Improving Voice and Data Service Provisioningin Cellular/WLAN Integrated Networks by Admission Control*

Wei Song†, Yu Cheng‡, Weihua Zhuang†, and Aladdin Saleh§†Dept. of Elec. & Comp. Eng., University of Waterloo, Waterloo, ON, Canada‡Dept. of Elec. & Comp. Eng., University of Toronto, Toronto, ON, Canada§Bell Canada, 5099 Creekbank Road, Mississauga, ON, Canada L4W 5N2

Abstract— In this paper, we study the voice and data serviceprovisioning in an integrated system of cellular and wirelesslocal area networks (WLANs). To maximize the overall resourceutilization of the integrated system, complementary quality ofservice (QoS) support capabilities of the two networks areexploited to serve voice and data traffic. As an essential resourceallocation aspect, admission control can be used to properly admitvoice and data calls to the overlaying cellular cells and WLANs.In this study, a generalized admission scheme is analyzed toinvestigate the dependence of resource utilization on admissionparameters, which vary with user mobility and traffic variability.By applying an effective QoS evaluation approach, the admissionparameters can be determined using a search algorithm.

I. INTRODUCTION

With complementary characteristics presented in the cellu-lar network and wireless local area network (WLAN), theirinterworking becomes a promising trend for next-generationwireless networks [1]. The cellular network provides ubiqui-tous coverage, while WLANs are deployed disjointly in hot-spot areas. In the area with WLAN coverage (referred toas double-coverage area in the following), both cellular andWLAN accesses are available. With this two-tier overlayingstructure, a new call in the double-coverage area can beadmitted either to a cell or WLAN. Proper admission decisionis also needed for vertical handoffs between a cell and WLAN,which is not necessary to maintain a call but mainly forload balancing and/or quality of service (QoS) improvement.The heterogeneous underlying QoS support of the integratednetwork can be exploited to maximize the resource utilization.

In the cellular network, the base station (BS) controlsaccess to the shared radio spectrum and reserves resources foradmitted calls. The centralized control and reservation-basedresource allocation enables fine-grained QoS. On the otherhand, the medium access control (MAC) in WLANs is usuallycontention-based, e.g., the distributed coordination function ofIEEE 802.11. Because of inevitable collision and backoff, thistype of MAC is difficult to support services with strict QoSrequirements such as real-time voice service, although it isefficient in serving bursty data traffic. With the complemen-tary QoS support, multi-class traffic load should be properlydistributed to the cells and WLANs by admission control.Two admission schemes are discussed in [2,3] for voice anddata services in a cellular/WLAN integrated network. In thispaper, we further generalize the above admission schemes andinvestigate the dependence of resource utilization on admission

*This research was supported by a research grant from the Bell UniversityLabs at the University of Waterloo.

parameters. A QoS evaluation approach based on momentgenerating functions (MGFs) is developed to consider highlyvariable data traffic [4] and user mobility within WLANs [5].

The remainder of this paper is organized as follows. InSection II, we describe the system model for this study. SectionIII discusses the admission control problem for cellular/WLANinterworking and the QoS evaluation approach to determine ad-mission parameters. Numerical results are presented in SectionIV. Section V concludes this research.

II. SYSTEM MODEL

We consider an interworking scenario with a simple topol-ogy, in which there is one WLAN located in each cell, referredto as a cell cluster. Within the interworked system, location-dependent factors such as user mobility, traffic distribution, andunderlying network support need to be taken into account.

A. Location-Dependent Mobility Model

As shown in [5], the indoor deployment and low usermobility within WLANs results in heavy-tailed user resi-dence time within a WLAN, denoted by Tw

r . An importantfeature of heavy-tailedness is the so-called “mice-elephants”phenomenon [6]. For the WLAN residence time, it implies thatmost users stay within a WLAN for a short time, while a smallfraction of the users have an extremely long residence time.To explore its impact on performance, we adopt a two-stagehyper-exponential distribution, which well captures the highvariability and makes the analysis tractable. The probabilitydensity function (PDF) of Tw

r with mean 1/ηw is given by

fT wr

(t) =a

a+ 1· 1

1a· 1

ηw

e−aηwt +1

a+ 1· 1

a · 1ηw

e−ηw

at, a ≥ 1. (1)

A large fraction aa+1

of the users stay within the WLAN fora mean time 1

a· 1

ηw , while the other 1a+1

of the users havea mean residence time of a · 1

ηw . The coefficient of varianceof Tw

r is Cv,T wr

=√a+ 1/a− 1. Increasing the parameter a

results in Twr with a higher variability. A simple algorithm

is proposed in [7] to approximate a large class of heavy-taileddistributions (including Pareto and Weibull distributions) withhyper-exponential distributions. Since the hyper-exponentialdistribution consists of a linear mixture of exponentials, thefollowing analysis can be extended to hyper-exponential dis-tributions with more exponential components, which moreaccurately approximate the original heavy-tailed distribution.

