob monitor the quantum bit error rate (QBER). If the QBER exceeds a certain threshold the made communication is just considered as not safe and therefore the secret key is discarded. The security threshold depends on the used protocol. The QBER and the RSIFT are considered the fundamental parameters to evaluate the performances of a quantum channel. This analysis has already been done for BB84 protocol [3]. The purpose of this paper is to extend the mentioned analysis to other three common protocols that are B92, EPR and SSP.

The remainder of this paper is organized as follows. Section II provides a review of 3 protocols used in addi-tion to BB84 protocol [3]. Section III outlines the four network topologies to be compared. The security thresh-old for every used protocol is determined in the Section IV. Section V provides a review of the physical princi-ples used for the simulations for each protocol. The re-

F. GARZIA ET AL. 167 sults of the comparison of the networks are presented in Sections 7 and 8, after having reported in Section 6 the parameters values employed. This is followed by a re-sults discussion, Section 4, and conclusions, Section5.

2. Protocols 2.1. BB84: First QKD Protocol The first protocol has been proposed in 1984 by Charles H. Bennett and Gilles Brassard [4], hence the name BB84 under which this protocol is recognized nowadays.

Alice wants to communicate Bob a bit sequence (qubits). The qubits are encoded with polarized photons. The protocol uses 4 polarization states: 0°, 90°, +45°, –45°.

These states are represented in the following way: horizontal |H> , vertical |V>, |45°> and |-45°>, where H≡0° and V≡90°. This states are assembled in 2 non-orthogonal basis : rectilinear (|H> ; |V>) and diago-nal (|+45°> ; |–45°>). The bases are maximally conjugate in the sense that any pair of vectors, one from each basis, has the same overlap: 1/2 .

Conventionally, one attributes the binary value 0 to states |H> and |45°>| and the value 1 to the other two states. In the first step, Alice sends individual photon to Bob in states chosen at random between the 4 basic states. Next, Bob measures the incoming photons in one of the two bases, chosen at random. If both Alice and Bob choose the same random basis, then Bob’s meas-urements have a deterministic outcome. If they do not choose the same basis, the outcome of his measurement becomes probabilistic. Once made all the measurements, Bob obtains a bit sequence said raw key. In the second step, Alice and Bob communicate over a public channel to compare the bases in which the qubits were encoded and measured. The qubits that are sent and measured in incompatible bases are discarded. The remaining qubits shared between Alice and Bob form the sifted keys. 2.2. B92 In an article of 1992 Charles Bennett proposed a new protocol, B92 [5,6].

The B92 quantum coding scheme is similar to the BB84 coding scheme but used only 2 out of the 4 BB84 states. It encodes classical bits in two non-orthogonal BB84 states. Since no measurement can distinguish two non-orthogonal quantum states, it is impossible to iden-tify the bit with certainty. Moreover, any attempt to learn the bit will modify the state in a noticeable way. This is the basic idea behind the quantum key distribution pro-tocol B92. By contrast to the BB84 case, the B92 coding scheme allows the receiver to learn whenever he gets the bit sent without further discussion with Alice. Since it

uses only 2 quantum states, the B92 coding scheme is sometimes easier to implement. However, the security it provides is more difficult to be established in certain experimental settings and very often turns out to be totally insecure. The polarization encoded version of B92 proceeds as follows for an idealized system.

Both the transmitter “Alice” and the receiver “Bob” generate an independent random bit sequence. Alice then transmits her random bit sequence to Bob using a clocked sequence of linearly polarized individual photons with polarization angles chosen according to her bit values as given by 0° ≡ 0 and 45° ≡ 1. In each time period, Bob makes a polarization measurement on an incoming pho-ton by orientating the transmission axis of his polarizer according to his bit value as given by –45° ≡ 0 and 90° ≡ 1. It can be seen that Bob detect only a photon (with probability one half) in the time slots where his polarizer is not crossed with that of Alice. We refer to these in-stances as “unambiguous” since when they occur, Alice and Bob can be sure that their polarization settings were not orthogonal and, consequently, that their bit values were the same (both 0 or both 1). Conversely, the instances in which Bob receives no photon are referred to as “ambiguous” since they can arise either from the cases where Alice’s and Bob’s polarisers were crossed or from the cases where the polarisers were not crossed, but Bob failed (with probability 1/2) to detect a photon. Bob then uses an authenticated public channel to inform Alice of the time slots in which he obtained an unambiguous result (1/4 on average) and they use the shared subset of their initial random bit sequences represented by these time slots as a key.

In this protocol, whose used values are shown in Ta-ble 1, we see that for the first and fourth bits Alice and Bob had different bit values, so that Bob doesn’t detect any bit in each case. However, for the second and third bits, Alice and Bob have the same bit values and the protocol is such that there is a probability of 50% that Bob detects a bit in each case. Of course, we cannot pre-dict in which of the two cases Bob detects the bit, but in this example he detects only third bit.

The B92 protocol is intrinsically less efficient than the given BB84 that, also in ideal conditions (when no bit of the raw key is to be deleted), only 1/4 of the impulses gives a key bits, while with BB84 protocol fraction is 1/2.

Table 1. An example of B92 protocol.

Alice’s sequence 1 0 1 0

Alice’s polarization +45° 0° +45° 0°

Bob’s polarization -45° -45° 90° 90°

Bob’s sequence 0 0 1 1

Bob’s bit detected No No Yes No


F. GARZIA ET AL. 168

This inefficiency is the price that Alice and Bob must pay for secrecy. 2.3. Six State Protocol (SSP) The Six State Protocol (SSP) is better well-known as the BB84 with the addition of 2 polarization states.

Because of the complex nature of his coefficients, Hilbert space 2-dimensional admits also a third base (cir-cular) conjugate to both the rectilinear and diagonal bases:

| 0 >=(|H>* | 01

2+i|V>*

1

2) (1)

1| >=(|H>*1

2-i|V>*

1

2) (2)

where i = 1 . In the SSP the polarization basis are determined by the

Poincarè sphere. Photons’ polarization is seen along the Cartesian axes where x = rectilinear base; y = diagonal base; z = circular base.

Thus, Alice sends a state randomly polarized in positive or negative x-, y-, or z-direction to Bob, who measures randomly in the x-, y- or z-basis. As in BB84 they com-municate over a public channel and keep only those cases in which their basis was the same.

While two states are enough and four states are stan-dard, a 6-state protocol respects much more the symmetry of the qubit state space. The six states constitute 3 bases; hence the probability that Alice and Bob chose the same basis is only of 1/3. But the symmetry of this protocol greatly simplifies the security analysis and reduces Eve’s optimal information gain for a given error rate QBER. If Eve measures every photon, the QBER is 33%, compared to 25% in the case of the BB84 protocol [1].

2.4. EPR

The protocols described up to now foresee that Alice sends the photons to Bob, where the state of the photon codifies the value of the bit to be transmitted. In the EPR protocol [7], each of the two parts receives a particle

|H>|V>

|0>

|1>

|-45°>

|45°>

z

x

y

Figure 1. Poincarè sphere.

belonging to a couple, produced by a third source. Ekert (1991) has devised a quantum protocol based on the properties of quantum correlated particles. Einstein, Podolsk and Rosen (EPR) [7] point out an interesting phenomenon in quantum mechanics. According to their theory, the EPR effect occurs when a pair of quantum mechanically correlated photons, called the entangled photons, is emitted from a source. The entanglement may arise out of conservation of angular momentum. As a result, each photon is in an undefined polarization. Yet, the two photons always give opposite polarizations when measured along the same basis. Since EPR pairs can be pairs of particles separated at great distances, this leads to what appears to be a paradoxical “action at a distance”. For example, it is possible to create a pair of photons (each of which we label below with the subscripts A and B, respectively) with correlated linear polarizations [8]. An example of such an entangled state is given by:

(A,B)= (|H>A|V>B – |V>A|H>B)*2

1 (3)

Einstein (1935) then states that such quantum correla-tion phenomena could be a strong indication that quan-tum mechanics is incomplete and that there exist “hidden variables”, inaccessible to experiments, which explain such “action at a distance”. Bell [9] gave a means for actually testing for locally hidden variable (LHV) theories. He proved that all such LHV theories must satisfy the Bell inequality. Quantum mechanics has been shown to violate the inequality. The EPR quantum protocol is a 3 state protocol that uses Bell’s inequality to detect the presence or absence of Eve as a hidden variable. We now describe a simplified version of this protocol in terms of the polarization states of an EPR photon pair.

An EPR pair is created at the source. One photon of the constructed EPR pair is sent to Alice, the other to Bob. Alice and Bob at random with equal probability separately and independently measure their respective photons. Alice chooses randomly one of the three measurement directions indicated in Figure 2 whereas Bob chooses a set of directions rotated by 45 [10].

Alice records her measured bit. On the other hand, Bob records the complement of his measured bit. This

z

x

(a) (b)

Figures 2. (a) Alice’s directions of measurement; b) Bob’s directions of measurement.


F. GARZIA ET AL. 169 procedure is repeated for as many EPR pairs as needed. Alice and Bob carry on a discussion over a public channel to determine the correct bases they used for measurement. Each of them then separates its respective bit sequences into two sub-sequences. One subsequence, called raw key, consists of those bits at which they used the same basis for measurement. The other subsequence, called rejected key, consists of all the remaining bits.

Unlike the BB84 and B92 protocols, the EPR protocol, instead of discarding rejected key, actually uses it to detect Eve’s presence. Alice and Bob now carry on a discussion over a public channel comparing their respec-tive rejected keys to determine whether or not Bell’s inequality is satisfied. If it is, Eve’s presence is detected. If not, then Eve is absent. In this way the probability that they happen to choose the same basis is reduced from 1/2 to 2/9 [1], but at the same time as they establish a key they collect enough data to test Bell inequality.

3. Topologies of Multi-User QKD Networks The first experimental implementation of QKD occurred in 1989 [11], when encryption keys were transmitted through 30 cm of air using polarization-encoded photons. It was shown that the use of orthogonal states on more than 10 km of optical fibre is impossible, according to the characteristics of the sources available at present [2, 12]. To allow transmissions at distances always longer, it is therefore necessary the use of systems different from the ones used before. In particular when using an interferometer we can encode qubits in an interferometric phase state.

For example we explain the implementation of BB84 using an interferometer. Alice encodes the photons with her phase modulator (PM) by randomly choosing one of four phase shifts: 0 and correspond to one basis set and /2 and 3/2 correspond to another basis set. She associ-ates 0 and /2 with qubit 0, and and 3/2 with qubit 1. Bob makes his measurement by randomly choosing be-tween a 0 or /2 phase shift. Only photons with a final phase shift of 0 or (the difference of Alice’s and Bob’s phase shifts) can interfere in Bob’s interferometer to produce a deterministic outcome. Any final phase shift /2 or 3/2 leads to a probabilistic outcome. Thus, whenever Bob measures correctly, qubit 0 is routed to Detector 1 (Det1) and qubit 1 to Detector 2 (Det2). Since Bob’s measurement consists of a random choice of basis, half of the measurement results is probabilistic. Therefore, after the qubit transmission, Bob confers with Alice about the appropriate basis choice. Any qubit measured in an incompatible basis is discarded and does not be-come part of the final key. This process creates the sifted key.

Now we introduce the four QKD network topologies to be compared [3]. These networks phase-encode the

qubits in optical fibre interferometers. The optical-ring network uses a Signac interferometer; all other topologies are implemented with unbalanced Mach–Zehnder inter-ferometers (MZIs). The unbalanced MZI is a modifica-tion of the standard MZI with improved interference stability. This improved stability comes at the expense of a 3-dB loss, since half of the photons transmitted through it are lost in the non-interfering path combinations of the interferometer [1]. This makes networks that use the unbalanced MZIs more loss, thus lowering their sifted key rate and increasing their QBER. The single-photon sources used in the network topologies and in the calcu-lations are modelled as highly attenuated laser pulses that are typically used in practice and contain an average of 0.1 photon per pulse. The single-photon detectors are also modelled as the response of gated avalanche photo-diodes operated in Geiger mode [13].

In general, Alice is defined as the user that provides the qubit information in the four bases, and Bob is de-fined as the user that chooses between the two non- thogonal basis sets. For the passive-star (Figure 3), wavelength-routed (Figure 5), and wavelength-addressed bus (Figure 6) topologies, Alice is the network controller.

TA

TA

TA

PM

PLS

Alice

1 x NSplitter

PM

Det 1

Det 2

Dan

Bob

Chris

N-thuser

Figure 3. Network topology of passive -star multi-user QKD network. (PLS: Pulsed laser source; TA: tuneable attenu-ator; PM: phase modulator; Det: detector.).

TA

PLS

CirculatorCoupler

Det 1

Det 2

Bob

Alice 1

Alice 2

Alice N

PM

Alice 3

PM

CW CCW

Figure 4. Network topology of optical-ring multi-user QKD network based on Signac interferometer.



TA

TA

TA

PM

TunablePLS

Alice

AWG

PM

Det 1

Det 2

Dan ( 3 )

Bob ( 1)

Chris ( 1)

N-thuser ( N )

Figure 5. Network topology of wavelength-routed multiuser QKD network. (AWG: arrayed-waveguide grating).

TATA

TAPM

TunablePLS

Alice

G G G

PM

Det 1 Det 2

Bob

PM

Det 1 Det 2

Chris

PM

Det 1 Det 2

N-th user

1 2 N

Figure 6. Network topology of wavelength-addressed bus multi-user QKD network. (G: fibre Bragg grating). She is equipped with an unbalanced MZI, a pulsed laser source (PLS), a tuneable attenuator (TA), and a four-state PM. The users at the receiving end (Bob, Chris, Dan, N-th user) choose between the two non-orthogonal bases. Each one of them has another unbalanced MZI, a two-state PM, and a pair of single-photon detectors (Det1 and Det2). The optical-ring network (Figure 4) is sig-nificantly different from the others. Here, Bob is the network controller and services multiple Alice. Bob’s setup consists of a laser source, two detectors, a two-state PM, and a circulator. Each Alice only possesses a four-state PM. 3.1. Passive-Star Network The topology of the passive-star QKD network is shown in Figure 3 [3]. A passive-star QKD network was first demonstrated to connect four users over 5.4 km of opti-cal fibre [14]. This topology is an extension of the two-user system, with Alice linked to receivers through a 1xN splitter. Due to the indivisible nature of the photon, each photon is randomly routed to a single user by the

1xN splitter. This topology can be easily implemented but suffers from the effective loss induced by the 1 splitter, which reduces the probability of photons to reach the detectors of any particular user. This reduction scales inversely as the number of users on the network. For example, a three-user network having a 1x2 splitter reduces the probability that a photon reach the desired receiver by one half and consequently acts as a 3-dB attenuator. A 17-user network containing a 1x16 splitter acts effectively like a 12-dB attenuator, and so on. Al-though this drawback can be partially mitigated by higher initial qubit rates, the routing of the photons to each user is inherently nondeterministic. For example, the mean detection rate at each user after a 1xN splitter is 1/Nth of the detection rate of a single Bob without the 1xN splitter. However, since the routing of photons to each user through the 1xN splitter is random, at any given time, some users receive photons at a rate above the mean detection rate of 1/Nth, and some users receive photons at a rate below the mean detection rate. This nondeterministic detection rate constrains the design of secure quantum networks by limiting the amount of in-formation that can be securely encrypted.

