spatialandtemporalfairnessinheterogeneoushsdpa-enabled ...overview of the current literature. in...

Hindawi Publishing CorporationEURASIP Journal on Wireless Communications and NetworkingVolume 2009, Article ID 682368, 12 pagesdoi:10.1155/2009/682368

Research Article

Spatial and Temporal Fairness in Heterogeneous HSDPA-EnabledUMTS Networks

Andreas Mader1 and Dirk Staehle2

1Department of Distributed Systems, University of Wuerzburg, Sanderring 2, 97070 Wurzburg, Germany2NEC Laboratories Europe, Kurfuersten-Anlage 36, 69115 Heidelberg, Germany

Correspondence should be addressed to Andreas Mader, [email protected]

Received 15 July 2008; Revised 27 November 2008; Accepted 29 December 2008

Recommended by Ekram Hossain

The system performance of an integrated UMTS network with both High-Speed Downlink Packet Access users and Release ’99 QoSusers depends on many factors like user location, number of users, interference, multipath propagation profile, and radio resourcesharing schemes. Additionally, the user behavior is an important factor; users of Internet best-effort applications tend to follow avolume-based behavior, meaning they stay in the system until the requested data is completely transmitted. In conjunction withthe opportunistic transmission scheme implemented in HSDPA, this has implications to the spatial distribution of active users aswell as to the time-average user and cell throughput. We investigate the relation between throughput, volume-based user behaviorand traffic dynamics with a simulation framework which allows the efficient modeling of large UMTS networks with both HSDPAand Release ’99 users. The framework comprises an HSDPA MAC/physical layer abstraction model and takes network aspects likeradio resource sharing and other-cell interference into account.

Copyright © 2009 A. Mader and D. Staehle. This is an open access article distributed under the Creative Commons AttributionLicense, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properlycited.

1. Introduction

Mobile network operators continue to deploy the High-Speed Downlink Packet Access (HSDPA) service in theirexisting Universal Mobile Telecommunication System(UMTS) networks. From the users perspective, the HSDPApromises high data rates (up to 14.4 Mbps with Release5) and low latency. From the perspective of an operator,HSDPA is hoped to play a key role for the much longed forbreakthrough of high-quality mobile data services. From atechnical perspective, HSDPA introduces a new paradigmto UMTS; instead of adapting the transmit power to theradio channel condition in order to ensure constant linkquality, HSDPA adapts the link quality to the radio channelconditions. This enables a more efficient use of scarceresources like transmit power, channelization codes, and alsohardware components.

The basic principle of the HSDPA is to adapt thelink to the instantaneous radio channel condition usingadaptive modulation and coding (AMC). HSDPA employs ashared channel, the High-Speed Downlink Shared channel

(HS-DSCH), which is used by all HSDPA users. With ashared channel, radio resources are occupied only if atransmission occurs, which enables a more efficient transportof bursty traffic. In each transport time interval (TTI), thescheduler located in the NodeB decides about the usersto be scheduled and about their data rate. The schedulingdecision can be either on behalf of channel quality indicator(CQI) reports from the user equipments (UE) to enableopportunistic scheduling schemes which use the air interfacemore efficiently, or simple nonopportunistic schemes likeround-robin can be used which shares the resources time fairamong the users.

An important aspect of HSDPA systems is the perceivedfairness of the connection metrics between the users. Thisis in contrast to pure UMTS Release ’99, where the circuit-switched design of the radio bearers guarantees equalQuality of Service (QoS) properties of all users of the sameservice class [1]. However, since in HSDPA the theoreticallyachievable data rate depends on the channel condition,the actual achieved data rates depend on user location,number of users, interference, scheduling discipline, and in

2 EURASIP Journal on Wireless Communications and Networking

integrated networks also on the number of dedicated channel(DCH) connections. In this work, we distinguish betweentwo fairness aspects. Spatial fairness refers to the spatialdistribution of the perceived data rates within a cell or sector.Temporal fairness refers to the long-term time-average userthroughput [2].

Our contribution is twofold: first, we propose a flow-level simulation framework which takes on the one handphysical layer aspects, scheduling disciplines, interference,and radio resource management schemes into account, butalso allows for simulation of large networks due to itsanalytical approach. Second, we investigate the impact ofthree well-known scheduling disciplines, namely round-robin, proportional fair, and Max C/I on the spatial userdistribution and on the system and user performance. Oneof our main findings is that Max C/I scheduling, althoughproviding sum-rate optimal rate allocations in static systemscenarios, performs worse than proportional fair schedulingif traffic dynamics are considered.

The remaining of this article is organized as follows:in the next section, we motivate our work and give anoverview of the current literature. In Section 3, we give abrief overview of the HSDPA. In Section 4, we explain radioresource sharing between DCH and HSDPA connections andformulate a model for the calculation of NodeB transmitpowers. In Section 6, a physical layer abstraction model forthe HSDPA is proposed which enables the calculation ofthe average throughputs per flow for different schedulingdisciplines. Simulation scenarios and numerical results arepresented in Section 7, followed by a conclusion in Section 8.

