472 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 2, NO. 3, MAY 2003
Wireless Access to the World Wide Web in anIntegrated CDMA System
Cristina Comaniciu, Narayan B. Mandayam, Senior Member, IEEE, David Famolari, andPrathima Agrawal, Fellow, IEEE
Abstract—In this paper, we propose a wireless access mecha-nism for web traffic packets in an integrated wireless code-divi-sion multiple-access system that has both voice and web traffic.The proposed scheme is a medium-access control layer/link layer(MAC/LL) scheduling algorithm that consists of a two level con-trol: admission control and packet scheduling. The admission con-trol restricts the number of users in the system such that quality-of-service requirements [target signal-to-interference ratio (SIR) anddelay] for both voice and web traffic can be met. The packet sched-uling balances the system interference on a slot-by-slot basis suchthat the target SIRs can be achieved for all users (voice and webbrowsing sessions) with a higher scheduling priority for voice. De-signing admission control for web users based on the average of-fered rate per session is difficult due to the high variations in theoffered load generated by heavy tailed distributions for web traffic.To overcome this problem, we propose an admission control algo-rithm that adaptively estimates the aggregate average load basedon load measurements using a sliding observation window.
Index Terms—Admission control, code-division multiple access(CDMA), packet scheduling, power control, web browsing sessions.
I. INTRODUCTION
I N RECENT years, there has been a great amount of activityfor developing the next generation wireless networks, which
are expected to provide a wide range of services, such as voice,data, video, and web traffic. Effective medium access control(MAC) protocols have to be implemented to handle conflictingquality-of-service (QoS) requests from different types oftraffic and to achieve efficient system resource utilization.There has been a lot of interest in designing access protocolsfor wireless multimedia services (e.g., [2], [3], [6], [9], [11],[12], and the references therein). However, very few papersconsider non Poisson traffic in wireless systems. For wirelinenetworks, wide-area network (WAN) traffic is modeled as aself-similar process [4]. Self-similarity is obtained as the effectof multiplexing manyON/OFFsources with heavy-tailedON/OFF
period lengths [4]. For the case of a single-cell code-division
Manuscript received April 27, 2001; revised October 9, 2001; accepted Oc-tober 30, 2001. The editor coordinating the review of this paper and approving itfor publication is M. Zorzi. This paper was presented in part at the ICME 2000and Vehicular Technology Conference, Fall 2000. This work was supported inpart by the National Science Foundation (NSF) under Grant NCR 97-06036, inpart by the NSF-KDI program under Grant IIS-98-72995, and in part by Tel-cordia Technologies.
C. Comaniciu and N. B. Mandayam are with the Wireless InformationNetwork Laboratory, Rutgers University, Piscataway, NJ 08854 USA (e-mail:[email protected]; [email protected]).
D. Famolari and P. Agrawal are with Telcordia Technologies, Morristown,NJ 07960-6438 USA (e-mail: [email protected]; [email protected]).
Digital Object Identifier 10.1109/TWC.2003.811051
multiple-access (CDMA) system, the number of web usersmay not be large enough for the assumption of self-similarityto hold. Hence, it is more appropriate to design an accessmechanism by considering models for the web browsingsessions [19], rather than using a self-similarity model for theaggregate traffic. To provide QoS for both voice and web usersin an integrated CDMA system, we propose a two level accesscontrol: at a call arrival time scale (admission control) and at atime slot scale (packet scheduling).
Designing admission control for web users based on theaverage offered rate per session is difficult due to the highvariations in the offered load generated by heavy taileddistributions for web traffic. To overcome this problem, theproposed admission control adaptively estimates the averageweb load aggregated from all World Wide Web (WWW)sessions currently admitted in the system. The aggregateaverage load estimation is based on load measurements usinga sliding observation window. Our work is motivated by [7]which discusses admission control for ATM networks. Gibbenset al.prove that significant statistical multiplexing gains can beachieved if the admission criterion is based on instantaneousload measurements and the admission threshold is derivedusing a decision theoretic approach.
At a slot time scale, resources are shared among web usersaccording to the residual capacity left over after subtracting thevoice contribution. Two different schemes for predicting the fu-ture voice load are proposed and compared. The “one-step pre-diction” method is similar with the prediction technique pro-posed in [6]. In [6], it is shown that a simple predictive schemeas the one used in this paper outperforms other low complexitynonpredictive methods such as the one discussed in [2]. How-ever, the analysis presented in [2] and [6] is not suitable forWWW admission control. The “mean value prediction” schemerepresents the case with no explicit access control and it is ana-lyzed for comparison purposes.
We design admission and packet scheduling based on the“noise rise” [1] condition, which was proved in [10] to satisfythe power control feasibility condition as well. In this paper, wefocus on the single-cell case with perfect power control. Undermore realistic assumptions (imperfect power control and mul-tiple cells), the admission control imposes supplementary con-straints [3], [17], [18]. We also show how the analysis can beextended for the imperfect power control and the multiple cellscenarios as well. We show that, although the analysis and thesystem capacity changes, the proposed admission and packetscheduling mechanisms remain the same. As a final observa-tion, we note that the analysis is general enough to include a
COMANICIU et al.: WIRELESS ACCESS TO THE WORLD WIDE WEB IN AN INTEGRATED CDMA SYSTEM 473
Fig. 1. Web browsing session.
variety of elastic data traffic models. In a related paper [8], wehave analyzed a more general case, where voice, data, and videotraffic are present in the CDMA system.