3.2. Optical-Ring Network Based on Signac Interferometer

Figure 4 shows the schematic diagram of the optical-ring network topology. A two-user QKD system based on the optical fibre Signac interferometer has been demon-strated [15]. This topology is significantly different from the topologies based on the unbalanced MZIs: the single- photon pulse enters the Signac interferometer through an optical circulator. This pulse splits into two parts in the 50/50 coupler, and each travels around the Signac loop in clockwise (CW) and counter clockwise (CCW) direc-tions, respectively. Any user on the loop that is commu-nicating with Bob modulates the pulse travelling in the CW direction. Bob modulates the pulse travelling in the CCW direction. The position of Bob’s PM is important since the pulse that it modulates must be returning from its round trip in the loop in order to prevent any informa-tion about Bob’s modulation choice from travelling through the loop. A timing and control mechanism must also be established so that only one Alice can modulate the photon at a time. Upon travelling around the loop, the pulses interfere in the coupler and enter one of two pho-ton detectors. Photons enter Detector 1 (Det1 in Figure 4) when they experience a phase shift between the CW and CCW pulses inside the Signac interferometer. On the other hand, they enter Detector 2 (Det2 in Figure 4) when they experience a 2 phase shift between the CW and CCW pulses inside the Signac interferometer. The Signac interferometer has the advantage of being free from thermal fluctuations since the counter propagating pulses pass through the exact same fibre paths inside the


F. GARZIA ET AL. 171 loop. Another potential advantage is that each user on the network, except Bob, contains only a single-PM and no photon detectors. This can simplify any deployment of a secure ring network using the Signac because Bob is the only user that requires the single-photon detectors. 3.3. Wavelength-Routed Network

The schematic diagram of the wavelength-routed net-work topology is depicted in Figure 5 [3]. This topology is implemented with unbalanced MZIs and is very similar in layout to the star network. The greater difference is that Alice has the ability to control which user receives the photons by employing a wavelength-routing scheme. Alice is equipped with a wavelength tuneable pulsed laser source (PLS) and the receivers are assigned their own wavelength channel. Alice transmits to a particular user by tuning her source to that user’s wavelength and the photons are routed via an arrayed waveguide grating (AWG). The advantage of this topology is that the inser-tion loss of the AWG is approximately uniform regard-less of the number of channels. Theoretically, the num-ber of users that this type of network supports is limited only by the channel spacing of the AWG and the band-width of the fibre. In addition, the single-photon detec-tors must be sensitive for the entire range of frequencies used in the network. This is not a concern as avalanche-photodiode (APD)-based single-photon de-tectors respond to a much broader spectrum than the band of wavelengths used in multi-wavelength networks.

Due to the principles of quantum mechanics described above, it is impossible for the spy Eve to gain perfect knowledge of the quantum state sent from Alice to Bob. Nevertheless, she can acquire some knowledge. Without interaction of a spy, each two-level quantum system carries 1 bit of information from Alice to Bob. When Eve gets hold of part of this information, she cannot prevent causing a disturbance to the state arriving at Bob’s side, and thus introducing a non-zero error rate. In principle, Bob can find out about this error rate and thus about the existence of a spy by communicating with Alice. The source for Eve’s knowledge is measurements performed on the signals (quantum states). The simplest eavesdrop-ping attack (intercept/resend) for Eve would be to meas-ure each signal just as Bob would do, and then to resend a signal to Bob which corresponds to the measurement result. Further we always have some detector noise, misalignments of detectors and so on. It should be pointed out that we cannot even in principle distinguish errors due to noise from errors due to eavesdropping activity. We therefore assume that all errors are due to eavesdropping. Another issue, not discussed here, is that of statistics. Eavesdroppers can be lucky: they create errors only on average, so in any specific realization the error rate might be zero (with probability exponentially small in the key length, of course). We are guided by the idea that a small error rate, for example 1 %, implies that an eavesdropper was not very active, while a big error rate is the signature of a serious eavesdropping attempt. But what is the meaning of “small” and “big”? From an information theoretic point of view, the natural measure of “knowledge” about some signal is represented by the Shannon information. It is measured in bits and can be defined for any two parties, the sender of the signal and the observer (receiver). In general terms, the knowledge of the observer consists of obtained measurement results and any additional gathered knowledge, like the an-nounced basis of signals in the BB84 protocol.

3.4. Wavelength-Addressed Bus Network

The wavelength-addressed bus network is also based on the unbalanced MZI setup and it is shown in Figure 6 [3].

Like the wavelength-routed network, this network also allows Alice to route her photons to a desired user by tuning the photons to be desired wavelength. In such a system, Alice is equipped with a tuneable PLS, and each receiver is assigned its own wavelength channel. Alice selects an intended receiver by tuning her source to that user’s wavelength and transmits the encoded photons along the bus. The receivers are connected to the bus line through a fibre Bragg grating (G), which allows them to retrieve only the photons, intended for them. These gratings are designed to reflect photons of a specific wavelength to a given user and transmit all others. The network accommodates multiple users by placing several fibre Bragg gratings in series along the bus. One of the merits of this topology is that it can be easily expanded to accommodate more users by simply tapping the bus and inserting a suitable grating.

4. Security Threshold

The QBER, which is indicative of the security and post-error-correction net key rate, is useful for assessing the performance of the network. High QBER values in QKD systems lower the net key rate during the error correction stage of the protocol [1]. In addition, high QBER allows an eavesdropper to gain more information about the transmitted keys at the expense of the legiti-mate receiver. It has been shown that for QBERs above a security threshold, an eavesdropper can actually gain more information than the legitimate receiver. If this happens, it is not possible to use any privacy-amplification technique. Therefore, when designing a QKD network, it is necessary to ensure that the baseline QBER is below this security threshold so that privacy amplification strategies may be used to eliminate any knowledge


F. GARZIA ET AL. 172 gained by Eve [1]. For QBERs under this threshold (QBERT), the Shannon information between Alice and Bob (IAB) is higher than that in Eva’s possession (IE ), while for superior values that of Eva is greater:

QBER < QBERT IAB > IE (4) QBER > QBERT IAB < IE (5)

Bounds on the obtainable Shannon information for eavesdropping on single bits can be found in the litera-ture for different protocols. Fuchs et al. give bounds for the BB84 [16] and the B92 protocol [17]. A bound for the Six State Protocol was also obtained [18]. These bounds are illustrated in Figures 7–9 for each of used protocol. Note the trade-off between Eve’s information gain and the disturbance she causes: more information for Eve means higher error rate for Bob. For reasonably low error rates Eve’s maximal information is smallest in the six-state protocol, as it uses the largest ensemble of input states.

Figure 7. Shannon Information (in normalized units) with B92 protocol.

Figure 8. Shannon Information (in normalized units) with SSP.

Figure 9. Shannon Information (in normalized units) with EPR protocol. Furthermore comparing Eva’s Shannon Information with the Shannon information between Alice and Bob, we are able to determinate the threshold for the QBER for each of the used protocols. 4.1. B92 Security Threshold for B92 protocol is:

QBERT 14% (6)

4.2. SSP Security Threshold for the Six State Protocol is:

QBERT 17% (7) 4.3. EPR Security Threshold for EPR protocol is:

QBERT 15% (8)

5. Key Parameters in QKD

QBER and RSIFT are two parameters used to gauge the performance of network topologies which offer QKD service. The QBER and sifted key rate equations that are used in the simulations are reviewed in this section. More detailed discussions on the physical principles underlying these equations are provided in references [1] and [19]. The sifted keys are those keys shared by Alice and Bob when they make compatible basis choices [19]:

RSIFT = q RRAW (9)

RRAW = fREP tLINK (raw key rate) (10) where q depends on protocol (for example, in BB84 pro-tocol q = 1/2 because half of the time Alice and Bob bases are not compatible), fREP is the repetition frequency, µ is the average number of photons per pulse, tLINK is the


F. GARZIA ET AL. 173 transmission coefficient of the link and is Bob’s detec-tion efficiency. The transmission coefficient is related to the loss lF (in dB per km) and length L (in km) of the fibre, the loss due to the number of users lN(N) (in dB), and the topology selected, by

tLINK = 10/10 TNF lNlLl (11) The topology choice introduces a topology loss con-

stant lT (in dB) that is an overhead of loss involved in working with a particular topology. This quantity is con-stant regardless of a network’s fibre length and number of users. The topology loss has 4 components: end-user losses arising from losses in the receiver’s interferometer, routing loss caused by the device that selects the user that receives the photon, the non-interfering path combi-nation loss in the unbalanced MZIs (for those topologies that use them), and miscellaneous losses, such as those caused by connectors and splices.

The QBER is defined as the number of wrong bits of the total number of received bits and is normally in the order of a few percent. In the following we use it ex-pressed as a function of rates [1]:

SIFT

error

errorSIFT

error

R

R

RR

RQBER

(12)

One can distinguish three different contributions to RERROR. The first one arises because of photons ending up in the wrong detector, due to imperfect interference or polarization contrast. The rate ROPT is given by the product of the sifted key rate and the probability POPT of a photon going in the wrong detector:

ROPT = RSIFT POPT (13)

This contribution can be considered, for a given set-up, as an intrinsic error rate indicating the suitability to use it for QKD. Imperfect phase matching in the interferometers results in reduced fringe visibilities that lead to an increased probability of routing photons to the wrong detectors. The probability of this type of error POPT is related to the fringe visibility (V) by:

POPT = 2

1 V (14)

The second contribution, RDARK, arises from the detec-tor dark counts (or from remaining environmental stray light in free space setups). This rate is independent of the bit rate and depends only on the characteristic of the photon counter [13]. Of course, only dark counts falling in a short time window when a photon is expected give rise to errors:

RDARK = k fREP PDARK (15)

where PDARK is the probability of registering a dark count per time-window and per detector, and the k factor is related to the fact that a dark count has a k % chance to happen with Alice and Bob having chosen incompatible bases (thus eliminated during sifting). Finally error counts can arise from uncorrelated photons, because of imperfect photon sources:

RACC = 2

1fREP tLINK PACC (16)

This factor appears only in systems based on entangled photons, where the photons belonging to different pairs, but arriving in the same time window, are not necessarily in the same state. The quantity PACC is the probability to find a second pair within the time window, knowing that a first one was created. The QBER can now be expressed as follows:

SIFT

ACCDARKOPT

R

RRRQBER

(17)

6. Parameter Values

The results are based on calculations assuming the following parameter values, which are held constant for each topology [1,13,14,20,21] :

Pulse repetition rate (fREP) 1 MHz Mean number of photon per pulse ( ) 0.1 Detector efficiency @1310 nm ( ) 20% Detector efficiency @1550 nm ( ) 10% Dark count probability (PDARK ) 10-5

Fringe visibility (V) 98% The transmission coefficient link tLINK varies from one

topology to another. The values used in the simulations that contribute to tLINK are outlined for each topology in Table 2. In the table the contributions to the topology losses are also shown; namely, the end-user loss, routing loss, non-interfering path combination loss, and miscel-laneous loss.

The end-user loss arises from the excess loss in the couplers and PM in the receiver’s interferometer. Routing loss is the loss in the device that routes the photons to each user. In the star, wavelength-routed, and bus networks, which are all based on the unbalanced MZI design, a 3-dB loss arises from non-interfering path combinations. The miscellaneous loss stems from losses Table 2. Losses contributing to the transmission coefficient tLINK for the 4 considered network topologies.

Loss Source Star Ring W.Routed Bus Topology Loss

End User Loss (dB)Routing Loss (dB)

Non interfer. path Loss (dB)Miscellaneous Loss (dB)

0.3 0.1 3.0 1.0

0.49 0.0 0.0 1.0

0.33.03.01.0

0.30.02

3.01.0

Total Topology Loss (dB) 4.4 1.49 7.3 4.32

Fiber Loss (dB/km) 0.35 @ 1310 nm 0.25 @ 1550 nm

User number Loss ( dB) 10log(N) 0.1N 0 0.2(N-1)


F. GARZIA ET AL. 174 such as those due to connectors, splices, and imperfections in the network all of which occur in practical optical network setups.

Now we are able to analyse the QBER and the RSIFT for every QKD protocol considered previously. The results of QBER for each topology are presented in sur-face plots which relate the QBER to the number of users and distance. The term “distance” is defined as the total fibre length used in the transmission of the photons. For the optical ring, it is the total length of the Signac loop. For all the other topologies, it is the total fibre length spanning Alice and Bob (or Chris, Dan, etc.). The system performances at the 1310-nm and 1550-nm telecommu-nications wavelength windows are shown in the results. The shaded regions in the QBER surface plots corre-spond to the combinations of distance and number of network users for which the QBER is less than QBERT. This threshold, previously mentioned in section IV, is the value below which secure key distribution can be per-formed on the network. Thus, the shape and area of the shaded regions allow one to easily determine the suit-ability of a given topology to support a given number of users. In addition, these plots also serve to show a net-work’s sensitivity to expanding the number of users.

7. QBER Performance

7.1. B92

As previously explained in section 2.2, B92 protocol is intrinsically less efficient than the given BB84 where, also in ideal conditions (when no bit of the raw key is to be deleted), only 1/4 of the impulses gives a key bits, while with BB84 this protocol fraction is 1/2. This in-efficiency is the price that Alice and Bob must pay for secrecy.

RSIFT = 4

1RRAW =

4

1 fREP tLINK (18)

ROPT =4

1 fREP tLINK POPT (19)

where

POPT = 2

1 V= 0.01 = 1% (20)

since fringe visibility is 98%.

RDARK = 2

1 fREP PDARK (21)

The 3-D QBER surface of the four topologies using a carrier wavelength of 1310 nm is illustrated in Figure 10, and the 3-D QBER surface at 1550 nm is illustrated in Figure 11. Table 3 and Table 4 summarize the informa-tion obtained from Figure 10 and Figure 11.

7.2. SSP

As previously explained in Subsection 2.3, the six states

(a)

(b)

(c)

(d)

Figure 10. B92-protocol topologies at 1310 nm. QBER sur-face as a function of users and distance. Users range: 2-128 users. Distance range: 0- 80 km. QBER<14% in shaded re-gion. (a) Passive Star B92 1310 nm; (b) Optical Ring B92 1310 nm; (c) Wavelength-routed B92 1310 nm; (d) Wave-length-addressed bus B92 1310 nm.



(a)

(b)

(c)

(d)

Figures 11: B92-protocol, topologies at 1550 nm. QBER surface as a function of users and distance. Users range: 2-128 users. Distance range: 0-80 km. QBER<14% in shaded region. (a) Passive Star B92 1550 nm; (b) Optical Ring B92 1550 nm; (c) Wavelength-routed B92 1550 nm; (d) Wavelength-addressed bus B92 1550 nm.

Table 3. B92-protocol Maximum number of users sup-ported by every topology at different distance for QBER <14%.

Distance(km)

Star 1310 nm / 1550 nm

Ring 1310 nm / 1550 nm

W.routed 1310 nm / 1550 nm

Bus 1310 nm / 1550nm

10 32/20 >128 / >128 >128/ >128 76/66

20 14/11 >128 / >128 >128/ >128 59/54

30 6/6 110/ 110 >128/ >128 41/41 40 2/3 75/85 >128/ >128 24/29 50 1/2 40/60 0/>128 6/ 16 60 0/1 5/35 0/0 0/4 70 0/0 0/10 0/0 0/0 80 0/0 0/0 0/0 0/0

Table 4. B92-protocol Maximum distance (km) supported by every topology for various number of users for QBER<14%.