2. Motivation and RelatedWork

The focus of this work is the impact of elastic flows on thesystem performance. We have to distinguish between QoSflows which require a fixed bandwidth, as for voice calls overDCH transport channels, and “best-effort” or elastic flowswhich adapt their bandwidth requirements to the currentlyavailable bandwidth. Such a flow may be an FTP transfer orthe combined elements of a web page including inline objectssuch as embedded videos, that may be transmitted in parallelTCP connections. A flow can be loosely defined as a coherentstream of data packets with the same destination address [3].An important distinction between the two types of flows isthat QoS flows typically follow a time-based traffic model,which means that the user wants to keep the connection for acertain time span. In contrast, elastic flows are volume-based,that is, the user is satisfied as soon as a certain data volume istransmitted. An effect in this context which is that of spatialinhomogeneity, which has been mentioned in [4] for systemswithout AMC, and has been further investigated in [5, 6]for pure single-cell HSDPA systems. Users with bad radioconditions experience lower data rates than users with betterradio conditions, leading to a spatial unfairness, which wedefine as the discrepancy between location-dependent userarrival probabilities and the observed residence probabilitiesin steady state. We investigate this effect in Section 7.1 fordifferent scheduling disciplines in a multicell scenario, that

is, with consideration of other-cell interference, and withlocation-dependent arrival rates.

A related point is the system performance and fairnessof the perceived data rates under different schedulingregimes. In the literature, a large number of fundamentalworks investigate the tradeoff between fairness and systemcapacity in a wireless systems with opportunistic scheduling.Examples can be found in [2, 7–10], where in [7] theconcept of multiuser diversity (MUD) in downlink directionhas been investigated, motivated by the findings in [11]for the uplink direction. For HSDPA systems, researchmainly concentrated on variations of the proportional fairscheduler developed for the 1xEV-DO system [12]. Differentapproaches exist to include QoS constraints on delay ordata rate into the scheduling decision [13–17]. The fairnessof different schedulers in HSDPA systems is investigated in[18, 19]. Both works conclude that Max C/I provides thehighest system throughput. We compare user and systemthroughput for round-robin, Max C/I, and proportionalfair scheduling. The results show that on the one hand, asexpected the two channel-aware schemes clearly outperformround-robin scheduling, but on the other hand, proportionalfair scheduling leads to a higher time-average throughputthan Max C/I scheduling. We discuss this result in detail inSection 7.2.

Statistically valid results for integrated UMTS networksrequire long simulation runs or analytical approaches. Anintuitive example is the DCH blocking probability; a DCHuser which is located far from the antenna is subject to stronginterference from surrounding NodeBs, he may thereforerequire a very high transmit power. If this user additionallyhas a long call time, the influence on the blocking probabilityis significant. Since such events occur not very often withreasonable loads, long simulation runs are required. Theresults in this work are therefore generated with a simulationframework based on [20, 21], that uses analytic methodsto approximate the effects of the physical layer and thescheduling discipline on flow level. This allows for accurateand time-efficient simulations of large UMTS networks.

3. SystemDescription

We consider a UMTS network where HSDPA and DCHconnections share the same radio resources, namely transmitpower and channelization codes. The core of the HSDPA isthe HS-DSCH, which uses up to 15 codes with spreadingfactor (SF) 16 in parallel. The HS-DSCH enables two typesof multiplexing; time multiplex by scheduling the subframesto different users, and code multiplex by assigning each usera nonoverlapping subset of the available codes. The latterrequires the configuration of additional High-Speed SharedControl Channels (HS-SCCHs). Throughout this work weassume that only one HS-SCCH is present, hence considertime multiplex only.

In contrast to dedicated channels, where the transmitpower is adapted to the propagation loss with fast powercontrol and thus enabling a more or less constant bit rate,the HS-DSCH adapts the channel to the propagation loss

EURASIP Journal on Wireless Communications and Networking 3

Scheduling

decisions

CQI reporting

2 msTTI

SF16

code

s

UE 1UE 2UE 3

Figure 1: Schematic view of the HSDPA transport channel.

with AMC. The UE sends CQI values to the NodeB. TheCQI is a discretization of the received signal-to-interferenceratio (SIR) at the UE and ranges from 0 (no transmissionpossible) to 30 (best quality). The scheduler in the NodeBthen chooses a transport format combination (TFC) suchthat a predefined target BLER, which is often chosen as 10%,is fullfilled if possible. The TFC contains information aboutthe modulation (QPSK or 16QAM), the number of usedcodes (from 1 to 15), and the coding rate resulting in a certaintransport block size (TBS) that defines the information bitstransmitted during a TTI. A number of tables in [22] definea unique mapping between CQI and TFC. This means thatwith an increasing CQI, the demand on code resources is alsoincreasing. This leads to cases where a high CQI is reportedto the NodeB, but the scheduler has to select a lower TBS dueto lacking code resources. A schematic view of the HSDPAfunctionality is shown in Figure 1.

4. Sharing Code and Power Resources betweenHSDPA and DCH

A key issue of the radio resource management in HSDPAenhanced UMTS networks is the sharing of code and powerresources between DCHs, signaling channels, common chan-nels, and finally channels required for the HSDPA, namely,the HS-DSCH and the HS-SCCH. The signaling channelsand common channels mostly require a fixed channelizationcode and a fixed power as for the pilot channel (CPICH)or the forward access channel (FACH). The DCHs aresubject to fast power control which means that their powerconsumption depends on the cell or system load thatdetermines the interference at the UE. The general level ofpower consumption depends on the processing gain and therequired target bit-energy-to-noise ratio (Eb/N0) of the radioaccess bearer (RAB).