On the other hand, the user residence time in the area of acell with only cellular access (referred to as cellular-only area),denoted by T c

r , is assumed to be exponentially distributed

©1-4244-0357-X/06/$20.00 2006 IEEEThis full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE GLOBECOM 2006 proceedings.

wcp

−1

wη

1

cηwc

p−

=

3η

12 1 pp −=

Denote sum of two independent exponentially distributed random

variables with parameters and .2η1ηi) 2η1

η

ii)1p

12 1 pp −=

Denote a random variable which with probability follows an exponential distribution with

parameter and with probability follows a generalized hyperexponential distribution with

parameter (the sum of two exponential random variables with parameters and ). 1η 1p

2p

3η 2

η1

η

:1

c

rT

:2

c

rT

cη cηwη

cc

p−

wη cη cηwη wη cη

1η 2η

1p

1η

wc

p−

cc

p− cc

p− cc

p−

wc

p− wc

p− wc

p−

cc

p− cc

p− cc

p− cc

p−

wc

p−

Figure 1. Modeling of user residence time in a cell.

with parameter ηc. Users moving out of the cellular-only areaenter neighboring cells with a probability pc−c and enter thecoverage of the overlaying WLAN in the target cell with aprobability pc−w = 1 − pc−c. As a result, the residence timeof users admitted in the cell follows more complicate phase-type distributions as shown in Fig. 1. Let T c

r1 and T cr2 denote

the residence time within the cell of a new or handoff calladmitted to the cell from the cellular-only area and that of anew call admitted to the cell from the double-coverage area,respectively. The MGFs of T c

r1 and T cr2 can be derived as

Φ1(s) =∑∞

i=1 (pc−w)i−1pc−c ηc

ηc−s[ψ(s)]i−1 (2)

Φ2(s) =∑∞

i=1 (pc−w)i−1pc−c[ψ(s)]i (3)

where ψ(·) is the MGF of (T cr + Tw

r ), given by

ψ(s) = E[es(T c

r +T wr )

]=

ηc

ηc − s

[a

a+ 1· aηw

aηw − s+

1

a+ 1·

1aηw

1aηw − s

].

B. Traffic Models for Voice and Data Services

As an important motivation for cellular/WLAN interwork-ing, multi-service support is an essential requirement for futurewireless networks. In this study, we consider both voice (e.g.,traditional telephony) and data (e.g., Web browsing) services,which are representative examples of the conversational andinteractive service class of the universal mobile telecommu-nication system (UMTS), respectively. The call-level QoS re-quirements are defined in terms of voice call blocking/droppingprobability (Qvb), data call blocking/dropping probability(Qdb), and data call throughput (Qdt). As widely applied inprevious works, voice and data call arrivals are assumed to beindependent Poisson processes. The voice call duration Tv isexponentially distributed with mean 1/µv. It is observed in [4]that the data traffic generated by Web browsing is bursty andone main reason for the high variability is the heavy-tailedfile size. To explore the impact of highly variable file sizeon performance, similar to the modeling of WLAN residencetime, the data call size (Ld) is modeled by a two-stage hyper-exponential distribution with mean fd and PDF given by

fLd(x) =

b

b+ 1· 1

1b· fd

e− b

fdx

+1

b+ 1· 1

b · fde− 1

b· 1

fdx, b ≥ 1. (4)

The coefficient of variance of Ld is then Cv,Ld=

√b+ 1/b− 1.