Number of users

Star 1310 nm / 1550 nm

Ring 1310 nm / 1550 nm


Bus 1310 nm / 1550 nm

20 15/10 55/66 44/50 42/47

40 7/0 50/58 44/50 31/31

60 2/0 44/50 44/50 19/15

80 0/0 38/42 44/50 8/0

100 0/0 32/34 44/50 0/0

120 0/0 27/26 44/50 0/0

constitute 3 bases, hence the probability that Alice and Bob chose the same basis is only of 1/3. This means that to determinate the sifted key, that Alice and Bob can share, an average of 2/3 of the received bits must be dis-carded. But the symmetry of this protocol greatly simpli-fies the security analysis and reduces Eve’s optimal information gain for a given error rate QBER.

RSIFT = 3

1RRAW =

3

1 fREP tLINK (22)

ROPT =3


where POPT = 2

1 V= 0.01 = 1%,

since fringe visibility is 98%.

RDARK = 3

2 fREP PDARK (24)

The 3-D QBER surface of the four topologies using a carrier wavelength of 1310 nm is illustrated in Figure 12, and the 3-D QBER surface at 1550 nm is illustrated in Figure 13. Table 5 and Table 6 summarize the informa-tion obtained from Figure 12 and Figure 13.

7.3. EPR

As previously explained in Subsection 2.4, in the EPR protocol, each of the two parts (Alice and Bob) receives a particle belonging to a couple, produced by a thirsource. Because this source is not perfect, it could generate



(a)

(b)

(c)

(d)

Figures 12. SSP, topologies at 1310 nm. QBER surface as a function of users and distance. Users range: 2-128 users. Distance range: 0- 80 km. QBER<17% in shaded region. (a) Passive Star SSP 1310 nm; (b) Optical Ring SSP 1310 nm; (c) Wavelength-routed SSP 1310 nm; (d) Wavelength-addressed bus SSP 1310 nm.

Table 5. SSP. Maximum number of users supported by every topology at different distance for QBER <17%.

Distance (km)

Star 1310 nm / 1550 nm

Ring 1310 nm / 1550 nm


Bus 1310 nm

/ 1550 nm

10 34/21 >128 / >128

>128 / >128 78/68

20 15/12 >128 / >128

>128 / >128 60/55

30 6/6 113 /113 >128 / >128 43/43

40 3/3 78 /88 >128 / >128 25/30

50 1/2 43 /63 0/ >128 8/18

60 0/1 8 /38 0/0 0/5

70 0/0 0 /13 0/0 0/0

80 0/0 0/0 0/0 0/0

Table 6. SSP. Maximum distance (km) supported by every topology for various number of users for QBER<17%.

Number of users

Star 1310 nm

/ 1550 nm

Ring 1310 nm / 1550 nm


Bus 1310 nm / 1550 nm

20 16/11 56/67 45/52 43/48

40 8/0 50/59 45/52 31/32

60 3/0 45/51 45/52 20/16

80 0/0 39/43 45/52 9/0

100 0/0 33/35 45/52 1/0

120 0/0 28/27 45/52 0/0

uncorrelated photons that generate error counts (RACC). The photons belonging to different pairs, not necessarily in the same state, could arrive in the same time window with probability PACC. Furthermore the EPR protocol, instead of discarding rejected key, actually uses it to detect Eve’s presence. By a discussion over a public channel, Alice and Bob compare their respective rejected keys to determine whether or not Bell’s inequality is satisfied. If it is, Eve’s presence is detected. If not, then Eve is absent. In this way the probability that they happen to choose the same basis is reduced from 1/2 to 2/9 [1], but at the same time as they establish a key they collect enough data to test Bell inequality.

RSIFT = 9

2RRAW =

9

2 fREP tLINK (25)

ROPT =9


where POPT = 1%.

RDARK = 9

7 fREP PDARK (27)

RACC = 2

1fREP tLINK PACC (28)



PACC = 2

1 2 = 000.5 (29)

The 3-D QBER surface of the four topologies using a carrier wavelength of 1310 nm is illustrated in Figure 14, and the 3-D QBER surface at 1550 nm is illustrated in Figure 15. Table 7 and Table 8 summarize the information obtained from Figure 14 and Figure 15. 8. RSIFT Performance To be able to make more visible the difference between the various topologies, we compare the sifted key rate of each topology as a function of distance for 4, 32, 64, 128 users. They were obviously assembled for every used protocol.

The Sifted Key Rate performances of the four network topologies are the same for each used protocol. We use a grading system ranging from 1–4, where 1 indicates the network topology with the best performance, and 4 indi-cates the network topology with the worst performance, to summarize the results of the comparison of the sifted key rate performance of the network topologies. This is shown in Table 9.

Another observation that is made is the distance (30 km) that the 1550- and 1310-nm sifted key rate lines for a particular network cross each other. This distance, which we conveniently call the crossover distance, is the same for all four topologies and determines when the sifted key rate values at 1550 nm are greater or less than the key rates at 1310 nm. For distances less than the crossover distance, the sifted key rate values at 1310 nm are always greater than at 1550 nm.

The situation reverses for distances beyond the cross-over distance so that the sifted key rate values at 1550 nm become greater.

9. Results Discussion

Passive star network, that at first glance appears to be the easiest to implement, turns out to be the worst net topology because:

1) supports the smallest number of users for any given distance;

2) is very sensitive to change in the distance and/or in number of the users;

3) has the lowest RSIFT . Furthermore it requires each user to have their own

interferometer and photo-detectors. From this point of view, the ring topology is the simpler design, requiring each user to have only one four-state.

Optical ring network is characterized from: 1) higher stability against polarization and phase

fluctuations than the other three topologies since each pulse travels through the same fibre length in both the CW and CCW directions [22];

(a)

(b)

(c)

(d)

Figure 13. EPR- protocol, topologies at 1310 nm. QBER surface as a function of users and distance. Users range: 2-128 users. Distance range: 0-80 km. QBER < 15% in shaded region. (a) Passive Star EPR 1310 nm; (b) Optical Ring EPR 1310 nm; (c) Wavelength-routed EPR 1310 nm; (d) Wavelength-addressed bus EPR 1310 nm.



(a)

(b)

(c)

(d)

Figure 14. EPR- protocol, topologies at 1550 nm. QBER surface as a function of users and distance. Users range: 2-128 users. Distance range: 0-80 km. QBER < 15% in shaded region. (a) Passive Star EPR 1550 nm; (b) Optical Ring EPR 1550 nm; (c) Wavelength-routed EPR 1550 nm; (d) Wavelength-addressed bus EPR 1550 nm.

Table 7. EPR-protocol Maximum number of users sup-ported by every topology at different distance for QBER < 15%.

Dis-tance (km)

Star 1310 nm /1550 nm

Ring 1310 nm /1550 nm


Bus 1310 nm /1550 nm

10 3/1 79 / 58 >128/ >128 26/16

20 1/1 44 / 33 0/ 0 8/3

30 0/0 9/9 0/ 0 0/0

40 0/0 0/0 0/0 0/0

50 0/0 0/0 0/0 0/0

60 0/0 0/0 0/0 0/0

70 0/0 0/0 0/0 0/0

80 0/0 0/0 0/0 0/0

Table 8. EPR-protocol Maximum distance (km) supported by every topology for various number of users for QBER < 15%.

Number of users

Star 1310 nm

/ 1550 nm

Ring 1310 nm / 1550 nm


Bus 1310 nm

/ 1550 nm

20 0/0 26/25 15/10 13/7

40 0/0 21/17 15/10 2/0

60 0/0 15/9 15/10 0/0

80 0/0 9/1 15/10 0/0

100 0/0 4/0 15/10 0/0

120 0/0 0/0 15/10 0/0

Table 9. Comparison of the Sifted Key Rate performance for the 4 network topologies.

Number of users

Passive star

Optical ring

W.routed Bus

4 4 1 3 2

32 4 1 2 3 64 4 1 1 3 128 3 2 1 4

2) lowest structure loss (1.49 dB, Table2); 3) lowest QBER with less than 64 users; 4) highest Sifted Key Rate with less than 64 users 5) being more susceptible to Trojan horse attacks

than systems based on the unbalanced MZI [3]. Wavelength network: 1) is the most suitable for networks with more than

64 users, because its Sifted key Rate is independent of the number of users on the network;

2) it may not be the best choice for networks that are not expected to expand beyond 64 users because since it has the highest structure loss (7.3 dB).



Figure 15. B92-protocol Sifted Key Rate versus distance for 4, 32, 64, 128 users.

Figure 16. SSP. Sifted Key Rate versus distance for 4, 32, 64, 128 users.



Figure 17. EPR-protocol. Sifted Key Rate versus distance for 4, 32, 64, 128 users.

Wavelength-addressed-bus network: 1) is the most favourable for networks with less

than 20 users because it can be easily expanded and has moderate structure loss (4.32 dB);

2) is unadvisable for networks with large number of users because it has a higher per-user loss than the ring network.

It has also been shown that there is a crossover dis-tance (30 km) that determines the optimum wavelength (1310 or 1550 nm) to use in the QKD network.

About QKD analyzed protocol only B92 and SSP turned out the most efficient. The EPR protocol is the less efficient. The difficulty to handle couples of particles without changing their correlation does not allow to obtain high performances.

The results obtained at the moment are the least encouraging for all the four net topologies. The maximum reachable distance was of 30 km with 9 user maximum using the Optical-Ring topology. Only for distances lower than 10 km it is possible to obtain sufficient performances avoiding however the Passive-Star topology. Six States Protocol and the B92 present praiseworthy results. The B92 is the protocol of QKD more used and allows to make less communications on public channel. Six State Protocol prevails on everybody because, having a security threshold of 17%, allows to have a high number of users also beyond the 60 km, furthermore it has the fastest Sifted Key Rate.

Commonly used technologies and techniques have been applied in order to evaluate the performances (QBER and Sifted Key Rate). To avoid that Eva can take some photons and measure their polarization without disturbing the one of the photons which arrives to Bob, it was considered an attenuation of the transmitted radia-tion, obtaining an average 0.1 photons per impulse. Commonly used technologies and techniques have been applied in order to evaluate the performances (QBER and Sifted Key Rate). Furthermore we considered pho-ton-detectors with efficiency of 10% at 1550 nm and 20% at 1310 nm. Obviously, this choice implies a non- neglectable reduction of the performances, but this com-promise solution has been chosen to operate a more realistic analysis. Technological improvements in single-photon detectors are able to reduce the number of photons lost and therefore to increase the system performances.

10. Conclusions

In this paper the performances of four passive optical network topologies in implementing multi-user QKD, using 3 protocols proposed by quantum cryptography (B92, EPR, and SSP) have been compared. The QBER and sifted key rate for each of these topologies have been analysed to determine their suitability for providing service to networks of various sizes.


F. GARZIA ET AL. 181 11. References

[1] N. Gisin, G. Ribordy, W. Tittel and H. Zbinden, “Quan- tum Cryptography,” Reveiw of Modern Physics, Vol. 2, Septemner 2001, pp. 1-57.

[2] W. K. Wootters and W. Zurek, “A Single Quantum Can- not be Cloned,” Nature, Vol. 299, London, 1982. pp. 802-803.

[3] P. Kumavor, A. Beal, S. Yelin, E. Donkor and B. Wang, “Comparison of Four Multi-user Quantum Key Distribu- tion Schemes Over Passive Optical Networks,” Journal of Lightwave Technology, Vol 23, No.1 January 2005, p.268.

[4] C. H. Bennett and G. Brassard, “Quantum Cryptography: Public Key Distribution and Coin Tossing,” IEEE Conference on Computers, Systems Signal Processing, Bangalore, 1984.

[5] C. Bennett, “Quantum Cryptography Using Any Two Non-orthogonal States,” Physical Review Letter, Vol. 68, 1992, p. 3121,

[6] K. J. Gordon, V. Fernandez, P. D. Townsend and G. S. Buller, “A Short Wavelength Giga Hertz Clocked Fiber-Optic Quantum Key Distribution System,” IEEE Journal of Quantum Electronics, Vol 40, No7, July 2004, pp. 900-908.

[7] A. Einstein, B. Podolsky and N. Rosen, “Can Quantum- Mechanical Description of Physical Reality be Considered Complete,” Physical Review, Vol. 41, May 1935, p. 777.

[8] M. M. Ishtiaq Khan and M. Sher, “Protocols for Secure Quantum Transmission: A Review of Recent Develop- ments,” Pakistan Journal of Information and Technology，Vol. 2, No. 3, 2003, pp. 265-276.

[9] J. Bell, “On the Einstein, Podolsky, Rosen Paradox,” Physics, Vol. 1, 1964, pp. 195-200.

[10] D. Bruß and N. Lutkenhaus, “Quantum Key Distribution: From Principles to Practicalities,” Vol. 2, September 1999.

[11] C. Bennett, F. Bessette, G. Brassard, L. Salvail and J. Smolin, “Experimental Quantum Cryptography,” Journal of Cryptography, Vol. 5, No. 1, 1992, pp. 3-28.

[12] A. Muller, J. Breguet and N. Gisin, “Experimental Demonstration of Quantum Cryptography Using Pola- rized Photons in Optical Fiber Over More Than 1 km,” Europhysics Letter, Vol. 23, 1993, pp. 383-388.

[13] D. Stucki et al., “Photon Counting for Quantum Key Distribution with Peltier Cooled InGaAs/InP APD’s,” Journal of Modern Optics, Vol. 48, No. 13, 2001, pp. 1967-1981.

[14] P. D. Townsend, “Quantum Cryptography on Multi-User Optical Fiber Networks,” Journal of Nature, Vol. 385, No.2, 1997, pp. 47-49.

[15] T. Nishioka, H. Ishizuka, T. Hasegawa and J. Abe, “Circular Type Quantum Key Distribution,” IEEE Photon Technical Letter, Vol. 14, No. 4, April 2002, pp. 576-578.

[16] C. A. Fuchs, N. Gisin, R. B. Griffiths, C. S. Niu and A. Peres, “Optimal Eavesdropping in Quantum Crypto-

graphy I,” Physical Review A, Vol. 56, 1997, pp. 1163.

[17] C. A. Fuchs and A. Peres, “Quantum State Disturbance VS. Information Gain: Uncertainty Relations for Quantum Information,” Physical Review A, Vol. 53, 1996, pp. 2038-2045.

[18] D. Bruß, “Optimal Eavesdropping in Quantum Crypto- graphy With Six States,” Physical Reveiw Letter, Vol. 81, 1998, p. 3018.

[19] D. Stucki, N. Gisin, O. Guinnard, G. Ribordi and H. Zbinden, “Quantum Key Distribution over 67 km with a Plug&Play System,” New Journal of Physics, Vol. 4, July 2002, pp. 1-8.

[20] P. A. Hiskett et al., “Eighty Kilometer Transmission Experiment Using an InGaAs/InP SPAD-based Quantum Cryptography Receiver Operating at 1.55 m,” Journal of Modern Optics, Vol. 48, No. 13, July 2001, pp. 1957-1966.

[21] D. S. Bethune and W. P. Risk, “Autocompensating Quan- tum Cryptography,” New Journal of Physics, Vol. 4, July 2002, pp. 1-15.

[22] X. Fang and R. O. Claus, “Polarization-Dependent All- fiber Wavelength Division Multiplexer Based on a Signac Interferometer,” Optical Letter, Vol. 20, No. 20, October 1995, pp. 2146-2148.

[23] P. D. Townsend, J. G. Rarity and P. R. Tapster, “Enhanced Single Photon Fringe Visibility in a 10 km-long Prototype Quantum Cryptography Channel,” Electronics Letter, Vol. 29, July 1993, pp. 1291-1293.

[24] C. Marand and P. D. Townsend, “Quantum Key Distri- bution over Distances as Long as 30 km,” Optical Letter, Vol. 20, No. 16, August 1995, pp. 1695-1697.