The paper is organized as follows. In Section II, we presentthe system and data models (voice and web traffic) used in thiswork. Section III describes the proposed access control withemphasis on the admission control in Section III-A and packetscheduling in Section III-B. Section III-C extends the analysisfor the multiple cell case, Section III-D presents simulation re-sults, and Section IV presents the conclusions.
II. SYSTEM AND DATA MODELS
We consider the scenario in which a set of power controlledmobile terminals are transmitting packets to a base stationin a single-cell, direct sequence CDMA system. Admissionand packet scheduling algorithms are proposed for the uplink,which is considered to be slotted, with a slot duration equal toa frame duration. Two main categories of traffic are discussed:web browsing sessions and voice. Each web session can varyits transmission rate by increasing or decreasing the spreadinggain. A web browsing session (see Fig. 1) [19] is comprisedof a number of packet calls (geometrically distributed)separated by reading times (geometrically distributed interar-rival time between packet calls). Each packet call consists of anumber of IP packets ( ), which is Pareto distributed, and hasa Geometrically distributed interarrival time between packetarrivals ( ). The length of a web traffic packet is fixed. Webbrowsing sessions arrive according to a Poisson distributionwith arrival rate . The above description is proposed in [19]for modeling large file uploads and also the reverse trafficcorresponding to WWW downloads. In our simulations, weassume that the new session requests also include sessionshanded off from other cells (arriving with a Poisson distributionas in [13]). No admission priority is assigned to handed offweb session connection requests. The dwell time of a WWWsession in the cell is modeled by an exponential distributionwith mean [13].
The QoS requirements for both voice and web users are bit-error rate (BER) and the access delay for packets. The BERtarget is mapped into a signal-to-interference ratio (SIR) target.The performance measure of interest is the SIR outage proba-
bility: the probability that the target SIR cannot be achieved. Interms of delay, voice users are delay intolerant and are givenpriority over web calls which are delay insensitive. Only an av-erage delay performance is required for the web traffic. Usingan analogy with a queuing system, the average delay experi-enced by web packets is directly determined by the aggregateweb load in the system. As a consequence, we consider as QoSrequirements for web browsing sessions the target SIR and theaggregate WWW traffic load.
As in [2] and [6], the voice activity is modeled as a continuousON/OFFMarkov process, with a transition ratefrom ON to OFF
and from OFF to ON states. The voice and data transmissionoccurs in slots of duration . is equal to a frame duration.Therefore, the cumulative voice activity process can be approx-imated as a discrete Markov chain. The transition probabilitiesfor voice users in the system and active voice users inthe th time slot are given by
(1)
where , .The stationary probability of state, in which users are
active, is given by
(2)
III. A CCESSCONTROL FORQOS GUARANTEES
The role of the access control is to balance the system uti-lization such that QoS requirements are met for all users. Weassume that SIR requirements for all calls in the system can bemet if the total system load is maintained below some threshold[1], [10]. Denoting the system bandwidth as, the “noise rise”condition can be expressed as
SIR SIR (3)
where and , represent the number of voice calls andWWW calls, and are the transmission rates for thethvoice and WWW call, and are the target SIRs forvoice and WWW, and is the system load. The noise rise co-efficient represents the ratio of the background noise powerspectral density to the total received interference power den-sity. Note that we do not distinguish between handed off websessions and new web session initiations in the cell, while ac-counting for the number of WWW calls .
Equation (3) should hold at all time scales. A two-level accesscontrol, implemented via admission control and packet sched-uling is proposed. The packet scheduling is based on (3) ex-pressed at the time scale of frames. It assigns resources for theweb traffic according to the residual capacity available afterthe voice traffic contribution has been subtracted. An imperfectscheduling resulting in higher system load constitutes an SIR
474 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 2, NO. 3, MAY 2003
outage since SIR targets cannot be met for the slots in which the“noise rise” condition does not hold. Thus, the SIR outage prob-ability characterizes the performance of the packet schedulingscheme.
The admission control guarantees the feasibility of the QoSspecifications for all users (SIRs and delay). It is based on the“noise rise” condition in (3) specified at the call arrival timescale. The admission control for web users is characterized by anew performance measure (the load outage probability), whichrepresents the probability that, after admission, the obtained ag-gregate WWW traffic load in the cell is higher than a prespeci-fied target value.
A. Admission Control
The admission control reserves resources for the aggregateweb traffic, based on voice traffic descriptors (average and peaktransmission rates) and on QoS requirements for the web ses-sions: average throughput and minimum throughput. In orderto reserve a minimum throughput for the WWW users ( ),while guaranteeing zero delay for the voice traffic for the worstcase in which all voice users transmit at peak rates ( ), themaximum load condition should hold
SIR SIR
(4)An average rate reservation for theth WWW sessioncan be guaranteed if the average load condition holds
SIR SIR (5)
where represents the average offered voice rate for user.
As in [8], a new call can be admitted in the system if (4) and(5) simultaneously hold.