A larger b corresponds to a data call with a higher variability.In this study, we consider code-division multiple access

(CDMA) cellular system with variable spreading gain. Thecapacity of the more congested downlink is analyzed based oncell load factor [8], which is limited by the total transmittedpower of the BS. The number of admitted users is limitedto bound the interference level and satisfy the requirementsfor signal-to-interference ratio. With the reservation-based re-source allocation, the restricted access mechanism [9] can beapplied for resource sharing between voice and data in thecell. Given the total cell bandwidth Wc, Nc

v voice calls atmaximum are accommodated to meet the requirement on voicecall blocking/dropping probability, while the other bandwidthis dedicated to data. All bandwidth unused by current voicetraffic is equally shared by ongoing data calls to exploittheir elasticity. Considering the interactive nature, each datacall needs to finish transfer within a time threshold and beguaranteed certain throughput. As a result, the number of datacalls admitted in the cell is limited by Nc

d .On the other hand, with the contention-based access, the

WLAN resources are shared in a complete-sharing mannerand there is no strict QoS guarantee for an individual call.Therefore, it is necessary to apply admission control to restrictthe numbers of admitted voice and data calls contendingfor access. In [2], we derive the WLAN capacity region interms of the maximum numbers of voice and data calls thatcan be simultaneously carried by the WLAN. Different QoSsupport is provided by the WLAN to voice and data callshaving different traffic characteristics and QoS requirements.The WLAN resource utilization changes with the amounts ofvoice and data traffic sharing the WLAN channel, which canbe controlled by limiting the maximum numbers of voice anddata calls admitted in the WLAN, denoted by Nw

v and Nwd ,

respectively. To ensure that the WLAN operates efficientlyand complements the cellular network effectively, Nw

v and Nwd

should be properly selected within the WLAN capacity region.

III. ADMISSION CONTROL FOR CELLULAR/WLANINTERWORKING

With the heterogeneous QoS support of the underlying inte-grated network, the voice and data traffic in a double-coveragearea should be properly directed to the cell and WLAN.Instead of applying complex criteria for network selection, thefollowing simple admission scheme is proposed to investigatethe dependence of resource utilization on traffic distributionand other important traffic and mobility parameters. For anincoming voice (data) call in the double-coverage area, witha probability θw

v (θwd ), it requests admission to the WLAN,

while it requests admission to the cell with a probabilityθcv = 1 − θw

v (θcd = 1 − θw

d ). Once the admission parameters θwv

and θwd (or θc

v and θcd) are determined for a given traffic load,

a mobile station can make a decision on its own and sendthe admission request to the corresponding target network.With the simplicity, the proposed admission scheme can beimplemented in a distributed manner. Also, as the network

©1-4244-0357-X/06/$20.00 2006 IEEEThis full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE GLOBECOM 2006 proceedings.

elements involved in the admission decision are limited to be asfew as possible, the signaling overhead is effectively reduced.

Due to the two-tier overlaying structure, the vertical handofffrom the cell to the overlaying WLAN is not necessary butoptional to maintain an ongoing call. Hence, the handoff trafficload to the WLAN can be controlled by properly adjusting theadmission parameters of the WLAN, e.g., by using a simpleguard channel policy. Let Gw

v (Gwd ) denote the number of

guard channels reserved in the WLAN for new voice (data)calls from the double-coverage area. Then, a handoff voice(data) call from the cell is admitted to the WLAN whenthe number of voice (data) calls in the WLAN is less thanMw

v = Nwv −Gw

v (Mwd = Nw

d −Gwd ), and is rejected otherwise.

A. QoS Evaluation for Voice and Data Services in the WLAN

To determine the admission parameters, the QoS metricsof interest need to be evaluated accurately and effectively.We begin with the WLAN QoS. Let λv1 and λv2 (λd1 andλd2) denote the mean arrival rates of new voice (data) calls inthe cellular-only area and double-coverage area, respectively.Then, the mean arrival rates of new voice and data calls tothe WLAN are λw

nv = θwv · λv2 and λw

nd = θwd · λd2, respectively.