[25] H. Zbinden, “Interferometry With Faraday Mirrors for Quantum Cryptography,” Electronics Letter, Vol. 33, 1997, pp. 586-588.

[26] H. Kosaka, A. Tomita, Y. Nambu, N. Kimura and K. Nakamura, “Single Photon Interference Experiment over 100 km for Quantum Cryptography System Using a Balanced Gated-Mode Photon Detector,” Electronics Letter, Vol. 39, No. 16, 2003, pp. 1199-1201.

[27] S. J. D. Phoenix et al., “Multi-User Quantum Crypto- graphy on Optical Networks,” Journal of Modern Optics, Vol. 42, No. 6, January 1995, pp. 1155–1163.

[28] A. Muller, T. Herzog, B. Huttner, W. Tittel, H. Zbinden, and N. Gisin, “Plug and Play Systems for Quantum Cryptography,” Applied Physics Letter, Vol. 70, No. 7, February 1997, pp. 793–795.

[29] E. Waks et al., “Secure Communication: Quantum Cryptography with a Photon Turnstile,” Nature, Vol. 420, London, December 2002, p. 762.

[30] E. Moreau et al., “Single-Mode Solid-State Single Photon

Source Based on Isolated Quantum Dots in Pillar

Microcavities,” Applied Physcis Letter, Vol. 79, No. 18,

October 2001, pp. 2865-2867.

[31] H. Bechmann-Pasquinucci and N. Gisin, “Incoherent and

Coherent Eavesdropping in the 6-state Protocol of


F. GARZIA ET AL.


182

Quantum Cryptography,” Physical Review A, Vol. 59, No.

6, 1998, pp. 1-11.

[32] A. Ekert, “Quantum Cryptography Based on Bell's Theorem,” Physical Review Letter, Vol. 67, 1991, pp. 661-663.

[33] R. J. Hughes, G. L. Morgan and C. G. Peterson, “Practical Quantum Key Distribution over a 48-km Optical Fiber Network,” Physics Division Los Alamos National Liberatory, NM 87545.



An Introduction to RFID Technology

Sanjay Ahuja, Pavan Potti School of Computing, University of North Florida, Jacksonville, USA

E-mail: sahuja, pavan.potti @unf.edu Received March 7, 2010; revised July 2, 2010; accepted July 30, 2010

Abstract RFID technology emerged some time back and was not used that much because of lack of standardization and high costs. Latest technologies have brought costs down and standards are being developed. Today RFID is mostly used as a medium for numerous tasks including managing supply chains, tracking livestock, preventing counterfeiting, controlling building access, and supporting automated checkout. The use of RFID is limited by security concerns and delays in standardization. This paper describes RFID technology and its applications in today’s world. Keywords: RFID, RFID Applications

1. Introduction According to Roy Want in [1], “Radio Frequency Identi-fication Technology (RFID) has moved from obscurity into main stream applications that help speed the handling of manufactured goods and materials”. Barcode is still the dominant player in supply chain industries and departmental stores. However RFID is replacing barcode technology and enjoys the major advantage of being independent of line of sight problems and scanning the objects from a distance. It offers the promise of reduced labor levels, enhanced visibility, and improved inventory management. Walmart has been one of the leaders in the large scale adoption of RFID technology [1,2]. RFID tags have a memory capacity of 16-64 Kbytes which is far more than the barcodes (1-100 bytes) [3] and can store additional data such as manufacturer name and product specifications.

The initial step of RFID was during World War II, when the British used it to identify whether planes belonged to “friend or foe”. Some technical problems resulted in the gunning down of allied planes and since then the use of RFID was limited to Defense and armed forces industries due to the cost factors. New advance-ments in science and technology have enabled usage in commercial applications. Large institutions, such as the US Department of Defense, have since implemented RFID which is now spreading to other organizations and industries [1]. Walmart is the second biggest user of RFID and investing significant resources to develop its applications.

Security problems still prevailing about RFID technology

is the fear that people can easily build RFID readers with lower costs and can read data from an RFID chip without knowledge and maybe even alter the data. For example, someone could use the RFID reader on an inexpensive product and upload the data to a chip that is on an expen-sive product, thereby getting the latter for a lower price. Another example is about retrieving data from unsecured RFID enabled mobiles.

RFID advantages can be briefly explained as follows: Reader can read and write data to RFID tags with

out direct contact and no line of sight problem. Data from the multiple RFID tags are accessed by

the reader by radio waves. No maintenance costs; RFID can work under different

environments and can be used effectively for over 10 years.

Fast read and write with the time taken for read/write being a few milliseconds.

Modern RFID tags are made with very good memory capacities ranging from 16-64 Kbytes which is many times more than a typical barcode.

RFID tags can work with GPRS and has been used for tracking.

RFID tags can also integrate with other technologies. For example, it is used with wireless sensor net-works for better connectivity.

The rest of the paper is organized as follows. RFID principles are discussed in Section 2, Section 3 discusses RFID applications, and Section 4 discusses RFID security and technical solutions. Conclusions are listed in Section 5.

S. AHUJA ET AL. 184 2. RFID Principles

Different types of RFID tags exist, but are broadly classified as active or passive. An active tag requires a power source and is either connected to a powered de-vice or to a battery and is often limited by the lifetime of its source. Being dependent on a powered source puts limitations on Active RFID tags. Cost, size, lifetime make them impractical for regular use. On the other side, Passive RFID is of interest because of the fact they are independent of power source and maintenance.

Passive RFID also have advantages of long life and being small enough to fit into a practical adhesive label. Hence passive RFID tags are used for many applications and this paper focuses more on passive RFID tags. A passive RFID tag consists of mainly three parts: an an-tenna, a semiconductor chip attached to the antenna, and some encapsulation to protect the tag from the environ-ment. As explained before, passive RFID tags don’t carry any powered device and became active only upon expo-sure to external energy. The RFID reader does the work for activating and communicating with the tag. The passive RFID tag antenna captures energy from the reader and is responsible for communicating the data between tag and reader. Roy Want states in [1], “Two fundamentally different RFID design approaches exist for transferring power from the reader to the tag: magnetic induction and electromagnetic (EM wave capture). These two designs take advantage of the EM properties associated with an RF Antenna – the near field and the far field”. Both technologies can transfer enough power to a remote tag, usually the power levels will be in the range of 10μW and 1 mW which is very minimal when compared to regular Intel 4 processor power levels 50W. Near-field is the most common approach used for implementing passive RFIDs, and used for near range communications. It has the physical limitations of range. The range of communication of near field technology depends upon the formula c/2Пef where c is the speed of light and f is the frequency. It has the limitation that frequency of operation increases as the distance decreases. One more limitation is the energy available for induction as a func-tion of distance. These physical limitations have led to far field communication and far field communications depend upon backscattering.

3. Applications of RFID RFID applications are very broad and open in nature. First we discuss daily use applications followed by a case study.

RFID is used as a medium for numerous tasks including managing supply chains, tracking livestock, preventing counterfeiting, controlling building access, supporting automated checkout etc. RFID is also used as a means of

providing security to differentiate pirated copies of video and audio discs by sticking RFID stickers to the discs.

Another widely popular example for RFID application is RFID based toll gates. Electronic payment of toll collecting using E-ZPass is a wide spread application. E-ZPass tags are RFID transponders attached to the car license plate and sends account information to the equipment built into lane-based or open toll collection lanes. The toll system will charge from a pre-entered credit card or sends a check. A latest enhancement to this technology is sending the bill details instantly to the user’s mobile phone. And this technique is also used to track stolen cars and other vehicles by police departments with the use of GPRS and RFID.

Another popular application of RFID is in animal tracking. Using RFID tags to track animals is not a new application, but it has evolved from the usage of detecting of missed cattle to the tracking of its movements and behavior. The RFID tags are even used to control out-breaks of animal diseases. Today technology has trans-formed into human implantation of RFID tags. RFID based wristbands and clothes embedded with RFID tags are used to track prisoners.

The RFID tags are also used in the health care industry; an RFID tag is used to store the patient’s medical history. RFID tag is scanned each time to know the developments and changes of the patient’s health condition and medi-cation. RFID tags are often used for medical transactions. RFID tags can also be used in airline industry to track the baggage of the passengers [4]. Walmart is conducting trials to explore a cart integrated with an RFID reader and a wireless mobile computer authorized to make payments as customers add items to the cart. The system displays prices and then authorizes a batch payment when the customer finishes shopping. If a customers RFID mobile is also tuned with credit details, the pay-ment is also done electronically.

Bluetooth is one potential option for providing connectivity, but its usage is hindered by the time it consumes for device discovery and service discovery processes [5]. Salminen et al. in [5] used RFID technology to enhance Bluetooth connection establishment and com-pared the results with and without using RFID and showed that their approach dramatically increase the performance. Even though Bluetooth is one of the leading means of communicating between devices, the limiting factor for it is the time it takes for device discovery process. And when the user is looking for a specific ser-vice offered by other Bluetooth enabled devices it takes more time and is often unnecessary. So the work in [5] authors suggests that the RFID system be used to as a means to initiate a Bluetooth communication channel between the user’s terminal and the services in the en-vironment. Establishing connection between two Blue-tooth devices is a two step process. The first step is to search for the devices in its neighborhood called Device


S. AHUJA ET AL. 185 Discovery, and the second step is to look for the available services and their characteristics called Service Discovery. So to decrease the time of communication, Salminen at al in [5] the stored address and the attributes of the pro-vided service in RFID tags so that the Bluetooth con-necting device is quickly aware of the services offered by other devices. A typical Bluetooth device takes about 10.24 seconds for connecting with other Bluetooth enabled devices and some times it exceeds that time with multiple Bluetooth devices in the environment. Com-pared with Bluetooth, RFID takes only a few milliseconds for communication which is much faster. Another research area for RFID is in the field of Wireless Sensor networks which are a mixture of both sensors and RFID tags and are used for better connectivity and communica-tion [6]. RFID is also used for Activity Recognition and Visual Tracking [7]. 4. RFID Security and Technical Solutions

4.1. RFID Security

The major and primary security concern of RFID is that anyone can access the RFID data because there is no line of sight problem and be able to gather data. In addition, people are cloning RFID tags and using them just as the way it was done for credit cards before. Preventing effective cloning of RFID tags is still an open and challenging problem. Criminals with RFID readers could scan crowds for high-value banknotes. And terrorists could scan digital passports to target specific nationalities.

Currently the research is on-going on RFID malware [8]. RFID malware can be grouped into three distinct categories: exploits, worms, and viruses. RFID exploits are traditional hacking attacks that are identical to those found on the Internet like buffer overflows, code inser-tion, and SQL injection attacks. RFID worms and viruses are simply RFID exploits that copy the original exploit code to newly appearing RFID tags. The main difference between the two is that RFID worms rely on network connections whereas RFID viruses do not. 4.2. Technical Solutions One of the problems of RFID tags is that customers often forget to remove the tags from clothes after purchase and this gives the chance of tracking customers. The better solution is to use EPC kill command as a pro-privacy technology after selling the products. Another alternative to prevent leaking of data from RFID tags is the use of cryptography as measure of privacy. This in turn results in an additional problem of key management and the level of encryption standards and its cost. A different approach is using Tag passwords so that a tag could emit important information only if receives the right password.

The dilemma is in the reader having to know the tag identity. Another solution is using a timer based mecha-nism that the causes the tag to change the password periodically with a predefined mechanism. Another solu-tion is the use of Blocker tags, i.e. using two tags and blocker tag creates an RF environment that is hostile to RFID readers. But a simple and effective solution to prevent leakage of data from RFID tags is differentiating the reader with their energy levels. This was based on assumption that criminals will maintain more distance than valid RFID readers and the power levels will be different.

For details on RFID security protocols, readers are referred to [9,10]. 5. Conclusions RFID is still in a developing phase and more is in the pipeline in terms of new applications. Among applica-tions already developed, RFID tags are being used in clothing for billing and security purposes. RFID tags are embedded inside animals for tracking purposes. RFID tags embedded in uniforms can be used to know the number of hours an employee spends to complete a par-ticular task. There are several associations that are pro-testing against the use of RFID to track people fearing the impact on people’s social life and privacy. Clearly the extent to which use RFID is to be used is still an open debate.

A lot of research on RFID tags is ongoing including on embedding these with other devices, especially mobile devices. RFID manufacturers and users are looking for proper standardization and regulation of RFID. As prices fall further and technological improvements con-tinue to occur, RFID technology is expected to become economically and technically more viable and impact our daily lives as more applications are developed. 6. References

[1] Roy Want, “An Introduction to RFID Technology,” IEEE Communications Society, Santa Clara, Vol. 5, No. 1, 2006, pp. 25-33.

[2] Ron Weinstein, “RFID: A Technical Overview and Its Application to the Enterprise,” IT Professional, Vol. 7, No. 3, June 2005, pp. 27-33.

[3] Klaus Finkenzeller, “RFID Handbook,” 2nd edition, John Wiley & Sons, Ciudad Real, 2003.

[4] Badri Nath, Franklin Reynolds and Roy Want, “RFID Technology and Applications,” IEEE Communications Society, Santa Clara, Vol. 5, No. 1, 2006, pp. 22-24.

[5] Timo Salminen, Simo Hosio and Jukka Riekki, “Enhanc-ing Bluetooth Connectivity with RFID,” Proceedings of the Fourth Annual IEEE International Conference on


S. AHUJA ET AL.


186

Pervasive Computing on Pervasive Computing and Communications, Pisa, 2006, pp. 6-41.

[6] Lei Zhang and Zhi Wang, “Integration of RFID into Wireless Sensor Networks: Architectures, Opportunities and Challenging Problems,” Proceedings of the 5th In-ternational Conference on Grid and Cooperative Com-puting Workshops, Hunan, 2006, pp.463-469.

[7] N. Krahntoever, J. Rittscher, P.Tu, K. Chean and T. Tomlin-son, “Activity Recognition Using Visual Track-ing and RFID,” Proceedings of the 7th IEEE Workshop on Applications for Computer Vision, Vol. 1, New York, 2005, pp. 494-500.

[8] Melanie R. Rieback, Bruno Crispo and Andrew S. Tanen- baum, “RFID Malware,” IEEE Security and Privacy, Vol. 4, No. 4, 2006, pp. 70-72.

[9] Hyun-Seok Kim, Jeong-Hyum Ob and Jin-Young Choi, “Security Analysis of RFID Authentication for Pervasive Systems using Model Checking,” Proceedings of the 30th Annual International Computer Software and Applica-tions Conference, Vol. 2, Chicago, 2006, pp. 195-202.

[10] Hyun-Seok Kim, Jung-Hyun Oh, Jin-Young Choi and Jin-Woo Kim, “The Vulnerabilities Analysis and Design of the Security Protocol for RFID System,” Proceedings of the 6th IEEE International Conference on Computer

and Information Technology, Seoul, 2006, p.152.

[11] Simon L. Garfinkel, Ari Juels and Ravi Pappu, “RFID Privacy: An Overview of Problems and Proposed Solu-tions,” IEEE Security and Privacy, Vol. 3, No. 3, Massa-chusetts, 2005, pp.34-43.

[12] F. Niederman, Richard G. Mathieu, R. Morley, and Ik-Whan Kwon, “Examining RFID Applications in Sup-ply Caching Management,” Communications of the ACM, Vol. 50, No. 7, July 2007, pp. 92-101.

[13] Z. Nochta, T. Staake and E. Fleisch, “Product Specific Security Features Based on RFID Technology,” Pro-ceedings of the International Symposium on Applications and the Internet Workshops, Phoenix, 2005, pp. 4-75.