By analyzing (5), it can be seen that it requires thea prioriknowledge of for each WWW session. In order to determine
, its relationship with the offered throughput per web sessionhas to be first established. Assuming that all web sessions havethe same SIR target, we can define the aggregate rate reservationavailable for WWW users ( ) as
SIR
SIR(6)
Using an analogy with a queuing system, the web traffic will beserviced by a queue having a rate of service, and input of-fered traffic , with being the offered rate for thethWWW session. The average delay experienced by the WWWpackets is directly influenced by the aggregate WWW load inthe system ( ). To avoid very large delays, a target load ()for the queue has to be imposed
(7)
Note that (7) is based on (5) and represents the admission con-dition for the web users. If a minimum throughput requirementis specified by the web sessions, (7) will be used in conjunctionwith (4) for web traffic admission. Without loss of generality,
throughout the remainder of the paper.We notice that (7) requires thea priori knowledge of of-
fered throughput per session. However, in a WWW session, thenumber of offered packets per packet call is Pareto distributed;therefore, is a random variable with a very large variance.As a consequence, the sample mean estimate obtained for thelife of a session will be far away from the mean of the randomvariable . Thus, we cannot rely on the statistical mean to de-scribe the offered rate per WWW session: The expected valuefor the average offered rate per WWW session will greatly overor under reserve resources for that particular session. Instead ofstatistically describing the aggregate WWW offered load, weestimate by measuring at each time slot the average load forthe last slots. If the difference between the target load andthe current measured average load [ ] for the th timeslot is greater than some threshold, new sessions can be ad-mitted. This threshold is chosen such that, with high probability,the load for the next slot is less that the target load plus an ac-ceptable performance tolerance parameter.
We summarize the admission control algorithm as follows.At the end of each time slot do the fololowing.
• Measure the new total offered load averaged over thelast slots. is the length of a moving averagewindow over which samples are collected for averageload measurement for slot. The measured average loadfor slot [ ] is given by
(8)
where and is the numberof IP packets arriving in slot . is computed atthe base station by adding up all arrival rate measurementsreceived from all active WWW terminals.
• Admit all new connection requests in the next time slot, if
(9)
Otherwise, reject all new connections for the next timeslot.
The admission control performance is characterized by the loadoutage probability, i.e., the probability that, after admission, theobtained aggregate WWW load in the cell is higher than a pre-specified value . The load for the next time slotcan be computed as the load previously measured in slotplusthe new offered load [denoted as ] from all sourcesstarting to transmit in slot :
(10)
To simplify the computation in (10), we make the conservativeassumption that no active packet call in slotwill end in slot
. Defining the admission of new sessions in theth time
COMANICIU et al.: WIRELESS ACCESS TO THE WORLD WIDE WEB IN AN INTEGRATED CDMA SYSTEM 475
slot as the event and , the loadoutage probability can be expressed as
(11)
The admission threshold is determined such that the loadoutage probability for any slot is maintained below a pre-scribed value
(12)
We now derive an expression for the load outage probability in(11) as follows. can be expressed as the product ofthe number of new packet calls offered in slot [denotedas ], and the average rate offered per packet call( ), divided by the reserved rate
(13)
The rate offered per packet call is a random variable specifiedas
(14)
where is the number of IP packets in a packet call,is thelength of an IP packet (measured in number of basic packetlengths), and represents the interarrival time between IPpackets.
The number of new packet calls offered in slot com-prises of new session arrivals and of new packet call initiationsfrom previously admitted sessions that are currently inactive.The number of new sessions in slot [denoted as ]is Poisson distributed with rate sessions per unit slot ,and the number of new packet call initiations in the next timeslot is a binomial distributed random variable . Fora given number of inactive web calls in the system,is characterized by the probability mass function
(15)
where represents the probability of a new packet call initia-tion in the next time slot.
Therefore, is a random variable expressed as
(16)
Equation (13) can now be rewritten as
(17)
is Pareto distributed [19] with the characteristic functiongiven by
(18)
Fig. 2. Dependence of load outage probability on the rate of new sessionarrivals (� ) (sessions/second) and on the number of inactive sessions in thecell (N ).
The load outage probability, conditioned on (the number ofinactive web sessions in the CDMA system),, , and canbe expressed as
if
if
otherwise
(19)
where is defined as
(20)
The derivation of (19) is presented in the Appendix.By averaging over the probability mass functions for the
number of new session arrivals ( , with parameter ),the interarrival time between packets (, with parameter )and the number of new packet call initiations from inactivesessions ( with parameters and ), we can express theload outage probability conditioned on the number of inactiveweb sessions in the system:
(21)
where is the discrete unit of time for the packet interar-rival. depends on the particular value that
may take, and on the number of inactive sessionscurrently in the system ( ). For notation simplicity, we denote
.Analyzing (19), we notice that, due to a very small interarrival
time between packets, variations in the factor have anegligible impact on the load outage probability value. Thus,the dependence of on , for any time slot , isnegligible. In Fig. 2, is plotted as a function of both thenumber of inactive sessions currently admitted in the cell ()
476 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 2, NO. 3, MAY 2003
Fig. 3. Dependence of the admission threshold on the rate of new sessionarrivals (� ) (sessions/second) and on the number of inactive sessions in thecell (N ).
and the arrival rate of new sessions (). For a fixed constraintfor ( ), the admission thresholdis plotted in Fig. 3. It can be seen that is either zero (forthe feasible region) or one (infeasible region). For the first case,in an arbitrary slot , all calls can be admitted, regardless ofhow close the current measurement is to the target load, pro-vided that , whereas for the second case no call canbe admitted for the given pair (, ). The load outage prob-ability is predominantly affected by the arrival rate of new ses-sions. To limit the arrival rate of new sessions to a desired target
, we need an auxiliary admis-sion mechanism. The base station continuously estimatesandbroadcasts an admission probability
(22)
Each arriving new session generates a Bernoulli random vari-able with this probability. If a “1” is generated, the session sendsa new call request to the base station; otherwise, it refrains fromrequesting a new connection and it is blocked. In our discus-sion, we assume that blocked arrivals are not allowed to retryto reconnect, although the case of infinite number of retries aswell as finite number of retries can easily be incorporated in theanalysis. After the preliminary arrival rate control, the numberof new connection requests is a Poisson random variable witharrival rate .