The voice traffic load to the WLAN also includes handoffcalls from the overlaying cell with a mean rate λc−w

hv . WithPoisson arrivals and exponential duration, voice calls in theWLAN can be modeled by an M/G/K/K queue with meanchannel occupancy time E[min(Tv , Tw

r )]. With the insensitivityof M/G/K/K queues [10], the steady-state probability of voicecalls in the WLAN is given by

πwv (j) = πw

v (0)∏j

l=1

λwv (l)

l · µwv

, j = 1, ..., Nwv (5)

where1

µwv

= E[min(Tv , Twr )] =

a

a+ 1· 1

aηw + µv+

1

a+ 1· 1

1aηw + µv

(6)

λwv (l) = λw

nv + λc−whv when l ≤Mw

v , and λwv (l) = λw

nv whenMw

v + 1 ≤ l ≤ Nwv . The voice call blocking probability of the

WLAN is then Bwv = πw

v (Nwv ), and the rejection probability for

handoff voice calls from the cell is Dwv =

∑Nwv

i=Mwvπw

v (i).Given that the data call size follows the hyper-exponential

distribution in (4), the data traffic to the WLAN can be viewedas two virtual classes [11] with Poisson arrival rates b

b+1λw

d (j)

and 1b+1

λwd (j), respectively, and exponentially distributed ser-

vice requirements with mean fd/b and b · fd, respectively,where λw

d (j) = λwnd + λc−w

hd when j ≤Mwd , and λw

d (j) = λwnd when

Mwd + 1 ≤ j ≤ Nw

d . With i voice calls and j data calls admittedin the WLAN, the average channel occupancy time of the twovirtual classes is respectively

1

µwd1(i, j)

=a

a+ 1· 1

aηw +χw

d(i,j)

fd/b

+1

a+ 1· 1

1aηw +

χwd

(i,j)

fd/b

1

µwd2(i, j)

=a

a+ 1· 1

aηw +χw

d(i,j)

b·fd

+1

a+ 1· 1

1aηw +

χwd

(i,j)

b·fd

where χwd (·) is the mean service rate for packets from a data

call. Under the assumption that the number of data callsfluctuates in a much smaller time scale than that of voice calls,

the analysis for data traffic can be approximately decomposedfrom that for voice. The steady-state probability of data calls inthe WLAN is then given by πw

d (j) =∑Nw

vi=0 π

wv (i)πw

d (j|i), where

πwd (j|i) = πw

d (0|i)j∏

l=1

[b

b+1λw

d (l)

l · µwd1(i, l)

+

1b+1

λwd (l)

l · µwd2(i, l)

], j = 1, ..., Nw

d .

Similarly, the data call blocking probability of the WLAN isobtained as Bw

d = πwd (Nw

d ), and the rejection probability forhandoff data calls from the cell is Dw

d =∑Nw

di=Mw

dπw

d (i). FromLittle’s law, the average throughput of each admitted data callin the WLAN is given by

Γwd = fd

/ Nwv∑

i=0

πwv (i)

∑Nwd

l=1 l · πwd (l|i)

λwnd · (1 −Bw

d ) + λc−whd · (1 −Dw

d ). (7)

B. QoS Evaluation for Voice and Data Services in the Cell

The location-dependent mobility in a cell results in differentchannel occupancy time for traffic in the cellular-only areaand double-coverage area. This necessitates differentiation oftraffic from different areas while studying the system perfor-mance. To effectively evaluate the QoS metrics while searchingfor the admission parameters, we circumvent the computationcomplexity of solving large-scale balance equations by usingMGFs. Depending on the WLAN state, the average channeloccupancy time of voice calls in the cellular-only area can bederived from (2) as

E[min(Tv , Tcr1)] =

1

µv− 1

µvpc−c ηc/(ηc + µv)

1 − pc−wψ(−µv)� 1

µcv1

(8)

when there is not enough free capacity in the WLAN for avoice call; and it is 1/(µv + ηc) when the incoming voice callcan be admitted to the WLAN. Similarly, for voice calls in thedouble-coverage area, the average channel occupancy time is1

µv− 1

µvψ(−µv) if there is free room for one more voice call

in the WLAN or obtained from (3) as

E[min(Tv , Tcr2)] =

1

µv− 1

µvpc−c ψ(−µv)

1 − pc−wψ(−µv)� 1

µcv2

(9)

otherwise. To simplify analysis, we take an average for themean service rates of voice calls in the cellular-only area anddouble-coverage area, which are respectively given by

µcv1 = Dw

v µcv1 + (1 −Dw

v ) · (µv + ηc)

µcv2 = Dw

v µcv2 + (1 −Dw

v )µv

1 − ψ(−µv). (10)

By modeling the voice traffic in the cell with a multi-serviceloss system, we can obtain the steady-state probability of voicecalls in the cell as

πcv(j) = πc

v(0)