[14] R. Chandramouli, T. Grance, R. Kuhn and S Landau, “Security Standards for the RFID Market,” IEEE Security and Privacy, Vol. 3, No. 6, McLean, 2005, pp. 85-89.

[15] K. Michael and L. McCathie, “The Pros and Cons of RFID in Supply Chain Management,” Proceedings of the International Conference on Mobile Business, Sydney, 2005, pp. 623-629.

[16] Melanie R. Rieback, B. Crispo and Andrew S. Ta-nenbaum, “The Evolution of RFID Security,” IEEE Per-vasive Computing, Vol. 5, No. 1, 2006, pp. 62-69.



A Security Transfer Model Based on Active Defense Strategy

Zheng Ying Wuhan Maritime Communication Research Institute, Wuhan, China

E-mail: [email protected] Received June 13, 2010; revised July 18, 2010; accepted July 25, 2010

Abstract This paper proposes a security transfer model founded on the active defense strategy. In the unit of security domains of dynamic overlaying routers, this model defines the intelligent agent/ management of network element together with the self-similar, hierarchical and distributed management structure. Furthermore, we use deceptive packets so that the attackers can not trace back to the encrypted data packets. Finally, according to the digested information from data packets, this model is capable of detecting attacks and tracing back to the attackers immediately. In the meantime, the overlaying routers in the security domain are dynamically administered. In summary, this model not only improves the security of data transfer on the web, but also enhances the effectiveness of the network management and switching efficiency of routers as well. Keywords: Active Defense, Digested Information, Routing Control, Deceptive Packets, Hierarchical

Management

1. Introduction To enhance the network security, transfer protocols, such as L2TP, IPSec, TLS/SSL, SOCKSv5 are widely used in the design of terminal system. However, the network attacks are not only aiming at terminal systems: when transferred on line, the data can also be easily sophisticated or disturbed by attackers, which makes the receiver unable to get the accurate packets; fixed transfer ports and addresses are more vulnerable when the communica-tion is going on between the sender and the receiver. Thereby, the security of network entity in the process of transferring data must be taken into consideration [1-22]. At present, there are mainly two kinds of the safety pre-cautions together with corresponding network security productions: open model (as data encryption); passive defense model (as firewall). As the network attack is no bordered, pervasive and secluded, the above-mentioned technologies can not completely solve the network secu-rity problems. Hence, further improvements, like proto-cols of IntServ/RSVP, DiffServ and MPLS, are proposed on purpose of satisfying security requirements. However, disappointingly, people are still searching for absolute solution to the network security problem concerning data transfer up to the present.

Therefore, we would like to bring forward a security

transfer model on base of active defense strategy (as shown in Figure 1). In this model, we define the intelli-gent agent /management of network element together with its self-similar, hierarchical and distributed manage-ment structure and the protocols of IP data packets and hierarchical routing management. Besides, it constructs the dynamic security management mechanism in the unit of security domain of dynamic overlaying routers, accordingly realize the active defense against network attacks. In conclusion, this model not only strengthens the network security on the data transfer, but also ad-vances the network management and packets switching efficiency of routers as well.

2. Topological Structure of the Security Transfer Model

As shown in Figure 1, the network is divided into several adjacent virtual autonomous communities, namely security communities, each with a particular number to represent them (the parts surrounded by curve in Figure 1). The diameter of every community is set from 4 to 7 hops. Each autonomous community possesses edge routers, overlaying routers, spare overlaying routers, and manage-ment nodes which administer malfunction, configuration,

Z. YING 188

E R

E R

E R

E R

E R

E R

E R

E R

E R

E R

E R E R

E R

E R

O R.

S O R

.

. .

M N

M NM N

M N

M N

M N M N

..

M S i ,1 M S i ,2 M S i ,j. . .

M S i+ 1 ,1

T S

Figure 1. Topological structure of network.

performance and security of routers within this network area. 4 to 7 neighboring communities make up one vir-tual security clustering domains equipped with particular clustering numbers. The management nodes in clustering domains are administered by the management server in the upper layer, making the agent processes capable of governing the agent processes in the lower layer, viz. intelligent agent and management of network element. Moreover, we would like to define 4 to 7 neighboring virtual security clustering domains as one virtual security group domain holding unique group number. The manage-ment server in the upper layer administers the servers within the group domain, while enabling its agent processes to run the processes in the lower layer. In this way, ana-logously, we can construct the structure layer by layer. Considering the similarity, we would like to focus on only four layers of the management structure [4].

Besides, information exchange can be carried out in the way of SNMPv3 between network nodes. Due to dispersive management method adopted in this protocol, the network traffic is obviously reduced. In the meantime, the security is enhanced in that exchanging information is encrypted and only authorized personnel can execute the network management function and access encrypted information.

Edge routers (ER) locate on the borderline of auto-nomous communities. Their functions include:

1) identification of the category of data packets entering and exiting communities, including ordinary packets, encrypted packets and deceptive packets, data packets switch;

2) digesting the information of security packets and checking the integral security of the local community and detection of the service quality parameters of the whole community according to the requirements of

managements nodes in the local community; 3) submitting the detected state information to the

management nodes in the local community. Overlaying routers (OR) are capable of : 1) identification of the category of the packets, data

packets switch; 2) checking the integrality or security performance of

the security packets via information digest according to the requirements of managements nodes together with further examination of the service quality parameters of the community ;

3) submitting the state information to the management servers in the local community.

Spare overlaying routers (SOR) are responsible for : 1) switching only ordinary data packets and deceptive

data packets; 2) checking the integrality or security performance of

the security packets via information digest according to the requirements of management nodes together with the examination of the service quality parameters of the community;

3) submitting the detected and monitored state infor-mation to the management servers in the local commu-nity.

Management nodes (MN) function in the following aspects:

1) punctually or randomly sending detection instruc-tion to ER, OR or SOR, and recording the state informa-tion about the security performance and service quality submitted by routers into table Le, Lo and Ls ;

2) dynamically administering all kinds of routers in the community, starting up SOR to replace the damaged OR and make necessary modification to table Lo and Ls

according to the selective principle of overlaying routers when OR are possibly attacked;

3) synthetically scheduling the security packets which


Z. YING 189

k

enter the community through routers, namely designating the transfer path of the packet within the community, evaluation of the service quality and security performance of the whole autonomous community ;

4) submitting the evaluation of state information to the management server in the upper layer.

The selective principles of overlaying routers are: Security: the router is not attacked after checking the

security; Connectivity: packets entering the community via any

ER can be switched to any other ER passing through OR in the community;

Least cost: use overlaying routers as few as possible which should meet the service requirements on the con-dition of accomplishing the connectivity, and the service requirements include parameters such as the throughput, bandwidth and utilization rate.

Management servers (MS), based on the information of the service quality and security performance of the whole autonomous communities submitted by MN in the lower layer, administers the management nodes in the clustering domain or the management servers in the low-er layer, synthetically schedule the security packets in the clustering domain in the meantime, namely specifying the community way within the domain. Besides, MS would evaluate the service quality and security perform-ance of the whole clustering domain and submit the state information to the management servers in the top layer.

Top server (TS), based on the information of the ser-vice quality and security performance of the whole clus-tering domain submitted by MN in the lower layer, syn-thetically schedule the security packets, namely specify-ing the way made up of cluster domains within the group.

In the 4-layer management structure, the router node can be denoted as :

, , ,. .i i j i jMS MN R

i represents the label of MSi, denotes the label of MN

in the clustering domain administered by MSi, k represents the label of routers in the community administered by MNi,j. For example, in the clustering domain administered by the third MS, the No. 5 router administered by the second MN can be symbolized as: .

j

3 3,2 3,2. .MS MN R ,5

3. Protocols of IP Data Packets

In the terminal system based on active defense model [5], data on transferring is distributed to several different ISP connections. For each data stream, certain amounts of fake deceptive packets (deceptive packets for short), on

purpose of active defense and offering security services of different degrees, are appended into the transfer queue where they are mixed with the original encrypted packets (encrypted packets for short) proportionally and stochas-tically. The deceptive packets are capable of beguiling the attacks in that attackers can not tell the original packets from the fake ones. Besides, the methods proposed in literary [6] can be adopted to evaluate the security per-formance and quality of service of autonomous commu-nities and routers according to the state parameters, e.g. throughput, bandwidth and utilization rate. Therefore, data packets on the net can be divided into three categories: ordinary packets, encrypted packets and deceptive packets, of which the latter two are defined as security packets.

To identify the type of data packets and guarantee their integrality, we propose a new format of IP packets as shown in Figure 2. Provided the “protocol” field val-ue in the IP header is 60 (according to RFC1700, 55~60 is unused), this format is employed in the packet, other-wise this packet is of ordinary kind. Moreover, the class field further indicates the type of the data packet: it is a encrypted packet if the field value turns out to be zero after decoded; in case the value changes to 1 after de-coded, it would be a deceptive packet. The access server initially figures out the abstract MD via hash functioning, then encrypts the class field TI, the abstract field MD and user info field M with the public key to this network community. Finally the encrypted data is encased into the user info section of the packet as Figure 2 shows. When the security packets enter into the community, the border router firstly decodes the encrypted parts with the secret key shared among local users, secondly calculates the outcome of hash functioning the user info security M, H(M). Thereby, we can see whether the packet has been attacked or not through comparing H(M) with MD, viz. checking out the integrality and authenticity of the pack-et. When the packet leaves the community, the border router encrypts the class field TI, the abstract field MD and user info field M with the public key to this network community, and then encases the encrypted data into the user info section which is displayed in Figure 2. Besides, the routers switch ordinary packets without special treatment.

The idea of data flow proposed by IPv6 is widely adopted in our security transfer model: the packets flow-ing from a given beginning node to a given end node are of the same traffic; routers on the track which corre-sponds with a specific traffic, are required to designate the security grade and quality of service for the packets of this traffic and process effective management towards

Class field Abstract fieldIP header User info field

Figure 2. Format of data packets.


Z. YING 190 security packets. Considering the facts that the probability of IP packets being separated is less than 0.25% [7] and the MTU discovery implemented by Network tries to avoid fragmenting data packets, we would like to rede-fine the fragment identifier field together with a neighboring unused bit as the flow identifier field and the offset field in the IP header as the security service field. The access server will choose a stochastic figure from 217

－1 numbers as the unique flow label for one packet with encasing it into the flow identifier field subsequently and put the service quality requirements into the security service field as well.

In the security transfer model, routing methods of dif-ferent strategies are adopted: for the original packets, the routing process is strictly executed by levels and grades and the security control over their transfer courses would be achieved; for the deceptive packets, relaxed routing by levels and grades is employed so that we can not only beguile the attacks, but also measure the network states using the packets. Similar to the means of routing using extensible header proposed in IPv6, we introduce the concept of mask and define several fields in the extensible section of the header as the descriptions of route way (as show in Figure 3)

The unused number 10001000 is chosen as the identifier, and this solution, similar to IPv6, however, is not always definite source routing. The length field, counted in 4-bytes, shows clearly the overall extent of the Options section. Concerning the pointer, it denotes the next address to search. This field is set 0 initially, and in-creases by 1 once the data flow passes one router until the number reaches Len-1(Len represents the value of length field). Lastly, as for the mask, which is generated by the management server on the community borderline, the eight bits correspond with 8 addresses in the direc-tory respectively. Nonzero bits denote that the corre-sponding addresses must follow the previous ones. Whereas, 0 bits suggest that the corresponding addresses do not always obey that rule.

Due to the fact that the network diameter of each autonomous community is set from 4 to 7 hops, the maximal length of the field of routing description is less than 40 bytes, viz. the maximal length of the extensible section of IP header. 4. Security Routing Control Protocol

In the security transfer model, we define the intelligent

0Iden tifie r L eng th P oin te r M ask

T he firs t IP ad dress

T he secon d IP address

. . .

8 16 24 3 1

Figure 3. Format of routing description.

agent/ management of network element and management architect of hierarchy and distributed structure. The active defense of the data packets transferring on the net is implemented through hierarchical and distributed dynamic routing control of security packets. Therefore, we would like to discuss the security routing manage-ment protocol on dynamic routing on hierarchical and distributed structure.

The source label, destination label and flow label are respectively denoted as: , , ,. .s s h s h pMS MN R ,

, , ,. .d d e d eMS MN R q and F.

1) routing control protocol in the TS layer Founded on the service quality and security state

information of the whole clustering domain, we establish

a dynamic routing maintenance table (as shown in Table 1). In this table, current routing state of this group domain are recorded, including the security state, desti-nation clustering domain ( destination management serv-er node ), the clustering domain of next hop ( the man-agement server node of next hop) and its cost value. TS designates the security transferring paths for security

packets according to the router table in the unit of clustering domain.

TSL

TSL

When the security packet, of which the flow label is F, is transferred to the destination address which belongs to a clustering domain administered by a different manage-ment server from the source, the transferring path is:

1 2

TSRoute(F,L ) , , ,s l l dMS MS MS MS

In the formula, sMS and dMS denotes source label

and destination label respectively;

represents the clustering domains on the transferring path administered by the management server MS . TS will

broadcast the routing information to correlative manage-ment server MS .

( 1, 2, )ilMS i

il

il

When the source and destination addresses belong to clustering domains administered by the same MSi, routing management for the packet is left to the MSi.

2) routing control protocol in the MS layer On the foundation of the whole service quality and security state parameters of the autonomous communities submitted by the administering MNi,j, a dynamic maintenance rout-

ing table is set up (as shown in Table 2). In that table, current routing states of the clustering do mains

MSL i

Table 1. Route table of management server nodes MSi.

Destination Next hop Cost Security state

1MS 1kMS

1kC 1kS

2MS 2kMS

2kC 2kS


Z. YING 191

Table 2. Route table of management nodes MNij.


,1.i iMS MN 1,.i i kMS MN 1,i kC 1,i kS

,2.i iMS MN 2,.i i kMS MN 2,i kC 2,i kS

jMS ,.pi i kMS MN

, pi kC , Pi kS

administered by MSi, including security state of MNi,j, destination community, the next community and its cost as well.

When the security packet, of which the flow label is F, is transferred to the destination address which belongs to a autonomous community administered by a different management node from the source, the transferring path is:

1 2

MS, ,Route(F,L ) . , . , .i

mi i l i i l i i lMS MN MS MN MS MN ,

.iMS , ( 1, 2, ,ji l )MN j m denote the label of autono-

mous communities administered by MNi, jl on the trans-

ferring path. MSi will broadcast routing information to correlative management nodes.

If the source and the destination addresses are belonging to communities administered by the same management nodes, MNi,j will handle the routing management.

3) route control protocol of the MN layer

On the foundation of the state info table ,

and , we can establish and dynamically

maintain the router table (as shown in Table 3)

of the autonomous communities administered by MNi,j. In this table, current security state of the router Ri,j,k , the destination router, the next overlaying router (including the edge router of the neighboring community) and its cost are recorded. Thereby, according to the route table

, MNi,j will appoint the transfer path in the unit of

routers within the administered community for security packets.

,MNeL i j

,o

MNL i j

,MNL i j

,s

MNL i j

,MNL i j

For the security packets, of which the flow label is F, the transfer path designated on base of route table

is : ,MNL i j

,

1

MN

, , , , , , , , ,

Route(F,L )

, , ,. . . . . .i j

ni i j i j l i i j i j l i i i lMS MN R MS MN R MS MN R

1

)

In the formula, , , , ( 1, 2, ,. .i i j i j lMS MN R n de-

notes the overlaying router labels administered by MNi,j

on the appointed transfer path. 1, , ,. .i i i lMS MN R , repre-

senting the entrance router label of the next community

Table 3. Route table of the nodes Ri,j,k in communities ad-ministered by MNij.