The HSDPA requires code and power resources. Codesare the channelization codes that are generated according tothe orthogonal variable spreading factor (OVSF) code tree.The number of codes that is available for a certain spreading

factor (SF) is equal to the spreading factor itself. A 384 kbpsDCH occupies an SF 8 channelization code. Accordingly,the maximum number of parallel 384 kbps users per sectoris theoretically 8. In practice, only 7 parallel 384 kbps usersare possible since the signaling and common channels alsorequire some code resources. Let us introduce an SF 512 codeas the basic code unit. Then, a DCH i with SF k occupiesci = 512/k code resources. An HSDPA code with SF 16requires cHS = 32 code resources. Let CDCH be the totalcode resources occupied by all DCHs, CCCH be the resourcesoccupied by signaling and common channels, and, CHS =nHS · cHS be the total number of code resources used by theHSDPA where nHS is the number of SF 16 codes allocatedto the HS-DSCH. The total number of code resources isequal to Ctot = 512. We consider adaptive code allocation[23, 24], which is illustrated in a simplified view (pilot andcontrol channels are omitted) in Figure 2 for both transmitpower and channelization codes. We further assume that thecodes are always optimally arranged in the code tree, andthat no code tree fragmentation occurs. The number of codesavailable for the HSDPA is then

nHS =⌊Ctot − CCCH − CDCH

cHS

⌋. (1)

Accordingly, the transmit power Tx,tot consists of aconstant part TCCH for common and signaling channels, apart TDCH for DCHs, and a part THS for the HS-DSCH. LetT∗ be the target transmit power at the NodeB. Then, the HS-DSCH power with adaptive power allocation is

THS = T∗ − TCCH − TDCH, (2)

whereT∗HS is the power reserved for the HS-DSCH, and TDCH

is the total DCH power averaged over some period of time.

5. Calculation of Downlink Transmit Powers

We define a UMTS network as a set L of NodeBs withassociated UEs, Mx. A DCH connection k corresponds to aradio bearer at NodeB x ∈ L with data rate Rk and coderesource requirements ck. Since the power consumed by theDCH connection is subject to power control, the receivedEb/N0 εk fluctuates around a target-Eb/N0 value ε∗k , whichis adjusted by the outer-loop power control such that thenegotiated QoS parameters like frame error rate are fulfilled.A common approximation for the average Eb/N0 value is

εk = W

Rk· Tk,x · dk,x

W ·N0 + Ik,oc + αi · Tx,tot · dk,x, (3)

where the orthogonality αk describes the impact of themultipath profile for DCH k, dk,x is the average path gainbetween NodeB x and UE k, W is the system chip rate, andN0 is the thermal noise density. We assume perfect powercontrol, that is, the mean Eb/N0 value meets exactly thetarget-Eb/N0 such that εk = ε∗k . The mean transmit powerrequirement of a DCH connection follows then as

Tk,x =ε∗k · Rk

W·(W ·N0 + Ik,oc

dk,x+ αk · Tx,tot

). (4)


Time TimeTr

ansm

itp

ower

Ch

ann

eliz

atio

nco

des

THSDPA

Tnon-HSDPA

Tmax Cmax

CHSDPA

Cnon-HSDPA

Adaptive radio resource allocation

Figure 2: Adaptive radio resource management scheme.

The average other-cell interference comprises thereceived powers of surrounding NodeBs such thatIk,oc = ∑

y∈L\x Ty,tot · dk,y. The total NodeB transmitpowers can be calculated with an equation system over allNodeBs. For that reason, we follow [25] and define the loadof NodeB x with respect to NodeB y as

ηx,y =∑

k∈Mx

ωk,y ,

with ωk,y =ε∗k · Rk

W·

⎧⎪⎨⎪⎩α, if L(k) = y,dk,y

dL(k), k, if L(k) /= y.

(5)

After some algebraic modifications, this allows us to formu-late the total DCH transmit power in a compact form as

Tx,DCH =∑y∈L

ηx,y · Ty,tot. (6)

In this equation, we neglect the thermal noise since in areasonable designed network its impact on the transmitpower requirements is minimal. Note also that the equationincludes the case y = x for the own-cell interference. For thetotal transmit power we introduce the boolean variable δy,HS

indicating whether at least one HSDPA flow is active in cellx. The total transmit power at NodeB x is then

Tx,tot = δx,HS · T∗x +(1− δx,HS

)

·(Tx,CCH +

∑y∈L

ηx,y · Ty,tot

).

(7)

This equation states that if the HS-DSCH is active, the totaltransmit power is equal to the target power. Otherwise, itconsist only of the DCH transmit power and the transmitpower for common channels. Introducing the vectors

V[x] = δx,HS · T∗x +(1− δx,HS

) · Tx,CCH, (8)

and matrix

M[x, y] = (1− δx,HS) · ηx,y (9)

leads to the matrix equation

T = V + M · T ⇐⇒ T = (I −M)−1 ·V , (10)

which provides the transmit powers of all NodeBs in thesystem. The matrix I is the identity matrix, and T is thevector of NodeB transmit powers Tx. The DCH and HSDPAtransmit powers are then calculated with (6) and (2).