The admission scheme proposed in this section controls onlyconnection requests from new session arrivals. The inactive callsalready admitted in the system are automatically granted accessat the time they become active. Future work may consider atighter WWW load control for which new packet call initiationrequests (from previously inactive sessions) are rejected when-ever .
We also note that admission priorities for the handed off callsmay be implemented by setting different admission thresholdsfor the new calls initiated in the current cell and the handed offcalls. This also represents future work and it is not addressed inthis paper.
If imperfections in the power control loops are taken into ac-count, the previous analysis has to be modified to derive a newexpression for . In this case, the received SIR for both voiceand WWW calls is a random variable. Applying the central limittheorem, the average load and the maximum load can be approx-imated to be Gaussian random variables, and therefore, (4) and(5) should hold in a probabilistic sense. Thus, we derive a newexpression for by imposing the condition
var(23)
where represents the outage probability target.Let us denote and var . We assume that
the reserved rate for WWW sessions is equally shared amongthe active web users; therefore, ( is the ratereservation for a particular web session). Denotingas thevoice transmission rate, the mean and variance for the averageload can be expressed as
SIR SIR
SIR var SIR(24)
Denoting SIR var SIR SIR var SIR ,and SIR
var SIR SIR SIR ,(24) and (23) yield the quadratic inequality
SIRvar SIR
(25)
The maximum value for, such that , is found to be
SIR var SIRSIR
(26)Since and given defined as in (26), from thefirst equation in (24), we derive the new reservation rate for theaggregate web traffic in the cell
SIRSIR
(27)
B. Packet Scheduling
The packet scheduling multiplexes the voice and web trafficsuch that the system resources are used as efficiently as pos-sible and all delay requirements are met. Voice sessions are al-ways given the highest priority and they are allowed to transmitwithout delay. The web packets are scheduled to transmit ac-cording to the available capacity (residual capacity) obtained bysubtracting the voice traffic contribution from the total systemcapacity. The WWW transmission rate is adaptively modifiedfrom one slot to another, based on resource availability. The re-source scheduling is done such that the “noise rise” [1] condi-tion holds at the slot (frame duration) time scale for each slot.Resource availability for WWW is determined by estimating thevoice load contribution for the next time slot. For the voice calls,the fluctuations in the load contribution are the result of changes
COMANICIU et al.: WIRELESS ACCESS TO THE WORLD WIDE WEB IN AN INTEGRATED CDMA SYSTEM 477
in voice activity. We propose and compare two schemes for pre-dicting the voice load for the next slot.
1) One-step prediction. The total estimated load forthe voice users at slot is determined based on the loadmeasured in the previous time slot . A simpleone-step estimator is used: .
2) Mean value prediction. This method assumes that thenumber of active voice users in each time slot is exactlyequal to the expected value for the number of active usersand computes the predicted voice load accordingly.
Using the above described prediction methods, the residual ca-pacity available for the WWW users is determined. Based onits value, the base station assigns transmission rates for the nexttime slot to each WWW user by traversing a circular list whichcontains all active WWW users’ ids, together with their last raterequest. The list is updated whenever a new rate request from aWWW user is received. The rate requested by a WWW user (de-noted as ) is computed every time slot by the requesting ter-minal according to the current number of basic WWW packetswaiting for transmission. A basic WWW packet is equivalent toa voice packet transmitted with rate . Denoting thenumber of basic WWW packets as , the new requested ratecan be expressed as
(28)
where is the limit imposed on the maximum rate that canbe requested.
A NULL rate request means that the WWW call became in-active, and as a consequence, the base station will remove theuser from the circular list used for scheduling.
If perfect power control is assumed, the residual load avail-able for WWW transmission in slot is computed as a functionof the estimated load for the voice calls ( ) as
(29)
The voice load estimate is obtained for the “one-step prediction”method as
SIR (30)
and for the “mean value prediction” method as
SIR (31)
is the number of active voice users in the previousslot, and is the voice activity coefficient.
The total aggregate rate that the WWW users can transmit inthe next time slot , given the fact that all WWW usershave the same BER requirement, is given by
SIR(32)
can be also determined for the imperfect power controlcase by using a similar approach to that described in the previoussection. For this case, the instantaneous load condition shouldhold in a probabilistic sense, and the residual capacity availablefor data transmission at each time slot can be derived in a similarmanner as the derivation for in Section III-A.
The choice of the maximum burst rate assignment ( )parameter affects the system performance. For practical rea-
sons, cannot be driven to infinity. Further, for a cellularsystem, high rate burst transmission may cause excessive in-terference to the neighboring cells, and the scheduling proce-dure has to account for this effect [3]. If , thesystem is equivalent to a queuing system for which the servicerate ( ) is obtained as a long-term average of the total aggre-gate rate available for WWW traffic in a particular slot, i.e.,
. Under the assumptionthat the data queue service process is ergodic,can also beexpressed as , where representsthe mean with respect to , which is the number of active voiceusers.