, , ,. .i i j i jMS MN R 1 1, , ,. .i i j i j kMS MN R 1, ,i j kC

1, ,i j kS

, , ,. .i i j i jMS MN R 2 2, , ,. .i i j i j kMS MN R

2, ,i j kC

2, ,i j kS

,.i iMS MN 1, , ,. .i i i lMS MN R 1, ,i lC

1, ,i lS

adjacent to the one governed by MNi,j, will not be in-cluded when the source and destination addresses are within communities administered by the same MNi,j. MNi,j will send the routing information and mask generated based on quantity of addresses to the entrance router. Subsequently, this router will orderly write the IP ad-

dresses, corresponding with items of ,

into the routing address list of the security packet F , and the mask into the mask district.

,MNRoute(F,L )i j

Founded on the above mentioned protocols, we would like to propose the process of routing control of security packets:

1) If the destination label and the source label of the security packets are within the autonomous communities administered by the same management node, routing management are enforced directly by the management node, otherwise go to 2);

2) The management node will submit the source label, destination label and flow label of the packet to the manage-ment server in the upper layer. When the source label and the destination label of the security packets are with-in clustering domains governed by the same management server, then the server will handle the routing management, send the routing information to the management node in the lower layer, and finally the routing information is sent to the entrance router by the management node; or else administer 3);

3) Management servers submit the source label, desti-nation label and flow label of the packets to the Top management server which will administer the routing management. Moreover, the routing information is passed from the top server to management servers in the lower layer, then to the management nodes in the layer below, and reaches the entrance routers at last.

To sum up, the active defense characteristics of the security transfer model proposed in our article are mainly demonstrated in the following aspects: firstly, we define the intelligent agent /management of network element together with the self-similar, hierarchical and distributed management structure and the protocols of IP data packets and hierarchical routing management. In addition, this paper puts forward the dynamic security management


Z. YING


192

mechanism in the unit of security domain of dynamic overlaying routers. Secondly, the encrypted packets are mixed with the fake ones before they are transferred, which makes the attackers incapable of tracing back to the encrypted packets sender. Thirdly, information digested from security packets is used to detect and localize the attacks duly and effectively in the unit of autonomous communities. Besides, due to the manage-ment mode of hierarchy and distribution adopted in the our proposal, routing management of each layer can adapt itself to different requirements, which not only mobilizes the management, but also greatly reduces the items recorded in router tables, thereby enhances the switching efficiency of routers. 5. References

[1] Nagswara S. V Rao, et al. “NetLets: Measurement-Based Routing Daemons for Low End-to-end Delays Over Net-works,” Computer Communications Vol. 26, No. 8, 2003, pp. 834-844.

[2] D. Nesset, “Factors Affecting Distributed System Secu-rity,” IEEE Transactions on Software Engineering, Vol. SE-13, No. 2, 1987, pp. 233-248.

[3] Haixin Duan and Jianping Wu, “Entity Security Architect Structure in Computer Network,” Transactions on Com-puter, Vol. 24, No. 8, August 2001, pp. 147-155.

[4] V Paxsion1 “End-to-end routing behavior in the Internet1 IEEE/ACM Transaction on Networking,” Vol. 5, No. 5, 1997, pp. 601-615.

[5] L. Matthew, “Mission-Critical Network Planning,” Artech House Inc., London, 2003.

[6] S. Muftic and M. Sloman, “Security Architecture for Distributed Systems,” Computer Communications, Vol. 17, No. 7, 1994, pp. 492-500.

[7] D. Nesset, “Factors Affecting Distributed System Secu-rity,” IEEE Transactions on Software Engineering, Vol.

13, No. 2, 1987, pp. 233-247.

[8] Joe, “Touch Dynamic Internet Overlay Deployment and Management Using the X-Bone,” Computer Networks, Vol. 57, No. 5, 2001, pp. 117-135.

[9] R. Forder, The Future of Defense Analysis,” Journal of Defense Science, Vol. 2, No. 1, 2000, pp. 215-226.

[10] Kimberly Holloman, “The Network Centric Operations Conceptual Framework,” Proceeding of Network Centric Warfare 2004 Conference, Washington D. C., 2004, pp. 3-12.

[11] A. S. Tannenbaum, “Computer Networks,” 4th Edition, Machine Press, Beijing, 2004.

[12] D. Comer and D. Stevens, “Internetworking with TCP/IP, Volume II: Design Implementation and Internals,” 2nd Edition, Prentice Hall, New Jersey, 1994.

[13] S. Giordano, M. Potts and M. Smirnov, “Advances in QoS,” IEEE Communications Magazine, Vol. 41, No. 1, 2003, pp. 137-141.

[14] S. Kent and R. Atkinson, “IP Security for the Internet Protocol,” 1998. http://www.ietf.org/rfc/ rfc24011.txt

[15] M. Leech, M. Ganis, Y. Lee, et al. “SOCKS protocol,” 1996. http://www.ietf.org/rfc/ rfc1928.txt

[16] T. Dierks and C. Allen. “The TLS Protocol,” 1999. http://www.ietf.org/rfc/ rfc2246.txt

[17] F. Chung, “Reliable Software and Communication I: An Overview,” IEEE Journal on Selected Areas in Commu-nications, Vol. 12, No. 1, 1994, pp. 23-32.

[18] B. Coan and D. Heyman, “Reliable Software and Com-munication III: Congestion Control and Network Reli-ability,” IEEE Journal on Selected Areas in Communica-tions, Vol. 12, No. 1, 2002, pp. 40-45.

[19] Elham Ghashghai and Ronald L. Rardin, “Using a Hybrid of Exact and Genetic Algorithms to Design Survivable Networks,” Computers and Operations Research, Vol. 29, No. 1, 2002, pp. 53-66.



Optimization of UMTS Network Planning Using Genetic Algorithms

Fabio Garzia, Cristina Perna, Roberto Cusani INFOCOM Department, SAPIENZA – University of Rome, Rome, Italy

E-mail: [email protected] Received May 8, 2010; revised June 8, 2010; accepted June 20, 2010

Abstract The continuously growing of cellular networks complexity, which followed the introduction of UMTS tech-nology, has reduced the usefulness of traditional design tools, making them quite unworthy. The purpose of this paper is to illustrate a design tool for UMTS optimized net planning based on genetic algorithms. In par-ticular, some utilities for 3G net designers, useful to respect important aspects (such as the environmental one) of the cellular network, are shown. Keywords: UMTS Network Planning, Genetic Algorithms

1. Introduction The extraordinary growth of mobile telecommunication sector of the last years has implied strong economical investments of enterprises that operate in this vital sector, in particular way from the net infrastructure point of view.

The development of third generation mobile commu-nication (3G) such as UMTS, with the related advanced allowed services, has increased the need of an efficient network planning that could keep into account all the aspects of complexity which are typical of this new technology, changing the traditional approach to this kind of problem [1-3].

In fact, even if the WCDMA techniques used by UMTS reduce the problems related to the frequency management, the capacity of the net represents a vital problem since the capacity of each radio cell is strongly related to the signal – interference ratio (SIR), that is a function of the number and of the kind of active users inside each communication cell [2,3].

The need of reduction of radiated power, due to envi-ronmental restrictions, and the need of guaranteeing a good quality of services, requires a capillary distribution of Radio Base Stations (BSs) on the territory to be covered. Nowadays, due to the reduced availability of BSs place-ment zones, it is necessary to seek new and efficient methods to optimize the cellular coverage services.

Different and interesting solutions have already been proposed. One of the most interesting is based on a tech-nique inspired to the natural evolution, represented by

the Genetic Algorithms (GAs) [4-21], which are good candidates, thanks to their versatility, to solve a complex and multi-parametric problem such as the considered one.

The purpose of this work is to illustrate a new GAs based method to solve the optimization coverage and capacity problem of UMTS system, keeping into account its specific features and the typical restrictions found in real situations, such as the environmental one. 2. The Genetic Algorithms Genetic algorithms are considered wide range numerical optimisation methods which use the natural processes of evolution and genetic recombination. Thanks to their versatility, they can be used in different application fields.

The algorithms encode each parameters of the problem to be optimised into a proper sequence (where the alphabet used is generally binary) called a gene, and combine the different genes to constitute a chromosome. A proper set of chromosomes, called population, under-goes the Darwinian processes of natural selection, mating and mutation, creating new generations, until it reaches the final optimal solution under the selective pressure of the desired fitness function.

GA optimisers, therefore, operate according to the following nine points:

1) encoding the solution parameters as genes; 2) creation of chromosomes as strings of genes; 3) initialisation of a starting population; 4) evaluation and assignment of fitness values to the

F. GARZIA ET AL. 194 individuals of the population;

5) reproduction by means of fitness-weighted selection of individuals belonging to the population;

6) recombination to produce recombined members; 7) mutation on the recombined members to produce

the members of the next generation. 8) evaluation and assignment of fitness values to the

individuals of the next generation; 9) convergence check. The coding is a mapping from the parameter space to

the chromosome space and it transforms the set of parameters, which is generally composed by real numbers, in a string characterized by a finite length. The parameters are coded into genes of the chromosome that allow the GA to evolve independently of the parameters themselves and therefore of the solution space.

Once created the chromosomes it is necessary choose the number of them which composes the initial popula-tion. This number strongly influences the efficiency of the algorithm in finding the optimal solution: a high number provides a better sampling of the solution space but slows the convergence.

Fitness function, or cost function, or object function provides a measure of the goodness of a given chromo-some and therefore the goodness of an individual within a population. Since the fitness function acts on the parameters themselves, it is necessary to decode the genes composing a given chromosome to calculate the fitness function of a certain individual of the population.

The reproduction takes place utilising a proper selec-tion strategy which uses the fitness function to choose a certain number of good candidates. The individuals are assigned a space of a roulette wheel that is proportional to they fitness: the higher the fitness, the larger is the space assigned on the wheel and the higher is the prob-ability to be selected at every wheel tournament. The tournament process is repeated until a reproduced popu-lation of N individuals is formed.

The recombination process selects at random two individuals of the reproduced population, called parents, crossing them to generate two new individuals called children. The simplest technique is represented by the single-point crossover, where, if the crossover probability overcome a fixed threshold, a random location in the parent’s chromosome is selected and the portion of the chromosome preceding the selected point is copied from parent A to child A, and from parent B to child B, while the portion of chromosome of parent A following the random selected point is placed in the corresponding positions in child B, and vice versa for the remaining portion of parent B chromosome.

If the crossover probability is below a fixed threshold, the whole chromosome of parent A is copied into child A, and the same happens for parent B and child B. The crossover is useful to rearrange genes to produce better combinations of them and therefore more fit individuals.

The recombination process has shown to be very important and it has been found that it should be applied with a probability varying between 0.6 and 0.8 to obtain the best results.

The mutation is used to survey parts of the solution space that are not represented by the current population. If the mutation probability overcomes a fixed threshold, an element in the string composing the chromosome is chosen at random and it is changed from 1 to 0 or vice versa, depending of its initial value. To obtain good results, it has been shown [4] that mutations must occur with a low probability varying between 0.01 and 0.1.

The converge check can use different criteria such as the absence of further improvements, the reaching of the desired goal or the reaching of a fixed maximum number of generations. 3. Definition of the Problem It is evident that, thanks to their versatility, GAs represent good candidates to solve the typical optimization problem of UMTS cellular net planning.

GAs have already been used for this kind of problem [5-21], even if their application is limited only to terri-tory coverage. On the contrary, in this paper, other parameters (such as SIR), that strongly influence the results in real situations, are considered, generating a powerful tool for optimal net planning.

Some general criteria have been adopted, without re-ducing the generality of the problem, which are:

1) it has been considered a suburban area whose dimensions are 3 km x 3 km with an inhomogeneous traffic distribution, even if the proposed algorithm is suitable for different shaped areas;

2) high gain BSs, placed at the same height, are con-sidered;

3) circular irradiation diagrams of BSs, instead of three-lobes diagrams, are considered. This assumption, made to simplify the implementation of the algorithm, does not influence the final result;

4) a consolidated electromagnetic propagation model [22] has been adopted;

5) the SIR has been calculated using the following formula [3]:

r

in out

PSIR SF

I I η

(1)

Figure 1. Operation scheme of a GA iteration.


F. GARZIA ET AL. 195 where SF is the Spreading Factor, Pr is the received power, Iin is the intra-cells interference, Iout is the in-ter-cells interference, η is the thermal noise. 4. Proposed Algorithms for Optimization

Problem Since a plenty of goals and restrictions must be respected in a UMTS net, the design can be made following dif-ferent criteria.

The designer can therefore have different optimization tools that allow him to consider, in each real situation, the predominant aspects.

For this reason, in this paper, the different mentioned real situations have been considered, showing the great flexibility of the proposed method. 4.1. Case 1 A situation without information about traffic level, without restrictions about the maximum number of BSs that can be used and without restrictions about their ter-ritorial placement is considered.

The goal of this case is the optimization of territorial coverage, neglecting the performance of the service as-pects.

To reach this target it is necessary to find a proper fit-ness function of GA and a proper chromosome.

The BSs are coded, inside the chromosome, by means of 2 double vectors, that represents the coordinates of each BS on the territory. To determine the length of the chromosome, related to the number of considered BSs, the minimum number of BSs necessary to ensure the coverage of a given percentage pT of the territory, is calculated as:

min T Tot BSn_bs p A C (2)

where ATot represents the area of the considered territory; pT is the percentage of territory that must be covered; CBS is the maximum coverage area of each BSs.

Due to the usual not regular shape of the territory to be covered and to the impossibility of perfectly matching the coverage diagram of near BSs, the value calculated by means of (2) may be not sufficient and it is necessary to consider a proper multiple n, generally equal to two. In the considered situation, we have n_bsmin=23.

Each gene of the chromosome, representing a BSs, is composed by a number k of variables equal to 3: 2 are used for the position of the BSs on the territory and 1 is used to represent the state of activation /deactivation of the BSs.

The length λ of the chromosome (in term of number of variables) in the considered situation is expressed by the following formula:

minλ n n_bs k (3)

Substituting the numerical values, we have: λ= 138. The fitness function (Ffit) to minimize is, in this situa-

tion: Tot Cov L

fitTot Tot min

A - A O n_bsF α β γ

A A n n_bs

(4)

where ACov is the sum of the coverage areas of the BSs placed on the territory, OL is the sum of the superposition areas of radiation diagram of BSs, n_bs is the number of BSs placed on the territory, α, β and γ are weight coeffi-cients that are varied as a function of the project goals. 4.2. Case 2 In real situations, the traffic inside a territory is not dis-tributed in a homogeneous way. The concentration users zones are named hot spots. It is evident that, to guarantee a certain QoS level, it is necessary to reduce, as more as possible, the intra-cells and inter-cells interference. As a consequence, placing a BS in a hot spot represents a first significant step in net optimization.

Given a non homogeneous traffic distribution and an initial numbers of BSs, calculated according to (2), the algorithm is capable of maximizing coverage and capacity and of minimizing cost.