[λv1 + λc−c

hv + λw−chv

µcv1

+λc

nv2

µcv2

]j/j!, j = 1, ..., Nc

v

where λcnv2(= θc

v · λv2) is the mean arrival rate of new voicecalls to the cell from the double-coverage area, λc−c

hv and λw−chv

are mean arrival rates of handoff voice calls from neighboringcells and the overlaying WLAN, respectively. The voice callblocking probability of the cell is then Bc

v = πcv(Nc

v).Similar to the QoS evaluation for data traffic in the WLAN,

both data calls admitted into the cell from the cellular-only

©1-4244-0357-X/06/$20.00 2006 IEEEThis full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE GLOBECOM 2006 proceedings.

area and those from the double-coverage area are respectivelydifferentiated into two virtual classes. Moreover, with therestricted access mechanism, the data call service rates becomedependent on both active voice and data calls in the cell, asall bandwidth unused by current voice traffic is shared equallyby active data calls. Given j voice calls and k data calls in thecell, similar to (10), the service rates of the two virtual classesof data calls in the cellular-only area can be approximated by

µc1d1(j, k) = Dw

d µc1d1(j, k) + (1 −Dw

d )[b · νc

d(j, k) + ηc]

µc1d2(j, k) = Dw

d µc1d2(j, k) + (1 −Dw

d )[νc

d(j, k)/b+ ηc]

(11)

where νcd(j, k) =

Rb,d

fdwith Rb,d being the bit service rate for a

data call and

µc1d1(j, k) =

b · νcd(j, k)

1 − Φ1(−b · νcd(j, k))

, µc1d2(j, k) =

νcd(j, k)/b

1 − Φ1(−νcd(j, k)/b)

.

Similarly, the service rates of the two virtual classes of datacalls admitted to the cell from the double-coverage area canbe obtained as

µc2d1(j, k) = Dw

d µc2d1(j, k) + (1 −Dw

d )b · νc

d(j, k)

1 − ψ(−b · νcd(j, k))

µc2d2(j, k) = Dw

d µc2d2(j, k) + (1 −Dw

d )νc

d(j, k)/b

1 − ψ(−νcd(j, k)/b)

(12)

where

µc2d1(j, k) =

b · νcd(j, k)

1 − Φ2(−b · νcd(j, k))

, µc2d2(j, k) =

νcd(j, k)/b

1 − Φ2(νcd(j, k)/b)

.

Given i voice calls in the cell, the equilibrium distribution ofthe symmetric queue [10] for data traffic is then given by

πcd(j|i) = πc

d(0|i) 1

j!

j∏l=1

[b

b+1λc

d1

µc1d1(i, l)

+

1b+1

λcd1

µc1d2(i, l)

+

bb+1

λcnd2

µc2d1(i, l)

+

1b+1

λcnd2

µc2d2(i, l)

]

i = 0, ..., Ncv , j = 1, ..., Nc

d (13)

where λcd1 = λd1+λc−c

hd +λw−chd , λc

nd2(= θcd·λd2) is the mean arrival

rate of new data calls to the cell from the double-coveragearea, λc−c

hd and λw−chd are the mean arrival rates of handoff

data calls from neighboring cells and the overlaying WLAN,respectively. Thus, the steady-state probability of data calls inthe cell is πc

d(j) =∑Nc

vi=0 π

cv(i)πc

d(j|i). The data call blockingprobability of the cell is Bc

d = πcd(Nc

d). From Little’s law, theaverage data call throughput in the cell is obtained as

Γcd = fd

/ Ncv∑

i=0

πcv(i)

∑Ncd

l=1 l · πcd(l|i)

(λcd1 + λc

nd2) · (1 −Bcd). (14)

IV. NUMERICAL RESULTS AND DISCUSSION

By applying the QoS evaluation given in Section IIIin a search algorithm similar to that of [2], we can ob-tain the best configuration for admission parameters θw

v andθwd to maximize the admissible traffic load with the given

cell/WLAN cluster. For a target traffic load, the correspondingvoice and data admission regions of the cell and WLAN(Nc

v , Ncd , N

wv , G

wv , N

wd , G

wd ) can be determined with the admis-

sion parameters θwv and θw

d . In this section, we analyze thenumerical results to investigate the dependence of resourceutilization on admission parameters and variability of mobilityand traffic. The system parameters are given in Table I.