If is too small, the system resources will be usedinefficiently. This happens because the total rate request for theWWW traffic in slot is limited to multiplied by thenumber of active WWW users ( ). Even though moreWWW packets are waiting in the queues and more resourcesmight be available, the web transmission rate in slotis atmost equal to .
Let be the number of WWW packets in user’s queue. If
and , thesystem resources are wasted for the time slot. By analogywith a queuing system, this has the effect of reducing the ser-vice rate of the queue ( ) to an effective service rate
.We wish to study the efficiency in allocating the system re-
sources for both “one-step prediction” and “mean value predic-tion” schemes. Specifically, we measure the efficiency as
(33)
We parameterize as a function of , whereis the transmission rate of an active voice user (basic rate),
. For the perfect power control case, we determinefor a given number of voice terminals ( ) and a given numberof active WWW terminals ( ) in the system, considering thatall voice terminals have SIR target requirement SIR, and allWWW sessions require an SIR target SIR SIR
i) one-step prediction
SIR(34)
ii) mean value prediction
SIR(35)
In Fig. 4, the efficiency in allocating the system resources ()is shown for the two proposed packet scheduling mechanisms,for a range of values for. The system’s parameters were chosenas: MHz, , voice calls, ,SIR dB, SIR dB, and and activeWWW calls. It can be seen that a smallreduces the efficiency
. Since only an average throughput requirement is specifiedfor the WWW users, the system should operate with the smallest
that gives . As we can see by comparing Fig. 4(a) and(b), this optimal value depends on the number of active WWWusers in the system. Therefore, if fewer WWW users are active,we will use the system resources more efficiently by allowinghigher WWW transmission rates.
478 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 2, NO. 3, MAY 2003
(a)
(b)
Fig. 4. Efficiency in allocating the system resources as a function of theR
constraint.
The packet scheduling performance is characterized by twoperformance measures: SIR outage probability and WWWtraffic throughput. If perfect power control is assumed, theworst case SIR outage probability can be determined for ,when WWW queues are full, and there are always WWWpackets to be transmitted. For a system having both voice andWWW users, the SIR outage is given by the errors in predictingthe voice load for the next time slot. Thus, considering the voicemodel described by (1) and (2) for the “one-step prediction,”we have
Prob
(36)
Fig. 5. Worst case SIR outage probability for perfect power control case.
For the “mean value prediction” method, an outage occurswhenever the number of active voice users is greater than themean number
(37)
Fig. 5 illustrates worst case SIR outage plots for the two predic-tion methods, as a function of the number of voice users in thesystem.
C. Extension to Multiple Cells
If the multiple cell case is considered, the impact of usersin neighboring cells have to be accounted for. As in [15], theaverage interference caused by an out-of-cell interferer, aggre-gated from all cells is equivalent to the interference caused byin-cell users. For a path loss exponent , . How-ever, the interference caused by the data users varies accordingto their location and their particular burstiness pattern. Thesevariations are averaged out for the admission control condi-tion, since this condition characterizes the system performanceon the long-term run. Hence, the interference created byout-of-cell sessions is equivalent, on average, with the inter-ference created by in cell users. However, for the packetscheduling, the interference variability has to be accounted for.A high WWW rate user in a neighboring cell is equivalent to
WWW users in the host cell, whereis the distance fromthe mobile user to its host cell base station [3]. As in [3],is expressed as the ratio of the path loss of the mobile to the hostcell to the path loss to the neighboring cell. Note that the packetscheduling also relies on instantaneous rate measurements fordata users in the neighboring cells.
We revisit the admission control conditions to account for theout-of-cell interference. As in the previous section, we assumethat no minimum throughput requirement is specified by theWWW users, i.e., . Therefore, for equal to thenumber of out-of-cell voice calls, the maximum load condition
COMANICIU et al.: WIRELESS ACCESS TO THE WORLD WIDE WEB IN AN INTEGRATED CDMA SYSTEM 479
TABLE INUMERICAL VALUES FORSIMULATIONS
limits the number of voice connections that can be supported bythe CDMA system
SIR SIR
(38)The average load condition has to account for the averageout-of-cell interference created by the voice users as well asfor the average aggregate out-of-cell interference generatedby the web users. A reasonable assumption is that all voiceconnections require the same target SIR, , and all webconnections same target SIR, . Assuming that all the cellsguarantee the same service rate for the web traffic, andthat the admission control scheme maintains the traffic loadclose to the target load , the average aggregate interferencecreated by the out-of cell WWW users can be expressed as
. is the numberof out-of-cell WWW sessions, and represents the numberof neighboring cells considered in the analysis. Hence, theaverage load condition can be derived to be
SIR SIR
(39)
From (39), we can solve for the maximum reservation rate percell available for WWW users
SIRSIR
(40)
The admission control algorithm is identical with the single-cellcase, with defined as in (40).
Since the interference caused by data users depends on theusers’ positions in the cell and varies according to the burstiness
pattern, the packet scheduling is based on neighbor coordinatedaccess control as in [3]. The residual load available for WWWtransmission in slot is estimated as
SIR
SIR (41)
where was defined in the previous section,representsthe voice activity factor, is the voice transmission rate,is the transmission rate of theth out-of-cell WWW interferer,
represents the number of active out-of-cell web sessions,and is the ratio of the path loss of theth out-of-cell WWWuser to its host cell to the path loss to the analyzed cell.,
, and are determined based on the information re-ceived from the neighboring base stations.