The fitness function to minimize in this case is: f

Tot Cov L Tot Cov

Tot Tot min Tot

F

A - A O n_bs U -Uα β γ δ

A A n n_bs U

(5) where UTot is the number of estimated users inside the territory and UCov is the number of users covered by the active BSs. 4.3. Case 3 In real situation, for environmental reasons, it is not possible to place BSs anywhere. In this case, only a lim-ited number of zones is available and it is necessary to find a function that accepts, as inputs, not only informa-tion concerning traffic but also information concerning the available installation zones (in particular their coor-dinates). The function must optimize the net considering these limitations that is a cost vinculum. Its structure is therefore equal to the one of (5) less the cost factor. 4.4. Case 4 Another crucial factor in UMTS system is represented by the radiated power (environmental restrictions), with particular respect to the QoS. Therefore the net needs, sometimes, to place the BSs on the territory to reduce, as more as possible, the emitted power, guaranteeing an acceptable level of QoS.

In this case the power of each BS is considered as in-


F. GARZIA ET AL. 196 put parameter (which can be properly changed), that in-fluences not only the coverage area but also the trans-mission capacity.

The fitness function is therefore: f

Tot Cov L Tot Tot Cov

Tot Tot Max Tot

F

A - A O P U -Uα β γ δ

A A n_srb P U

(6) where PTot is the total power of BSs and PMax in the

maximum power radiated by each BS.

5. Performance of the Algorithms and Results

In the following the results of each situation considered above are shown. 5.1. Case 1 Purpose of case 1 is the optimization of the net considering only the coverage of the territory, keeping into account the cost factor. The results obtained are shown in the following.

A first situation has been obtained considering the following values for the weights of fitness function: α = 0.6, β=0.1, γ=0.3. The results are shown in Figure 2. It is possible to see that the presence of a strong cost compo-nent has heavily penalized the coverage maximization.

A second situation has been obtained considering the following values for the weights of fitness function α=1, β=0, γ=0, that is to maximize coverage considering the cost as a quasi-neglectable factor.

Due to the structure of fitness function, it always tends to limit the number of BSs on the territory, evaluating each time if the coverage gain justify the increase of the number of BSs.

(a)

(b)

Figure 2. (a) Initial situation. Units are expressed in kilo- Meters; (b) Final results after 300 generations. Units are expressed in kilo-meters.

(a)

(b)

Figure 3. (a) Initial situation. Units are expressed in kilo- meters; (b) Final results after 300 generations. Units are expressed in kilo-meters.


F. GARZIA ET AL. 197 5.2. Case 2

In this case, given a non homogenous traffic distribution, the fitness function tends to maximize capacity and cov-erage, trying anyway to reduce costs.

A first situation has been obtained considering the fol-lowing values for the weights of fitness function: α=0.3, β=0.1, γ=0.2 e δ=0.5, which is to consider mainly the capacity component. The results are shown in Figure 4.

A second situation has been obtained considering the following values for the weights of fitness function: α=0.5, β=0.1, γ=0.2 e δ=0.3, which is to consider com-plementary situation with respect to the previous one. The results are shown in Figure 5.

5.3. Case 3

In this case, given a limited numbers of zones to place

(a)

(b)

Figure 4. (a) Initial situation. Units are expressed in kilo-meters. The dots represent the users to be reached by the wireless net; (b) Final results after 300 generations. The dots represent the users to be reached by the wireless net. Units are expressed in kilo-meters.

(a)

(b)

Figure 5. (a) Initial situation. Units are expressed in kilo-metres. The dots represent the users to be reached by the wireless net; (b) Final results after 300 generations. The dots represent the users to be reached by the wireless net. Units are expressed in kilo-meters. BSs (environmental restrictions) and a limited number of BSs (26 for example), the maximum coverage is desired. The obtained results are shown in Figure 6. 5.4. Case 4

In this situation, the maximization of coverage and capacity is desired, with a reduction of the emitted power (environmental restrictions).

The results are shown of Figure 7. It is possible to see that the GA places the BSs in the zones where the traffic density is higher, to reduce, as more as possible, the radi-ated power, reducing, obviously, also the coverage area of the BSs, as it is possible to see from Figure 7.



(a)

(b)

Figure 6. (a) Initial situation. Units are expressed in kilo-meters. The dots represent the users to be reached by the wireless net; (b) Final results after 600 generations. The dots represent the users to be reached by the wireless net. Units are expressed in kilo-meters.

(a)

(b)

Figure 7. (a) Initial situation. Units are expressed in kilo- meters. The dots represent the users to be reached by the wireless net; (b) Final results after 1000 generations. The dots represent the users to be reached by the wireless net. Units are expressed in kilo-meters. 5.5. Results The results are shown in Table 1, where it is possible to see that in the most of considered situations, the obtained solutions are satisfying from both coverage and capacity point of view.

The results demonstrate that the GA ensures always high quality results, whose performances increase with the precision of input data. In particular, a significant reduction of number of BSs is always present (cost reduction) even if their initial num-ber is not a given data. This number is always a bit greater than the minimum number of BSs of the consid-ered territory, calculated with (2), due to the impossibil-ity of perfectly matching the circular radiation diagrams of near BSs.

Is also possible to see a certain variability from the coverage point of view while a quasi constant behaviour from the capacity point of view.

The computation time is also quite short, since the most of good solutions are obtained after about 200 ÷ 1.000 generations of GA as a function of the considered situation: the other subsequent generations give only

Table 1. Performance of each considered situation.

Fitness function (Case)

Number of BSs Coverage Capacity

1 A 23 89,3% - 1 B 27 98.1% - 2 A 25 92.3% 99.06% 2 B 25 96.8% 98.12% 3 26 96.9% 98.75% 4 31 94.9% 98.75%


F. GARZIA ET AL.


199

little improvement of quality of solutions.

7. Conclusions

A genetic algorithm based technique to optimize the de-sign of UMTS cellular nets has been presented.

The proposed method considers most of the limits im-posed by the installation of the BSs necessary to guarantee an optimal service, also including environmental restric-tions.

Even if some simplifications were made, the considered technique is capable of ensuring good results from any point of view, representing a useful tool for UMTS initial optimization. 8. References

[1] J. C. S. Cheung, M. A. Beach and J. McGeehan, “Network Planning for Third-Generation Mobile Radio Systems,” IEEE Communications Magazine, Vol. 32, No. 11, No-vember 1994, pp. 54-59.

[2] E. Berruto, M. Gudmundson, R. Menolascino, W.Mohr, and M. Pizarroso, “Research Activities on UMTS Radio Interface, Network Architectures and Planning,” IEEE Communications Magazine, Vol. 36, No. 2, February 1998, pp. 82-95.

[3] E. Amaldi, A. Capone and F. Malucelli, “Planning UMTS Base Station Location: Optimization Models with Power Control and Algorithms,” IEEE Transactions on Wireless Communications, Vol. 2, No. 5, September 2003, pp. 939-952.

[4] D. E. Goldberg, “Genetic Algorithms in Search, Optimiza-tion and Machine Learning,” Addison Wesley, Massa-chusetts, 1989.

[5] L. Nagi and L. Farkas, “Indoor Base Station Location Optimization Using Genetic Algorithms,” IEEE Interna-tional Symposium on Personal, Indoor and Mobile Communications, Vol. 2, London, 2000, pp. 843-846.

[6] B. Di Chiara, R. Sorrentino, M. Strappini and L. Tarricone, “Genetic Optimization of Radio Base-Station Size and Location Using a GIS-Based Frame Work: Experimental Validation,” IEEE International Symposium of Antennas and Propagation Society, Vol. 2, 2003, pp. 863-866.

[7] J. K. Han, B. S. Park, Y. S. Choi and H. K. Park, “Ge-netic Approach with a New Representation for Base Sta-tion Placement in Mobile Communications,” IEEE Ve-hicular Technology Conference, Vol. 4, No. 54ND, 2001, pp. 2703-2707.

[8] R. Danesfahani, F. Razzazi and M. R. Shahbazi, “An Active Contour Based Algorithm for Cell Planning,” IEEE International Conference on Communications Technology and Applications, Beijing, 2009, pp. 122-126.

[9] R. S. Rambally and A. Maharajh, “Cell Planning Using Genetic Algorithm and Tabu Search,” International Con-ference on the Application of Digital Information and Web Technologies, London, 2009, pp. 640-645.

[10] K. Lieska, E. Laitinen and J. Lahteenmaki, “Radio Coverage Optimization with Genetic Algorithms,” IEEE International Symposium on Personal, Indoor and Mobile Radio Communications, Vol. 1, 1998, pp. 318-322.

[11] A. Esposito, L. Tarricone and M. Zappatore, “Software Agents: Introduction and Application to Optimum 3G Network Planning,” IEEE Antennas and Propagation Magazine, 2009, Vol.51, No. 4, pp. 147-155.

[12] L. Shaobo, P. Weijie, Y. Guanci and C. Linna, “Optimi-zation of 3G Wireless Network Using Genetic Program-ming,” International Symposium on Computational Intel-ligence and Design, Vol. 2, 2009, pp.131-134.

[13] C. Maple, G. Liang and J. Zhang, “Parallel Genetic Algorithms for Third Generation Mobile Network Plan-ning,” International Conference on Parallel Computing in Electrical Engineering, Dresden, 2004, pp. 229-236.

[14] I. Laki, L.Farkas and L. Nagy, “Cell Planning in Mobile Communication Systems Using SGA Optimization,” International conference on Trends in Communications, Vol.1, 2001, pp.124-127.

[15] D. B. Webb, “Base Station Design for Sector Coverage Using a Genetic Algorithm with Method of Moments,” IEEE International Symposium of Antennas and Propa-gation Society, Vol.4, 2004, pp. 4396-4399.

[16] A. Molina, A. R. Nix, and G. E. Athanasiadou, “Optimized Base-Station Location Algorithm for Next Generation Microcellular Networks,” Electronics Letters, Vol.36, No. 7, 2000, pp. 668-669.

[17] G. Cerri, R. De Leo, D. Micheli and P. Russo, “Base- Station Network Planning Including Environmental Im-pact Control,” IEEE Proceedings on Communications, Vol. 151, No. 3, 2004, pp. 197-203.

[18] A. Molina, G. E. Athanasiadou and A. R. Nix, “The Auto-matic Location of Base-Stations for Optimized Cellular Coverage: A New Combinatorial Approach,” IEEE In-ternational Conference on Vehicular Technology, Vol. 1, Orlando, 1999, pp. 606-610.

[19] N. Weaicker, G. Szabo, K. Weicker and P. Widmayer, “Evolutionary Multi-Objective Optimization for Base Station Transmitter Placement with Frequency Assign-ment,” IEEE Transactions on Evolutionary Computations, Vol.7, No. 2, 2003, pp.189-203.

[20] G. Fangqing, L. Hailin and L. Ming, “Evolutionary Algo-rithm for the Radio Planning and Coverage Optimization of 3G Cellular Networks,” International Conference on Computational Intelligence and Security, Vol.2, Wash-ington, D. C., 2009, pp. 109-113.

[21] J. Munyaneza, A. Kurine and B. Van Wyk, “Optimization of Antenna Placement in 3G Networks Using Genetic Algorithms,” International Conference on Broadband Communications, Information Technology & Biomedical Applications, Chicago, 2008, pp. 30-37.

[22] M. Hata, “Empirical Formula for Propagation Loss in Land Mobile Radio Services,” IEEE Transactions on Vehicular Technology, Vol. VT-29, No. 3, 1980, pp. 317-325.

Communications and Network, 2010, 2, 200-206 doi:10.4236/cn.2010.23029 Published Online May 2010 (http://www.SciRP.org/journal/cn)


Performance Evaluation of Java Web Services: A Developer’s Perspective

Sanjay Ahuja, Je-Loon Yang School of Computing, University of North Florida, Jacksonville, Florida, USA

E-mail: sahuja, [email protected] Received March 7, 2010; revised July 2, 2010; accepted July 3, 2010

Abstract As the population of the Internet grows rapidly the development of web technologies becomes extremely important. For the evolvement of Web 2.0, web services are essential. Web services are programs that allow computers of different platforms on the web to interactively communicate without the need of extra data for human reading interfaces and formats such as web page structures. Since web service is a future trend for the growth of internet, the tools that are used for development is also important. Although there are many choices of web service frameworks to choose from, developers should choose the framework that fits best to their application based on performance, time and effort for the framework. In this project, we chose four common frameworks to compare them in both qualitative and quantitative metrics. After running the tests, the results are statistically analyzed by SAS. Keywords: Web Service, Framework, Performance, Java, Developer

1. Introduction

For going on trips to other states or countries, the person usually requires to buy airplane tickets, rent a car, and make reservations for hotels to stay at. When dealing with airplane tickets, most of the time the person even has to buy several tickets for some stops instead of a ticket that takes the person directly to the final destina-tion. Looking up the airplane arrival and departure times to connect each flight might take hours of searching and planning. What if there was a virtual agent that could do this all in just a few seconds? So usually people would look for agents to do this for them. But what if this agent was actually is actually a virtual agent online. If the per-son just enters the location he wants to start from, the destination, the desired time for departure or arrival, and all the information required into the computer, in a snap, the computer shows all the results for the person to choose from and purchase the tickets. Even better, such virtual agents could have possibilities to also show information of the car rentals and hotels at the destination and reserve them for you. By using this type of virtual agent, could save much effort and time and can also be more accurate than human agents. Such technology relies on the development and widespread of web services.

Instead of developing a web service application from scratch, there are many open source frameworks that

make development much easier. Which of these frame-works would be a better choice for web service application development? This study compares four popular open source frameworks both qualitatively and quantitatively by doing several tests and analysis. The four frameworks are Apache Axis, JBossWS, XFire, and Hessian. More introductions of web services are done in Section 2. Section 3 describes the four frameworks that are used in this study. In Section 4, the metrics that are used to measure the frameworks are explained in more detail. Section 5 introduces the statistical analysis methods that are use to analyze the measured results. In Section 6, the test results are shown and analyzed. The conclusions are in Section 7. 2. Web Service Frameworks

Since web services are designed to transfer data in com-mon ways, several companies and groups developed web service frameworks for the convenience of web service developers so that they do not need to write a complete web service from scratch. Some of the popular frame-works are Apache Axis, JBossWS, Codehaus XFire, and Caucho Hessian. In this section, these frameworks are going to be introduced.

S. AHUJA ET AL. 201 2.1. Apache Axis Apache Axis (Apache Extensible Interaction System) is an open source, Java and XML based Web service framework created by the Apache Software Foundation (ASF). The foundation is a non-profit corporation that mainly produces software for network use, such as serv-ers and frameworks for servers. Their projects are well known to be collaborative, consensus based development process and free or open source software. The Apache Axis package has an implementation of a SOAP server and API’s for generating and deploying web service applications. The SOAP engine constructs SOAP processors like clients, servers, and gateways. This allows the servers and clients to communicate through SOAP messages. The API supports a variety of lan-guages. Besides the Java version, a C++ implementation is also available. It allows developers to construct their application in a variety of ways. The easiest method only requires changing the file name extension from “.java” to “.jws”. The downside of such a method is lacks flexibility for further configuration.

2.2. JBossWS

JBossWS is JBoss' implementation of J2EE compatible web services. The framework is designed to fit better in the overall JBoss architecture and is generally more suitable for the specific J2EE requirements for web ser-vices. Instead of using the traditional Apache server for this framework, JBoss has a server of its own, and is suggested that the framework is used on this server to get best performance. Similar to ASF, the JBoss community is a group of people that focus on open source projects. Their projects emphasize on the development of Java Enterprise Middleware, which are software that act like bridge between applications, operating systems or both.

2.3. Codehaus XFire

Codehaus XFire is a next-generation java SOAP frame-work. It is a free and open source SOAP framework that allows you to implement web services with great ease and simplicity. It also provides many features identified in web service specifications, which are not yet available in most commercial or open source tools. It is claimed to have higher performance since it is built on a low memory StAX (Streaming API for XML) based model but there is no data to document this fact.