6. HSDPA Physical Layer Model

Consider an HS-DSCH with power THS = ΔHS ·Ttot and nHS

parallel codes allocated to the HS-DSCH. Accordingly, theSIR at UE i for a RAKE receiver with perfect maximum ratiocombining is equal to

γi = ΔHS ·∑p∈P

Ttot · di,p,x

W ·N0 + Ioc,i +∑

r∈P \pTx,tot · di,r,x,

(11)

where di,p,x is the instantaneous propagation gain of signalpath p ∈ P . The UE measures the SIR and maps it to themaximum CQI with a transmission format that achieves aframe error rate of 10%. In [26] the following relation of SIRand CQI q is given:

q = max(

0, min(

30,⌊

SIR[dB]1.02

+ 16.62⌋))

. (12)

The CQI-value q defines the maximum possible TBSv(q), that can be transmitted in one TTI. It also defines thenumber of required parallel codes nHS(q). If the number ofavailable codes nHS is less than nHS(q), the scheduler selectsthe maximum possible TBS value according to nHS. Thismeans that an optimal usage of resources is only possibleif the transmission format according to the reported CQIutilizes all available codes. If too few code resources areavailable, power resources are wasted, and if too few powerresources are available, the CQI is too small to utilize allavailable codes. The reported CQI value depends essentiallyon the multipath profile, the users’ location, the availableHS-DSCH power, and the other-cell power. The number ofcodes required for a certain CQI value depends on the CQIcategory.

Above equations give the CQI and TBS for a concreteinstance of the propagation gains in particular of themultipath component power. For a simplified simulationand evaluation of the HSDPA performance, an approximatemodel for the HSDPA bandwidth similar to the orthogo-nality factor model for DCH is required. The orthogonalityfactor [27] is used to determine the signal-to-interferenceratio for a DCH i as

γi = W

Ri· Tx · dx,i

Ii,other + α · Ii,own, (13)

where W/Rk is the processing gain, Ii,other is the other-cell interference, and Ii,own = Tx,tot · dx,i is the own-cell


interference. The orthogonality factor α specifies the part ofthe power received from the own cell that contributes to theinterference due to multipath propagation. It captures theimpact of the multipath profile in a single value between 0.05and 0.4 depending on the multipath profile. For a deeperdiscussion of the orthogonality factor model please refer to[28–30] and the references therein.

Actually, the values γk, Iown, and Iother are mean valuesaveraged over the short-term fading. More precisely, weshould write (13) as

E[γi] = W

Ri· Tx,i · dx,i

E[Ii,other

]+ α · E[Ii,own

]

= W

Ri· Tx,i

Tx,tot· 1E[Ii,other

]/E[Ii,own

]+ α

.

(14)

The orthogonality factor model is not applicable to theHSDPA since it only yields the mean SIR. However, for theevaluation of the average HSDPA data rate of a UE at acertain location, the distribution of the reported CQI valuesis required. The essential assumption of the orthogonalityfactor model is that the mean normalized SIR, that is, the lastfraction in (14), is a function of the ratio Σ of average other-cell received power and average own-cell received power (orshort other-to-own-cell power ratio)

Σi =E[Ii,other

]E[Ii,own

] =∑

y /= x Ty,tot · dy,i

Tx,tot · dx,i. (15)

In [20], the orthogonality factor model is enhanced toyield not only the mean but also the standard deviation ofthe SIR in decibel scale as a function of Σi. Assuming thatthe distribution of the SIR follows a normal distribution thatis entirely characterized by its mean and standard deviation,the distribution of the reported CQI values, pCQI(q), isobtained from the cumulative density function (CDF) ofthe distribution of the SIR. Truncating the CQI distributionaccording to the available codes for the HS-DSCH yields thedistribution of the TBS as

pTBS(v) =

⎧⎪⎪⎨⎪⎪⎩pCQI(v(q)), if v(q) < v∗,

30∑q=v∗

pCQI(q), else,(16)

where v∗ is the maximum allowed TBS according to theavailable code resources. Accordingly, we denote the CDF ofthe CQI and TBS values with PCQI(q) and PTBS(v).

The physical layer abstraction model gives also insightsinto the impact of system parameters like multipath channelprofile, number of available codes and, UE category. Figure 3shows the gross data rate, that is, the throughput a single UEwould achieve, depending on the other-to-own-interferenceratio for the ITU Vehicular A, Pedestrian A, and VehicularB multipath propagation models. A profile with a strongdominating path, like in Pedestrian A, enables indeed veryhigh data rates up to 13 Mbps. In contrast, profiles with arelatively strong second path, like Vehicular A and VehicularB, lead to significantly lower data rates due to a higher

14

12

10

8

6

4

2

0

Mea

nT

BS

(kbi

t)

−30 −20 −10 0

Other-to-own cell power ratio Σ (dB)

ITU Pedestrian AITU Vehicular AITU Pedestrian B

Figure 3: Gross data rate for different channel profiles.

14

12

10

8

6

4

2

0

Mea

nT

BS

(kbi

t)

−30 −20 −10 0

Other-to-own cell power ratio Σ (dB)

UE cat. 10, Ped. AUE cat. 9, Ped. AUE cat. 7-8, Ped. AUE cat. 1–6, Ped. A

UE cat. 1–10, Ped. BUE cat. 11-12, Ped. BUE cat. 11-12, Ped. A

Figure 4: Gross data rates for different UE categories.

intersymbol interference. In fact, with these two models,it is sufficient to provide five SF 16 codes for the HS-DSCH. Figure 4 shows the gross data rates for different UEcategories, which reflect the capability for 16QAM, numberof parallel codes and, interscheduling time. Interesting isthat UEs without QAM 16 support (categories 11 and 12)have significantly lower data rates than UEs with QAM 16,although the transport block sizes are identically (categories1–6).