D. Simulation Results and Observations
The system performance is experimentally determined byusing a system simulator implemented in OPNET [20]. Theparameters of interest are average WWW load, access delay forWWW traffic packets, and percentage of SIR outage, measuredfor sample paths of duration min. For our simulations,we abstract the physical channel model as follows. We simulatethe system under the assumption of perfect power control, forwhich each user’s received SIR is deterministic and equal toits target SIR. Imperfections in the power control loops can behandled by modeling the received SIR as a random variable [see(23)–(27) in Section III-A]. A reasonable assumption is to con-sider the received SIR to be a lognormal random variable [1].This assumption is motivated by the fact that, in practice, SIRestimates are obtained using averages over a couple of frames
480 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 2, NO. 3, MAY 2003
[16]. Thus, the effect of Rayleigh fading is averaged out, andonly shadow fading has to be considered. While this impactsthe analysis, it does not alter the admission control and packetscheduling mechanisms proposed here. Therefore, to illustratethe performance of the proposed web admission control/packetscheduling scheme, the simplified model based on the perfectpower control assumption suffices. The numerical values usedfor the system parameters in the simulations are mostly basedon the cdma2000 specifications [19] and are given in Table I.
The admission control restricts the number of new connectionrequests allowed in the system. New session requests are com-prised of two types of arrivals: sessions handed off from othercells (arriving with a Poisson distribution as in [13]) and newsessions initiations in the current cell (Poisson distributed). Thetotal number of new session requests is, therefore, Poisson dis-tributed with a cumulated arrival rate session/s. TheWWW sessions that become inactive are allowed to maintain thepoint-to-point protocol (ppp) connection and are automaticallygranted access at the time they become active. The time neces-sary to establish a new connection for both inactive calls and newadmitted sessions is neglected in simulation. The residence time(dwell time) of a WWW session in the cell is assumed to be mod-eled by an exponential distribution with mean. As in [13],for our simulations, we consider min. In order to showthat the mobility model has no impact on the proposed accesscontrol performance, another value minutes is consid-
ered as well. Both proposed packet scheduling methods are im-plemented. Figs. 6–10 illustrate the simulation results as samplepaths for (from top to bottom) the number of WWW sessionscurrently admitted in the cell, the number of currently activeWWW sessions, the measured aggregate data offered load, andthe number of WWW packets arriving in a slot.
Fig. 6 represents the case in which the admission control isused in conjunction with the “one-step prediction” method, andFig. 7 illustrates the case in which the “mean value prediction”method is used as a packet scheduling mechanism. For bothcases, min and the length of the moving average windowis slots 10 min. Comparing the results ob-tained for these two cases in terms of percentage of outage andaverage aggregate offered load (Table II), it can be seen that forboth cases, the measured average load for a simulation time of40 min is close to the specified target . Also, althoughboth cases give a smaller outage percentage than the worst caseoutage probability derived in Section III-B (because of the smallaverage WWW load), the “one-step prediction” performs better,resulting in a better control of the SIR outage probability, com-pared with the “mean value prediction” method.
Figs. 8 and 9 illustrate the performance of the admissioncontrol for different lengths of the moving average window,
slots 1 min, and slots20 min, respectively. Both cases implement the “one-stepprediction” method and use min. We notice that
COMANICIU et al.: WIRELESS ACCESS TO THE WORLD WIDE WEB IN AN INTEGRATED CDMA SYSTEM 481
a smaller length for the averaging window leads to morevariability in the measured load, and to a less tight admissioncontrol, resulting in a higher average offered load measured forthe duration of the simulation (see Table II). Also, comparingthe average offered load obtained at the end of the simulationfor the three cases illustrated in Figs. 6, 8, and 9, we noticethat the best results are given for the largest window size. Thetradeoff is that the largest window size requires more memoryto store and process the samples.
Finally, comparing Figs. 6 and 10, for the same simulationconditions but for different values for , 3, and 10 min, respec-tively, we see that the system performance is similar, but for thecase illustrated in Fig. 10, more inactive calls are maintainingppp connections in the simulated cell.
Analyzing the web packets’ access delays (scheduling de-lays) (see Table II) obtained from simulations, it can be seenthat the delay is influenced by both the average load value ob-tained for the duration of the simulation and by the particularaverage load trace obtained. Although in general, it may seemthat a lower average load value gives a lower access delay, insome cases, higher instantaneous measured loads for longer pe-riods of time lead to higher access delays. This happens be-cause a larger percentage of WWW packets experience higherdelays due to increased congestion (see Fig. 11). By comparingFigs. 6 and 8, we observe that although higher load variationsin Fig. 9 lead to a higher average load value measured at the
end of the simulation, congestion occurs for shorter periods oftime yielding lower access delays for the web packets. In con-clusion, the WWW access delay performance will be highly de-pendent on the particular realization (trace) of the simulation,but a tight control on the average aggregate load will keep thedelay close to some specified target, which can be experimen-tally determined.
We mention that the access delay represents only delay in-troduced by the packet scheduling scheme. We have not con-sidered the interaction of the proposed admission/packet sched-uling scheme with transmission control protocol (TCP). Any ac-cess scheme will introduce a variability in the round-trip time(RTT) which may cause session time-outs and, therefore, de-crease the overall throughput due to unnecessary retransmis-sions [14]. Thus, the end-to-end delay experienced by WWWpackets is greatly influenced by the interaction of the accessscheme with TCP, which is beyond the scope of this work.