2.4. Hessian

The Hessian binary web service protocol makes developing web services simple and usable without requiring a large framework so that developers would not need to spend more time and effort to learn an alphabet soup of protocols.

Since it is a binary protocol, it works well on sending binary data without any need to extend the protocol with attachments. J2ME devices like cell-phones PDAs can use Hessian to connect to web services with better performance, because it is a small protocol. Hessian was named after the Hessian cloth, which is the British term for Burlap. It was named this way because burlap is simple, practical, and useful, but extremely ordinary material, which is like the characteristics of the Hessian protocol.

3. Evaluation Metrics

Different factors are considered when comparing the four frameworks in this project. Some metrics are to deter-mine the performance and efficiency; some are to show the transparency and abstraction. This section explains these metrics.

3.1. Latency

In terms of network, latency is an expression of how much time it takes for data to be sent back to a request. This includes the time for the request to be sent to the server, the time the server spends on processing the task, and the time for the results to be sent back. The network latency is contributed by many factors, such as propagation, transmission, modem and router processing, and storage delays. The propagation is the time it takes for an object, such as data, to transfer from a location to another in the speed of light. Transmission is the delay from the medium like optical fiber or wireless networks. Modems and routers take time to check the headers of a packet. The storage delay is the time it takes for the actual hardware storage, such as hard drives, to store the received data. In this project, the latency is tested with different scenarios such as requesting 1, 2, 3, 4, and 5 MB of data, and 1, 5, 10, 15, 20 clients simultaneous requesting data. From the results of such testing, trends can be found and compared for each framework.

3.2. Throughput

Throughput is the amount of clients or data processed within a certain unit of time, like a second. It is highly related to latency, since scenarios with high latency would result in low throughput, and scenarios with low latency would result in high throughput. However, by viewing the latency graph, we can only tell the trends of response time, while we can determine the most efficient scenario for a framework through viewing a throughput graph. 3.3. Memory Usage In computing, memory is data storage to temporarily store data for calculations of the computer. There is a


S. AHUJA ET AL. 202 wide variety of memory, such as cache memory, flash memory, random access memory (RAM), virtual memory, etc. Either memory, they are all limited in server due to the cost and space. A framework that uses less memory would have an advantage of allowing higher capacity for the server. 3.4. CPU Usage Central Processing Unit (CPU), also known as processor, is a component in a computer used to interpret program instructions and process data. Although it is only able to process one task at a time, when there are multiple tasks to be done, instead of finishing a task then going to an-other, the CPU is designed to switch to other tasks before finishing one if necessary, so that it would act like if it is executing multiple tasks at the same time. However, large tasks might occupy a lot of CPU time, which de-creases the time scheduled for other tasks. A framework that uses less CPU would allow the server to have more time to execute other tasks.

3.5. Source Lines of Code

The source lines of code (SLOC) used in a framework can indicate the transparency and abstraction of the framework. The main goal of a framework is to save the developer’s time and effort by not having to write the entire code from scratch. Thus the less lines of code that is required for a framework, the more time and effort is saved by this framework. However, lines of code cannot be exactly accurate since some lines might be long while some lines are short. So the number of files and size of files also is a consideration.

4. Statistical Analysis Methods

After retrieving the test data to compare the performances, we need a method to analyze the results. By simply calculating the average response times and making them into a graphs is not sufficient for the analysis. Looking at the average response times 1.5 seconds and 1.6 seconds, we can not be sure if that is a great difference or not. Therefore, statistical analysis methods are required to tell if the difference is significant or not. In this project, the general linear model (GLM) [9] and two-way analysis of variance (two-way ANOVA) is used for statistical analysis. Furthermore, the Statistical Analysis System (SAS) [10] is used as a tool for aiding the calculations of the statistical analyses required.

4.1. The SAS System The SAS system is statistical analysis software that has a wide variety of statistical modules and procedures. They

use a fourth-generation programming language (4GL) for their code and the programs are composed by three main components – the data step, the procedure step, and the macro language. The data step is for entering data, like inserting the data in the code or read from data files. The procedure step is the use of the statistical methods and models to analysis the data that was read in the data step. The macro language is for decreasing the redundancy of functions that are used again and again throughout the program.

4.2. The GLM Model

The GLM model is a statistical linear model that is used in general cases. It is the foundation of many statistical analyses, such as t-test, ANOVA, Analysis of Covariance (ANCOVA), etc. The easiest case to understand how the GLM model works is the two-variable case. The goal of this analysis is to use a way to accurately describe the information in this plot. Using the GLM model, we try to find a straight line that is closest to all the dots in the plot. This line would be written as: y = b0 + b1x + e, where y is the y-axis variable, x is the x-axis variable, b0 is the intercept (the value of y when x equals 0), b1 is the slope of the straight line, and e is the error. By solving b0 and b1, we can get information about this linear line that describes the dots in the plot. In other cases with more than two variables, the formula can be extended as: y = b0 + b1x1 + b2x2 + b3x3 + ... + bnxn + e, where n is the number of variables of the situation. But the mechanism for solving such problems is the same.

5. Results and Analyses In order to get best results from SAS analysis, each case is tested twenty times. Since four frameworks with five different amounts of clients are tested, there are twenty different cases. Adding the twenty test times for the twenty different cases would result 400 data sets to be calculated by SAS. Besides the amount of clients, the size of data is also considered. With also five different sizes of data sent, there would be twenty cases with a total of 400 data sets. The response time is measured by recording the time right before invoking the web service and recording the time right after the data requested is received then subtracting the time difference.

5.1. Results 5.1.1. Client Scenarios For testing the four different frameworks in different scenarios, web service applications to send out data are created. The five scenarios based on amount of clients for testing performance of the four frameworks are 1 client, 5 clients, 10 clients, 15 clients, and 20 clients each retrieving 1 MB of data. The average response time for


S. AHUJA ET AL.


203

results to throughput, Figure 2 shows the average clients per second for each scenario and framework.

each scenario and each framework is recorded for analysis. The results are as shown in Figure 1. By calculating the

Time vs Client

0

5000

10000

15000

20000

25000

1 Client 5 Clients 10 Clients 15 Clients 20 Clients

Res

pons

e T

ime

(Mil

lise

cond

s)

Apache Axis

Resin Hessian

JBossWS

Codehaus Xfire

Figure 1. Latency in client scenarios.

Client per Second

0.000

1.000

2.000

3.000

4.000

5.000

6.000

1 Client 5 Clients 10 Clients 15 Clients 20 Clients

Clien

t pe

r S

econ

d

Apache Axis

Resin Hessian

JBossWS

Codehaus Xfire

Figure 2. Throughput in client scenarios.

S. AHUJA ET AL.204

Figure 2 shows the most efficient client scenario for each framework. Apache Axis can deal with 4.993 clients per second after reaching the scenarios with 10 clients or more. Resin Hessian can deal with 4.807 clients per second in those same scenarios. JBossWS deals with 0.943 clients per second in every scenario. Codehaus XFire seems to

work most efficiently around the scenario of 5 clients, dealing about 2.892 clients per second. 5.1.2. Data Size Scenarios As for the results and average throughput for the five scenarios based on different data size, they are shown in Figures 3 and 4.

Time vs MB

0

1000

2000

3000

4000

5000

6000

1 2 3 4 5

MB Sent

Res

pons

e T

ime

(Milli

seco

nds)

Apache Axis

Resin Hessian

JBossWS

Codehaus Xfire

Figure 3. Latency in data size scenarios. MB per Seconds

0

0.5

1

1.5

2

2.5

3

3.5

4

1 2 3 4 5

MB Sent

MB

per

Sec

ond

Apach Axis

Resin Hessian

JBossWS

Codehaus XFire

Figure 4. Throughput in data size scenarios.


S. AHUJA ET AL. 205

Figure 4 shows the most efficient data size scenario for each framework. All frameworks reach their best performance after data with 2 MB or more are sent. Apache Axis has an average of 3.617 MB/s, JBossWS with 1.287 MB/s, Codehaus XFire with 1.240, and Resin Hessian with 1.017.

From all the graphs above, it would seem like Apache Axis has the best performance in all scenarios, but fur-ther analysis should be done by SAS.

5.2. Analyses

Obviously, the amount of respond time highly depends on the choice of framework, the quantity of data trans-ferred, and the number of clients that are invoking tasks from the web service. That makes these three factors significant to the test results, which is the response time. But before doing further analysis, we have to use the GLM model to make sure if the interactions of the three factors are also significant factors. If the interactions are not, we can use the Tukey method to do multiple com-parison and directly see which framework has better performances in all cases and which has worse; if the interactions are significant factors, than we would need to analyze the results case by case. 5.2.1. Client Scenarios First analyzing the results from the client scenarios, we use the SAS system to see the significance of each factor.

It turns out not only are the amount of clients and choice of framework significant factors, but also is the interaction between them. This means if one of the frameworks is significantly faster than the some others in the some scenarios, the framework would not necessarily be faster than those in other scenarios. Thus, SAS can not directly compare all frameworks in all scenarios.

A pair-wise comparison from the General Linear Model (GLM) procedure is used in this case. The first framework is compared to the second in the first scenario, then the first to third, first to fourth, second to third, second to fourth, and third to fourth. So there would be six com-parisons in each scenario.

Table 1 should be read a scenario at a time, i.e. when we are looking at the 1-client scenario, we ignore the data in the 5-client, 10-client, 15-client, and 20-client scenario. Groups with lower alphabets have lower re-

Table 1. Response time comparison for client scenarios.

1

Client 5

Clients 10

Clients 15

Clients 20

ClientsApache

Axis A A A A A

Hessian A A A A A JBossWS A B C C C Codehaus

XFire A A B B B

sponse time, which means better performance. In the 1-client scenario, all frameworks are put into group A, meaning they all have approximately the same performance in this scenario. In the 5-client scenario, JBossWS is put into group B while the others are in group A. This means in this scenario, JBossWS has worse performance than the others, while the others still are about the same. In the last three scenarios, Apache Axis and Resin Hessian are faster than Codehaus XFire, and Codehaus XFire is faster than JBossWS.

Although from the SAS analysis results the better performance of frameworks is a case by case matter, as the number of clients increase, the performance com-parison trends to be the same, being Apache Axis and Resin Hessian better than Codehaus XFire, and Codehaus XFire better than JBossWS. 5.2.2. Data Size Scenarios The process of analyzing performance based on data size is just the same as the analyzing it based on client amount. First, the interaction between data size and choice of framework is determined.

It turns out that the interaction between data size and choice of framework is also a significant factor. There-fore, the same pair-wise comparison procedure is used again.

When sending 1 MB, Apache Axis is better than Codehaus XFire, which is better than JBossWS, and that is better than Resin Hessian. In the second scenario, the comparison is almost the same but performances of Codehaus XFire and JBossWS are equivalent. In the last three scenarios, JBossWS and Codehaus XFire switch places, making JBossWS faster. Although each scenario is a different case, as the data size increases, the per-formance comparison trend to be the same, being Apache Axis the best, JBossWS the second, Codehaus XFire the third, and Resin Hessian the last. 5.2.3. Others Other metrics such as memory usage, CPU usage, and source lines of code (SLOC) are also tested in this pro-ject. Table 3 shows memory and CPU used on the web service application created using each framework.

Since the web service application created using four frameworks all barely use any CPU at all, CPU usage is not a main factor in this case now. Comparing the memory

Table 2. Response time comparison for client scenarios.

1 MB 2 MB 3 MB 4 MB 5 MB

Apache Axis A A A A A

Resin Hessian D C D D D

JBossWS C B B B B

Codehaus XFire

B B C C C


S. AHUJA ET AL.


206

Table 3. Memory and CPU usage of four frameworks.

Memory Usage CPU Usage

Apache Axis 13% 0%

Resin Hessian 8.7% 0%

JBossWS 16% 0%

Codehaus XFire 13% 0%

Table 4. SLOC of application of four frameworks.

SLOC Server Side Client Side

Apache Axis 64 120

Resin Hessian 70 85

JBossWS 94 127

Codehaus XFire 48 128

usages, Resin Hessian uses the least, which is almost half of JBossWS. Apache Axis and Codehaus XFire use an intermediate level of memory. The SLOC of web ser-vices created using each framework are as shown in Ta-ble 4.

JBossWS requires the most lines of code, and Resin Hessian requires the least. Such a web service application with only one trivial function requires a little amount of code, so the SLOC difference will be that crucial. But if these frameworks are used to create real world large applications, this 42% difference can mean thousands or more of lines, which greatly increase the effort, time, and errors of an application development. 6. Conclusions For web applications that require communication through the network between computers of different platforms, web service would be a good choice since it is designed based on a platform-independent language – XML. Instead of developing web services from scratch, using existing frameworks can greatly increase the pro-ductivity and lower the time and effort that developers spend on learning the details of web services. From the test results of this project, Apache Axis has overall best performance. When processing with small amount of data, Hessian performs just as well as Apache Axis. In contrast, it has the poorest performance of the four frameworks when processing larger amounts of data.

However, Hessian requires the least amount of code and uses the least memory and CPU. Thus, for developing a small application with small amounts of data being processed such as mobile devices, Hessian would be a great choice due to its high performance and low price. If developing a big application that processes high amount of data is the case, Apache Axis would be a better solution. The benefits of JBossWS are that it is more compatible with other JBoss projects or applications based on JBoss Application Server (JBoss AS). 7. References

[1] A. Ching and A. Wagner, “Understanding Performance Testing,” Technical Report, Microsoft Developer Net-work, February 2001.

[2] F. Curbera, M. Duftler, R. Khalaf, W. Nagy, N. Mukhi, and S. Weerawarana, “Unraveling the Web Services Web: An Introduction to SOAP, WSDL, and UDDI,” IEEE In-ternet Computing, New York, March 2002, pp. 86-93.

[3] F. Curbera, R. Khalaf, N. Mukhi, S. Tai and S. Weer-awarana, “The Next Step in Web Services,” Communica-tions of the ACM, New York, October 2003, pp. 29-34.

[4] S. Decker, S. Melnik, F. van Harmelen, D. Fensel, M. Klein, J. Broekstra, M. Erdmann and I. Horrocks, “The Semantic Web: The Roles of XML and RDF,” IEEE In-ternet Computing, New York, September-October 2000, pp. 63-73.

[5] J. Hendler, “Agents and the Semantic Web,” IEEE Intel-ligent Systems, Maryland, March-April 2001, pp. 30-37.

[6] A. McIlraith Sheila, T. C. Son and H. Zeng, “Semantic Web Services,” IEEE Intelligent Systems, New York, March-April 2001, pp. 46-53.

[7] S. Narayanan and S. A. McIlraith, “Simulation, Verifica-tion and Automated Composition of Web Services,” In-ternational World Wide Web Conference, Honolulu, May 2002, pp. 77-88.

[8] J. Yang and M. P. Papazoglou, “Web Component: A Substrate for Web Service Reuse and Composition,” Lecture Notes in Computer Science, Vol. 2348, 2002, pp. 21-36.

[9] General Linear Model (GLM), http://www.statsoft.com/textbook/general-linear-models/

[10] Statistical Analysis with SAS/STAT Software, http://www.sas.com/technologies/analytics/statistics/stat/index.html

http://www.statsoft.com/textbook/general-linear-models/

http://www.sas.com/technologies/analytics/statistics/stat/index.html

http://www.sas.com/technologies/analytics/statistics/stat/index.html

Date post:	10-Oct-2020
Category:	Documents
Upload:	others
View:	3 times
Download:	0 times

9 771949 242004 02 · assume the binary phase shift keying (BPSK) modulation. The result however...

Documents