6.1. Scheduling. The scheduler in the NodeB has a largeinfluence on the user-level and system-level performance of


the HSDPA. Several proposals exist for HSDPA scheduling,from which we considered three of the most commonschemes. The channel-blind round-robin scheme selectsusers consecutively for transmission. The MaxTBS-schedulerchooses always the user with the currently best possibleTBS, including restrictions due to code resources. Finally,the proportional fair scheduler selects the user whichhas the proportionally best TBS in relation to its pastthroughput.

Channel-aware schedulers like MaxTBS and propor-tional fair benefit from multiuser diversity [7]. With anincreasing number of users in a cell, the probability tosee at least one user with good radio conditions alsoincreases. If “strong” users are favored by the scheduler,the aggregated cell throughput increases. Exploitation ofmultiuser diversity is therefore in the end beneficial for theoverall system capacity, also because reduced transmissiontimes for volume-based users leads to longer time periodswhere the HS-DSCH is switched off—which in turn reducesinterference.

6.1.1. Round-Robin Scheduling. The round-robin schedulerselects the users consecutively for transmission. In a suf-ficiently long time interval, the probability that a user kis selected is therefore approximately 1/|M|. Round-robinis a channel-blind scheduling discipline, which means thatthe average throughput of each mobile depends only onits channel condition and the number of users in thecell, but not on the channel conditions of other users.Consequently, the cell throughput does not benefit frommultiuser diversity. However, round-robin is robust and doesnot suffer from any convergence issues like proportional fairscheduling in some cases [31], and it is easy to implementdue to its simple principle. Round-robin is an allocation-fair scheduling discipline in the sense that, to every user, thesame amount of radio resources in terms of codes and powerare allocated. This approach is often sufficient to preventstarvation of users at the cell edge.

6.1.2. MaxTBS Scheduling. With MaxTBS (or Max C/I)scheduling, the user with the currently best TBS is scheduled.This scheduling discipline maximizes the sum-rate capacity(in our context the cell throughput) given the saturated case,that is, all users have at least one packet to transmit [32, 33].If two or more users have the maximum possible TBS, arandom user out of this set is selected with equal probability.In contrast to round-robin scheduling, the throughput ofa user depends not only on its own location, but alsoon the location of the other users. In [6], this schedulingdiscipline is modeled as a priority queue, where locationscloser to the NodeB have higher priority than locationsfarther away. However, it is also possible to calculate theaverage throughput directly from the TBS distributions ofthe users. In this work we use the formulation we developedin [21]. MaxTBS strongly favors the user with the bestchannel quality. This implicates that users with weak radioconditions are penalized and perceive on average very lowdata rates, leading to unfair rate allocations. We show in the

next section how this behavior negatively affects the averagethroughput if traffic dynamics are considered.

6.1.3. Proportional Fair Scheduling. Proportional fair (PF)scheduling is a scheduling discipline which has been devel-oped for the 1xEv-DO-system in the downlink [12]. Thebasic principle is to allocate each user proportional to its linkquality and its past throughput. This is achieved by selectingthe user that has the best instantaneous relative throughputover its past throughput, which is often calculated witha sliding window approach. However, different versions ofPF scheduling exist. The most fundamental difference isthe way how the past throughput is calculated. The firstvariant updates the past throughput every scheduling periodregardless whether the user has been scheduled or not, thesecond variant updates the past throughput only if the user isindeed chosen for transmission. The difference between bothversions is that in the first case the mean throughput of a useris proportional to its channel quality only, while in the secondcase it is also related to the generated traffic. In [31, 34]it is argued that both variants approximately lead to thesame results in case of statistically identical fades and infinitebacklogs. The second assumption is reasonable during theinterevent time, while the first assumption is contradicted bythe fact that the shape of the CQI distribution depends on thelevel of received other-cell interference. A direct formulationof the flow-average throughput and a comparison betweenboth variants can be found in [21].

7. Flow-Level Performance Results

UMTS networks are dynamic systems because of the mutualdependency among the transmit powers of different cells.This means that a well-designed performance evaluationhas to consider networks with a reasonable size in orderto capture these effects and their impact on flow-levelperformance properly. We consider two different types ofnetworks: a 19-NodeB hexagonal layout with a NodeBdistance of 1.2 km, and an irregular layout with 22 NodeBswhich is generated from a Voronoi tessellation. The networkareas are partitioned into area elements with an edge lengthof 25 m. Figure 5 shows the irregular network with antennalocations (dots) and arrival cluster centers (stars). In thehexagonal layout, user arrive according to a homogeneousPoisson process such that arrival rates are equal for all areaelements. In the irregular network, users arrive according toa clustered Poisson process as described in [25] and shown inFigure 6; the total arrival rate λ f in an area element f resultsfrom the superposition of circular clusters with constantarrival rates. In the irregular network therefore not only thelayout but also the arrival process is heterogeneous.

Results are generated with a time-dynamic simulationwhich considers the HSDPA data traffic of a user as a flowwith a certain data volume. The network area is discretizedinto a set of area elements with an edge length of 25 m.The time axis is divided in interevent times. We assume thatbetween two events the users stay roughly within an areaelement.


1

23

4

56

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

7

6

5

4

3

2

1

(km

)

1 2 3 4 5 6 7

(km)

Figure 5: Irregular network layout. Dots indicate NodeB (antenna)locations, stars mark cluster centers.

7

6

5

4

3

2

1

(km

)

1 2 3 4 5 6 7

(km)

Figure 6: Inhomogeneous arrival densities. Darker colors indicatehigher probability of arrival.