482 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 2, NO. 3, MAY 2003
Fig. 11. Percentage of web traffic packets experiencing a scheduling delay lessthan the value on the abscissa.
IV. CONCLUSION
In this paper, we proposed a wireless access mechanism forweb traffic packets in an integrated wireless CDMA system thathas both voice and web traffic. The access scheme is based ona two level control: admission control and packet scheduling.
At the call arrival time scale, the admission control guaranteesthe QoS requirements in terms of bit error rates and delays forall users. At a slot time scale, a packet-scheduling mechanismdelivers QoS by scheduling the available resources among voiceand WWW users.
The difficulty in designing admission control for WWWtraffic comes from the fact that an accurate estimate for theaverage offered rate per web session is hard to determine. Theweb traffic model comprises heavy tailed distributions whichresult in high variations in the offered load per session. Thus,the expected value for the average offered rate per WWWsession will greatly over or under reserve resources for thatparticular web session. As a solution to this problem, we pro-pose an admission control algorithm that adaptively estimatesthe aggregate average offered load for the web traffic, based onload measurements using a sliding observation window. Theperformance of the admission control is measured by the loadoutage probability, which is the probability that after admission,the average load is maintained below a prescribed value.
Two different packet scheduling mechanisms are proposedand compared. The performance of the packet schedulingscheme is characterized by the SIR outage probability, whichgives the percentage of time slots in which the SIR targetscould not be met for all users (voice and WWW) in the system.
The simulation results showed that the proposed admissioncontrol maintains the average aggregate load close to a desiredtarget, therefore guaranteeing bounded average delay for theweb packets. At the expense of increased implementation com-plexity, a tighter control of the average aggregate WWW loadcan be obtained if a larger moving average window is used. Interms of SIR outage probability, both analysis and simulationssuggest that the SIR outage probability is better controlled bythe “one-step prediction” based packet scheduling mechanism.
APPENDIX
Derivation of the conditioned load outage probability for-mula, when the conditioning is with respect to (the numberof inactive web sessions in the CDMA system),, and :
We start our derivation using the expressions for the loadoutage probability (11) and for the new offered load (17). Withrespect to the random variables, , , and , the outagecondition becomes
(42)
Rearranging (42) and denoting , theoutage condition is given as
(43)
We distinguish two cases.
1) . Since ,(43) is valid for all . Thus
if
(44)
COMANICIU et al.: WIRELESS ACCESS TO THE WORLD WIDE WEB IN AN INTEGRATED CDMA SYSTEM 483
2) (43) can be further ex-pressed as
(45)
Since is Pareto distributed, if, . Otherwise,
using (18), we compute the conditioned load outage probabilityas
(46)
REFERENCES
[1] A. M. Viterbi and A. J. Viterbi, “Erlang capacity of a power controlledcellular CDMA system,”IEEE J. Select. Areas Commun., vol. 11, pp.892–900, Aug. 1993.
[2] A. Sampath and J. M. Holtzman, “Access control of data in integratedvoice/data CDMA systems: Benefits and tradeoffs,”IEEE J. Select.Areas Commun., vol. 15, pp. 1511–1526, Oct. 1997.
[3] S. Kumar and S. Nanda, “High data-rate packet communications for cel-lular networks using CDMA: Algorithms and performance,”IEEE J. Se-lect. Areas Commun., vol. 17, pp. 472–492, Mar. 1999.
[4] M. E. Crovela and A. Bestavros, “Self-similarity in world wide webtraffic: Evidence and possible causes,”IEEE/ACM Trans. Networking,vol. 5, pp. 835–846, Dec. 1997.
[5] A. Erramilli, O. Narayan, and W. Wilinger, “Experimental queuing anal-ysis with long-range dependent packet traffic,”IEEE/ACM Trans. Net-working, vol. 4, pp. 209–223, Apr. 1996.
[6] C. Comaniciu and N. B. Mandayam, “Delta modulation based predictionfor access control in integrated voice/data CDMA systems,”IEEE J.Select. Areas Commun., vol. 18, pp. 112–122, Jan. 2000.
[7] R. J. Gibbens, F. B. Kelly, and P. B. Key, “A decision-theoretic ap-proach to call admission control in ATM networks,”IEEE J. Select.Areas Commun., vol. 13, pp. 1101–1113, Aug. 1995.
[8] C. Comaniciu, N. B. Mandayam, D. Famolari, and P. Agrawal, “QoSguarantees for third generation (3G) CDMA systems via admission andflow control,” in Proc. IEEE Vehicular Technology Conf. (VTC), vol. 1,Boston, Sept. 2000, pp. 249–256.
[9] N. Dimitriou, R. Tafazolli, and G. Sfikas, “Quality of service for multi-media CDMA,” IEEE Commun. Mag., vol. 38, pp. 88–94, July 2000.
[10] A. Sampath, N. B. Mandayam, and J. M. Holtzman, “Erlang capacity ofa power controlled integrated voice and data CDMA system,” inIEEEVehicular Technology Conf. (VTC), vol. 3, 1997, pp. 1557–1561.
[11] Y. Bao and A. Sethi, “Performance-driven adaptive admission controlfor multimedia applications,” inProc. IEEE Int. Conf. Communications(ICC), Vancouver, BC, Canada, June 1999, pp. 199–203.