TABLE I

SYSTEM PARAMETERS

Parameter Value Parameter Valueλv1 0.1 calls/s λv2 0.2 calls/s

(ηc)−1 10 min (ηw)−1 14 min

pc−c 0.76 pc−w 0.24

(µv)−1 140 s fd 64K bytes

Wc 3.84 Mchips/s Qdt 200 Kbps

Qvb 0.01 Qdb 0.01

Fig. 2 shows the dependence of the acceptable data trafficload (λd = λd1 + λd2) on the admission parameter θw

v , whichis the probability that an incoming voice call in the double-coverage area requests admission to the WLAN. The resultsare obtained with Gw

v = Nwv , i.e., no voice handoff from the

cell to the overlaying WLAN. It can be seen that there existoptimal values of θw

v that maximize the acceptable traffic loadand achieve a maximum utilization. Similar phenomenon isobserved in [2]. It is actually a result of properly balancing thevoice and data traffic load in the cell and WLAN. On one hand,when more voice traffic in the double-coverage area is directedto the WLAN (i.e., a larger θw

v ), more cellular bandwidthis available to admit data calls. With a smaller bandwidth,the cell is actually the bottleneck of the whole integratedsystem for data traffic. Hence, with the load balancing ofthe WLAN, the congestion of the cell and in turn the wholesystem is effectively relieved. On the other hand, with alarger θw

v , the maximum number of voice calls allowed inthe WLAN (Nw

v ) should also be larger to meet the voicecall blocking/dropping probability requirement. However, theWLAN is very inefficient in supporting voice traffic. Dueto the small coverage of the WLAN, not only can frequentvertical handoffs between the cell and WLAN degrade thevoice quality and increase the risk of call dropping, but alsoit is detrimental to multiplexing gain by breaking a call intomore service stages. Although the voice traffic load to the cellis reduced to an extent, the number of data calls accommodatedby the WLAN is also significantly reduced. As a result, thetotal acceptable traffic load starts to decrease when θw

v is largerthan a threshold. Hence, θw

v should be large enough to balancethe voice traffic load from the cell and also small enough toavoid the inefficient states of the WLAN in voice support.

It is observed from the curves in Fig. 2 that, with a largerparameter a, the acceptable data traffic load (λd) is larger. Thatis, a higher utilization is achievable when the variability ofuser mobility in the double-coverage area is higher. When a

is 1.0, 4.0, and 6.0, the values of θwv achieving the highest

utilization are 0.44, 0.68, and 0.84, respectively. A largerparameter a indicates more users staying within the WLAN fora shorter time. As shown in Fig. 3, given a fixed Nw

v (i.e., themaximum number of voice calls allowed in the WLAN), whenthe parameter a is larger, a larger fraction of the voice callsin the double-coverage area can be carried by the WLAN andrelieved from the cell. The data call throughput in the cell isthen higher and more traffic is acceptable with QoS guarantee.

The data call variability also affects the admission param-

©1-4244-0357-X/06/$20.00 2006 IEEEThis full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE GLOBECOM 2006 proceedings.

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

2

2.5

3

3.5

4

4.5

5

5.5

6

6.5

7

Fraction of voice traffic load to the WLAN (θvw )

Acc

epta

ble

data

traf

fic lo

ad (

λd

) (c

alls

/s)

Mobility variability parameter a = 1.0Mobility variability parameter a = 4.0Mobility variability parameter a = 6.0

Figure 2. Dependence of maximum acceptable traffic load (λd) on fractionof voice traffic carried by the WLAN (θw

v ) with different mobility variability.

0 5 10 15 20 25 300

0.2

0.4

0.6

0.8

1

Maximum number of voice calls allowed in the WLAN (Nvw )

Fra

ctio

n of

voi

ce tr

affic

load

to th

e W

LAN

(θ vw )

Mobility variability parameter a = 1.0Mobility variability parameter a = 4.0Mobility variability parameter a = 6.0

Figure 3. Dependence of maximum number of voice calls allowed in theWLAN (Nw

v ) on fraction of voice traffic carried by the WLAN (θwd ) with

different mobility variability.

eters and resource utilization. As shown in Fig. 4, more datatraffic is acceptable with a larger value of b, which indicates ahigher variability of the data call size. With a fixed mean value,when the variability of the data call size is higher, more datacalls have a smaller size and less have an extremely larger size.Hence, more data calls have a shorter channel occupancy timeand can be carried by the WLAN with a high throughput. Also,more data calls can finish within the WLAN instead of havingto hand over to the cell when users move out of the WLAN.The traffic load of the small-bandwidth cell is then effectivelyreduced, which results in a higher utilization. From Fig. 2-4,we can conclude that the first-order mean values themselvesare not enough to determine the admission parameters. Withthe MGF-based approach, we can take into account higher-order statistics with reasonable complexity.