We consider two types of events: arrival events, that is,the arrival of a new user into the system, and departureevents, which may occur if an HSDPA user has receivedall its data or if the call time of DCH user is reached.On arrival of a new user, admission control for DCH andHSDPA is performed. The admission control for DCHconnections is threshold-based. An incoming connectionis blocked if the total transmit power including the newconnection exceeds the target transmit power, or if theavailable code resources are not sufficient. For this purpose,

the required transmit power is calculated at the servingNodeB under the worst-case assumption that all NodeBstransmit with the target power in order to prevent possibleoutage. For the HSDPA, we assume a count-based admissioncontrol which restricts the maximum number of concurrentconnections to a fixed value. If the incoming connection isadmitted into the system, the call time or the data volume,depending on the user type, is calculated according to therespective distribution parameters. We assume exponentiallydistributed call times with mean E[T] = 120 s for DCH usersand exponentially distributed flow sizes with mean volumeE[V] = 100 KB for HSDPA users. The arrival rate of theDCH users is determined from the offered DCH code loaddefined as

ρc =∑s∈S

λsμs· csCtot

, (17)

where μs = 1/E[Ts], and the index s denotes the service classof the radio bearer.

On each event, the system variables are recalculatedif necessary. If the event is generated by a DCH arrivalor departure, HSDPA code resources in the relevant cellsare decreased or increased according to the DCH coderequirements. Additionally, the total transmit powers areupdated for all NodeBs in order to capture the new inter-ference situation. Transmit power recalculation is also doneif the HS-DSCH is switched on or off because of HSDPAuser arrivals or departures. In all cases, the data volumetransmitted by HSDPA users within the past intereventtime is subtracted from their remaining data volumes. NewHSDPA data rates are calculated, taking the new radioresource and interference situation into account. Finally, theexpected departure times of the HSDPA users are updatedaccording to the remaining data volumes and data rates.

7.1. Volume-Based Traffic Model and Spatial Fairness. Asmentioned before, an important distinction between QoSand elastic flows is that QoS flows typically follow a time-based traffic model, which means that the user wants tokeep the connection a certain time span, for example, for thetime of a conversation. In contrast, elastic flows are volume-based, that is, the user leaves the system as soon as a certaindata volume is transmitted. In reality, the user behavior isa mixture between both models, depending on factors likeuser satisfaction, pricing models, type of content. However,the two models can be seen as the extremes of the actual userbehavior.

A time-based traffic model implicates that the number ofcurrently active users is independent of the perceived datarates. Moreover, the spatial distribution of the number ofusers is corresponding to the spatial arrival process; if usersarrive with arrival rate λ, the number of concurrently activeusers in steady-state follows according to Little as λ/μ, if noblocking occurs.

A volume-based traffic model means that users stayin the system until their service demands are fullfilled.Therefore, the number of active users depends on theassigned data rates. In HSDPA systems, the data rate depends


on the channel quality, which means that users with lowaverage channel qualities stay longer in the system thanthose with good channel qualities. Since the average channelquality is dominated by the other-cell interference, usersat the cell edges stay longer in the system than users inthe center of the cell. This implies that the spatial arrivalprocess and the spatial steady state distribution are notdirectly related anymore, a fact that complicates planningof HSPDA networks significantly. One reason is that MonteCarlo methods [35] now have to estimate the spatial userpopulation for every snapshot, which is difficult withoutknowledge of the the currently ongoing flows. With round-robin scheduling, a direct formulation of the mean transfertime was found in [5, 24], since in that case the data ratesof the users only depend on the number of users and theirposition, but are otherwise independent of each other.

We now clarify the effect of spatial heterogeneity withsome example scenarios. Figure 7 shows the arrival proba-bility and the residency probability versus the distance tothe antenna for cell number 2 from the irregular scenario.The arrival probability describes the probability that a userarrives in this cell at a certain point, while the residenceprobability reflects the spatial distribution of the users in thecell in steady state. The spiky shape of the curves is due to thediscretization of the cell area into area elements. It is obviousthat arrival and residence probabilities are not equal, and thatthe magnitude of the deviation depends on the schedulingdiscipline. MaxTBS scheduling shows the highest deviation,since users close to the antenna leave the system much earlierthan users farther away. An interesting result is that residenceprobabilities with proportional fair scheduling fir slightlybetter to the arrival probabilities if compared to round-robinscheduling. We will see later that this effect comes from thefact that the proportional fair scheduler favors users on thecell edges.

Figure 8 shows the corresponding ratio between arrivaland residence probability in the same cell. With time-basedusers, the ratio would be equal to one at all distances. Withvolume-based users and MaxTBS-scheduling, the probabilityto meet a user at the cell edge is four times higher than thearrival probability at the same location.

The deviation of arrival and residence probabilities isthe result of spatial unfairness regarding the data rateallocation. This is demonstrated in Figure 9, which showsthe average user throughput depending on the distanceto the antenna. MaxTBS-scheduling favors strongly userin the cell center, and thus shows the highest degree ofunfairness. Proportional fair and round-robin schedulinglead to more balanced results. The difference between round-robin and proportional fair reflects the scheduling gain dueto multiuser diversity. Note that the gain of the proportionalfair scheduler over the round-robin scheduler is nearlyindependent of the distance.