[12] S. Choi and K. G. Shin, “An uplink CDMA system architecture withdiverse QoS guarantees for heterogeneous traffic,”IEEE/ACM Trans.Networking, vol. 7, pp. 616–628, Oct. 1999.
[13] L. Ortigoza-Guerrero and A. H. Aghvami, “A prioritized handoff dy-namic channel allocation strategy for PCS,”IEEE Trans. Veh. Technol.,vol. 48, pp. 1203–1215, July 1999.
[14] S. Ramakrishna and J. M. Holtzman, “Interaction of TCP and data accesscontrol in an integrated voice/data CDMA system,”Mobile NetworksAppl. (MONET), vol. 3, no. 4, pp. 409–417, 1998.
[15] K. S. Gilhousen, I. M. Jacobs, R. Padovani, A. J. Viterbi, L. A. Weaver,and C. E. Wheatley, III, “On the capacity of a cellular CDMA system,”IEEE Trans. Veh. Technol., vol. 40, pp. 303–312, May 1991.
[16] D. Ramakrishna, N. B. Mandayam, and R. D. Yates, “Subspace based es-timation of the signal-to-interference ratio for CDMA cellular systems,”IEEE Trans. Veh. Technol., vol. 49, pp. 1732–1742, Sept. 2000.
[17] M. Andersin, Z. Rosberg, and J. Zander, “Soft and safe admission controlin power-controlled mobile systems,”IEEE/ACM Trans. Networking,vol. 5, pp. 255–265, Apr. 1997.
[18] D. Kim, “Efficient interactive admission control in power-controlledmobile systems,”IEEE Trans. Veh. Technol., vol. 49, pp. 1017–1028,May 2000.
[19] The cdma2000 ITU-R Candidate Submission, Apr. 1998.[20] OPNET,Modeler Manuals—Modeling, vol. 1 and 2, version 6.
Cristina Comaniciu received the M.S. degreein electronics from the Polytechnic University ofBucharest, Bucharest, Romania, in 1993, and thePh.D. degree in electrical and computer engineeringfrom Rutgers University, Piscataway, NJ, in De-cember 2001.
From 1998 to 2001, she was with the Wireless In-formation Network Laboratory (WINLAB), RutgersUniversity, working on integrated access controland detection algorithms for multimedia CDMAsystems. She is currently a Postdoctoral Fellow in
the Electrical Engineering Department, Princeton University, Princeton, NJ.Her research interests include radio resource management, admission controlfor multimedia wireless systems, multiuser detection, and modeling andperformance analysis for wireless systems.
Narayan B. Mandayam (S’90–M’95–SM’00)received the B.Tech. (honors) degree in 1989 fromthe Indian Institute of Technology, Kharagpur, India,and the M.S. and Ph.D. degrees from Rice Univer-sity, Houston, TX, in 1991 and 1994, respectively,all in electrical engineering.
Since 1994, he has been at the Wireless InformationNetwork Laboratory (WINLAB), Rutgers University,Piscataway, NJ, where he is currently an AssociateProfessor in the Department of Electrical andComputer Engineering and also serves as Associate
Director. He also served as the Interim Director of WINLAB from January to July2001. His research interests are in various aspects of wireless data transmissionincluding interference cancellation, wireless system modeling and performance,multiaccess protocols, and radio resource managementwith emphasis on pricing.
Dr. Mandayam received the Institute Silver Medal from the Indian Instituteof Technology in 1989 and the National Science Foundation CAREER Awardin 1998. He was selected by the National Academy of Engineering in 1999 forthe Annual Symposium on Frontiers of Engineering. He serves as an AssociateEditor for IEEE COMMUNICATIONS LETTERS.
David Famolari received the B.S. and M.S. degreesin electrical engineering from Rutgers University,Piscataway, NJ, in 1996 and 1999, respectively.
From 1997 to 1999, he was a part of the WirelessInformation Network Laboratory (WINLAB), Rut-gers University, where he carried out research in theareas of third-generation cellular networks, dynamicradio resource management, and power control forCDMA systems. In 1999, he joined the Applied Re-search Department, Telcordia Technologies, Morris-town, NJ, where he is involved in wireless IP mul-
timedia systems and distributed Internet appliance networking. His current re-search interests include mobility management for IP services, personal area net-working technology, and mobile computing.
Prathima Agrawal (S’74–M’77–SM’85–F’89)received the Ph.D. degree in electrical engineeringfrom the University of Southern California, LosAngeles.
She is Assistant Vice President of the InternetArchitecture Research Laboratory and ExecutiveDirector of the Computer Networking ResearchDepartment, Telcordia Technologies (formerlyBellcore), Morristown, NJ. She is also an AdjunctProfessor of electrical and computer engineering,Rutgers University, Piscataway, NJ. She worked
for 20 years at AT&T/Lucent Bell Laboratories, Murray Hill, NJ, whereshe was Head of the Networked Computing Research Department. Herresearch interests are computer networks, mobile and wireless computing, andcommunication systems and parallel processing. She has published over 150papers and has received or applied for more than 50 U. S. patents.
Dr. Agrawal received the Distinguished Member of Technical Staff Award ofAT&T Bell Laboratories in 1985 and the Telcordia CEO Award in 2000. She isa Member of the ACM. She was the recipient of the IEEE Computer Society’sDistinguished Service Award in 1990 and the IEEE Third Millennium Medal in2000. From 1998 to 2000, she Chaired the IEEE Fellow Selection Committee.