In fact, the WLAN-first scheme in [2] is a special caseof the general admission scheme with Gw

v = 0, while theservice-differentiated admission scheme proposed in [3] isequivalent to setting Gw

v = Nwv . Although the specific admission

schemes in [2,3] may be different, the utilization can both bemaximized by properly adjusting the admission parameters ofthe corresponding schemes to reach the best traffic distribution.

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

2.5

3

3.5

4

4.5

5

5.5

6

6.5

Fraction of data traffic load to the WLAN (θdw )

Acc

epta

ble

data

traf

fic lo

ad (

λd

) (c

alls

/s)

Data call variability parameter b = 1.0Data call variability parameter b = 8.0Data call variability parameter b = 16.0

Figure 4. Dependence of maximum acceptable traffic load (λd) on fractionof data traffic carried by the WLAN (θw

d ) with different data call variability.

V. CONCLUSION

In this paper, we study how to improve voice and datasupport in the cellular/WLAN integrated network by applyingadmission control. To take advantage of the complementaryQoS provisioning capabilities of the two networks, the voiceand data traffic should be properly admitted to the cells andWLANs. A generalized admission scheme is proposed toinvestigate the dependence of resource utilization on admissionparameters and traffic distribution. A QoS evaluation approachbased on MGFs is applied to effectively determine the ad-mission parameters. It is observed from numerical results thatthe overall utilization can be maximized when a balance isachieved in distributing the voice and data traffic load to theoverlaying cells and WLANs. The best configuration varieswith the traffic and mobility variability factors.

REFERENCES

[1] A. K. Salkintzis, “Interworking techniques and architectures forWLAN/3G integration toward 4G mobile data networks,” IEEE WirelessCommun. Mag., vol. 11, no. 3, pp. 50–61, June 2004.

[2] W. Song, H. Jiang, and W. Zhuang, “Call admission control for integratedvoice/data services in cellular/WLAN interworking,” in Proc. IEEE ICC,June 2006.

[3] W. Song and W. Zhuang, “QoS provisioning via admission controlin cellular/wireless LAN interworking,” in Proc. 2nd Int’l. Conf. onBroadband Networks (Broadnets), vol. 1, Oct. 2005, pp. 585–592.

[4] M. E. Crovella and A. Bestavros, “Self-similarity in World Wide Webtraffic: evidence and possible causes,” IEEE/ACM Trans. Networking,vol. 5, no. 6, pp. 835–846, Dec. 1997.

[5] S. Thajchayapong and J. Peha, “Mobility patterns in microcellularwireless networks,” in IEEE WCNC, vol. 3, Mar. 2003, pp. 1963–1968.

[6] N. Benameur, S. B. Fredj, F. Delcoigne, S. Oueslati-Boulahia, and J. W.Roberts, “Integrated admission control for streaming and elastic traffic,”in Proc. 2nd Int’l. Workshop on Quality of Future Internet Services, Sep.2001, pp. 69–81.

[7] A. Feldmann and W. Whitt, “Fitting mixtures of exponentials to long-tail distributions to analyze network performance models,” PerformanceEvaluation, vol. 31, no. 3-4, pp. 245–279, Jan. 1998.

[8] J. Perez-Romero, O. Sallent, R. Agusti, and M. A. Diaz-Guerra, RadioResource Management Strategies in UMTS. New York: Wiley, 2005.

[9] M. Naghshineh and A. Acampora, “QoS provisioning in micro-cellularnetworks supporting multiple classes of traffic,” Wireless Networks,vol. 2, no. 3, pp. 195–203, Aug. 1996.

[10] F. P. Kelly, Reversibility and Stochastic Networks. Wiley, 1979.[11] M. Marsan, “Performance analysis of hierarchical cellular networks with

generally distributed call holding times and dwell times,” IEEE Trans.Wireless Commun., vol. 3, no. 1, pp. 248–257, Jan. 2004.

©1-4244-0357-X/06/$20.00 2006 IEEEThis full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE GLOBECOM 2006 proceedings.