Finally, in Figure 10, the same statistic for the center cellof the homogeneous scenario is shown, but in a scenario witha higher DCH load of ρc = 0.6. Here, the lack of resourcesleads to low throughputs, such that the aforementionedfavoring of user at the cell edge with proportional fairscheduling is clearly visible. This is caused by the higher

0.05

0.04

0.03

0.02

0.01

0

Pro

babi

lity

0 200 400 600 800 1000 1200

Distance to antenna (m)

Proportional-FairMaxTBS

Round-RobinArrival probability

Figure 7: Arrival and residence probabilities for cell 2 in theirregular network with inhomogeneous user arrivals and DCHoffered load ρc = 0.4. The black line with diamond markersindicates the user arrival probability.

5

4

3

2

1

0

Rat

ioar

riva

l/re

side

nce

prob

abili

ty

0 200 400 600 800 1000 1200


Proportional-FairMaxTBSRound-Robin

Figure 8: Ratio between arrival and residence probabilities.MaxTBS-scheduling leads to the highest inhomogeneity.

variance of the TBS distribution of users which experiencemore other-cell interference than users close to the antenna,see also [36] for a discussion of this effect.

7.2. Impact of Scheduling Disciplines. We now investigatethe impact of different scheduling disciplines on the overallperformance of the network. We consider the homogeneousscenario with hexagonal cell layout and increase the offeredDCH load from 0.1 to 0.8.


2000

1500

1000

500

0

Mea

nda

tara

te(k

bps)

0 200 400 600 800 1000 1200



Figure 9: Mean throughput versus distance to antenna with offeredDCH load ρc = 0.4 for cell 2 of the irregular scenario.

600

500

400

300

200

100

Mea

nda

tara

te(k

bps)

0 100 200 300 400 500 600



Figure 10: Mean throughput versus distance to antenna for thecenter cell of the hexagonal scenario with offered DCH loadρc = 0.6.

Figure 11 shows the resulting time-average cell and userthroughput versus the offered DCH load. As expected, thechannel-aware scheduling disciplines lead to better resultsthan the channel-blind round-robin discipline, regardlessof the DCH load. However, with higher DCH load,the difference between the scheduling disciplines becomessmaller, since the lack of code resources prevents an efficientexploitation of multiuser diversity. An interesting result isthat proportional-fair scheduling leads to higher throughput

2500

2000

1500

1000

500

0

Tim

e-av

erag

eth

rou

ghpu

t(k

bps)

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8

Offered DCH code load ρc

Cell throughput

User throughput


Figure 11: Time-average user and cell throughput versus offeredDCH load for different scheduling disciplines.

curves than MaxTBS-scheduling, which is at a first glancecounter intuitive. MaxTBS-scheduling maximizes cumulateddata rates (the sum-rate) for a static scenario, that is, for afixed number of ongoing flows and consequently also duringany interevent time [32]. This also means that MaxTBS-scheduling always leads to a higher cell throughput thanproportional-fair scheduling if we consider the same snap-shot for both schedulers, reflecting the well known tradeoffbetween system capacity (defined as cell throughput) andfairness of data rate allocation (see, e.g., [10]).

However, this unfairness means that in cases wherethe differences between the average channel conditions arelarge, the MaxTBS scheduler has a strong tendency tooverproportionally favor the best user, such that the datarates of the remaining UEs are very low. These users stayvery long in the system which is then reflected in thetime-average cell and user throughput. With proportional-fair scheduling the data rate of users with good channelconditions is lower, however this is compensated with lowersojourn times of users with bad channel conditions. Notethat in principle this also holds for round-robin scheduling,but channel-blindness overweights this effect such that theaverage throughput is indeed lower.

In the literature, some numerical results seem to con-tradict the results presented here. In [37, 38], the systemthroughput for round-robin, proportional fair and Max C/I(i.e., MaxTBS) is shown, and it is concluded that Max C/Ischeduling provides the highest average cell throughput.However, the results apply to static scenarios with persistentdata flows for a fixed number of users. In such a scenario,MaxTBS scheduling is optimal, but it is not comparable withthe flow-level throughput in system with traffic dynamics.In [19], users arrive according to a Poisson process andrequest 100 KB of data, which is incidentally the same


average amount of data as in our scenario. However, usersare dropped from the system if they stay longer than 12.5seconds in the system, such that the time-average usersojourn time is reduced. So, in fact this study employsa mixture between time-and volume-based traffic model.Consequently, the results show a small performance gainfor Max C/I scheduling. Similarly, in [18] users are droppedfrom the system if their throughput is lower than 9.6 kbps.It is not clear over which time span the throughput ismeasured, but the dropping of low-bandwidth users skewsthe time-average throughput to the benefit of the Max C/Ischeduler.

Figure 12 shows the CDF of the user and cell throughputsfor an offered DCH load of ρc = 0.4. The CDF of the MaxTBSscheduler confirms the time-average throughput curves; alarge portion of the probability weight is on very low datarates, but in the same time the higher quantiles, for example,for 0.8, are higher than for proportional fair and round-robinscheduling. In terms of fairness, it is remarkable that theshape of the curves for Round-robin and proportional-fairare similar with exception of a small peak for low data ratesfor the proportional fair scheduler. Also note the stair-likeshape of cell-throughput CDF for low data rates, which iscaused by preemption from DCH connections.

Figure 13 exemplarily demonstrates the behavior of thethree schedulers for scenario with three users which havefixed data volumes and Σ-values of −20 dB, −10 dB, and0 dB. The figure shows the remaining total data volumeversus time. Figure 14 shows the corresponding data rates.With MaxTBS scheduling, the first and second users leavethe system faster than with the other disciplines (indicatedby the vertical dashed lines), but the remaining data volumeof the “worst” user with Σ = 0 dB is so large that in total, theproportional-fair scheduler needs less time to transport thewhole data volume. Note that it depends on channel profileand cell layout how large the advantage of the proportional-fair scheduler is and whether it exists at all.

8. Conclusion and Outlook

We investigated spatial and temporal fairness aspects ofintegrated HSDPA-enhanced UMTS networks on flow level.Results have been generated with a flow-level simulationwhich considers the network-wide interference situationand its impact on DCH transmit powers and HSDPA datarates. The latter are calculated with a physical layer abstrac-tion model which considers code resources, multipath-propagation, HS-DSCH transmit power, and differentscheduling disciplines.

The numerical results have been generated within two-network scenarios: a homogeneous scenario with hexagonalcells and equal arrival rates over the whole space, and aninhomogeneous scenario with irregular-shaped cells andlocation-dependent arrival densities. An expected result isthat the shared-bandwidth approach of the HSDPA transportchannel leads to spatial user residence probabilities whichare different to the corresponding arrival probabilities. Thedegree of unfairness depends on the employed scheduling

1

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0

CD

Fof

thro

ugh

put

0 500 1000 1500 2000 2500 3000 3500 4000

Throughput (kbps)

Cell throughput

User throughput


Figure 12: CDF of user and cell throughput for an offered DCHload of ρc = 0.4.

25

20

15

10

5

0

Tota

lvol

um

e(k

bit)

0 2 4 6 8 10 12 14

Time (s)

MaxTBSProportional-FairRound-Robin

Figure 13: Total remaining data volume versus time for a three-user scenario with fixed data volume. Vertical dashed lines indicatedepartures.

discipline; “greedy” scheduling disciplines like MaxTBSlead to a high unfairness, while channel-blind round-robinscheduling and proportional fair scheduling show similarresults. However, proportional-fair scheduling has a nearlyconstant relative gain in terms of throughput over round-robin scheduling independent of the distance to the antennaand of the arrival densities.

A further objective of this paper is to understand theflow-level performance of different scheduling disciplines.


3500

3000

2500

2000

1500

1000

500

0

Tota

lcel

lbit

rate

(kbp

s)

0 2 4 6 8 10 12 14Time (s)

MaxTBSProportional-FairRound-Robin

Figure 14: Corresponding cell throughput versus time.

The comparison between round-robin, proportional fair,and MaxTBS scheduling showed that, remarkably, propor-tional fair scheduling has a slight performance gain interms of average cell and user throughput. The reason isthat although MaxTBS-scheduling maximizes the sum ratewithin a static scenario, traffic dynamics, and the highunfairness of the data rate allocation with MaxTBS favorsin the end proportional fair scheduling. This shows that theconsideration of traffic dynamics is a crucial point of theperformance evaluation of shared bandwidth systems, andit encourages further investigations of the relation betweenphysical layer parameters and flow-level performance.

References

[1] “Quality of service (QoS) concept and architecture,” Tech.Rep. TS 23.107 V6.1.0, 3GPP, Valbonne, France, March 2004.

[2] X. Liu, E. K. P. Chong, and N. B. Shroff, “Optimal opportunis-tic scheduling in wireless networks,” in Proceedings of the 58thIEEE Vehicular Technology Conference (VTC ’03), vol. 3, pp.1417–1421, Orlando, Fla, USA, October 2003.

[3] J. W. Roberts, “A survey on statistical bandwidth sharing,”Computer Networks, vol. 45, no. 3, pp. 319–332, 2004.

[4] S. Lu, V. Bharghavan, and R. Srikant, “Fair scheduling in wire-less packet networks,” IEEE/ACM Transactions on Networking,vol. 7, no. 4, pp. 473–489, 1999.

[5] R. Litjens, J. van den Berg, and M. Fleuren, “Spatial trafficheterogeneity in HSDPA networks and its impact on networkplanning,” in Proceedings of the 19th International TeletrafficCongress (ITC ’05), pp. 653–666, Bejing, China, August-September 2005.

[6] H. van den Berg, R. Litjens, and J. Laverman, “HSDPAflow level performance: the impact of key system andtraffic aspects,” in Proceedings of the 7th ACM InternationalSymposium on Modeling, Analysis and Simulation of Wirelessand Mobile Systems (MSWiM ’04), pp. 283–292, Venice, Italy,October 2004.

[7] P. Viswanath, D. N. C. Tse, and R. Laroia, “Opportunisticbeamforming using dumb antennas,” IEEE Transactions onInformation Theory, vol. 48, no. 6, pp. 1277–1294, 2002.

[8] X. Liu, E. K. P. Chong, and N. B. Shroff, “A frameworkfor opportunistic scheduling in wireless networks,” ComputerNetworks, vol. 41, no. 4, pp. 451–474, 2003.

[9] Y. Liu and E. Knightly, “Opportunistic fair scheduling overmultiple wireless channels,” in Proceedings of the 22nd AnnualJoint Conference of the IEEE Computer and CommunicationsSocieties (INFOCOM ’03), vol. 2, pp. 1106–1115, San Fran-cisco, Calif, USA, March-April 2003.

[10] L. Yang, M. Kang, and M.-S. Alouini, “On the capacity-fairnesstradeoff in multiuser diversity systems,” IEEE Transactions onVehicular Technology, vol. 56, no. 4, part 1, pp. 1901–1907,2007.

[11] R. Knopp and P. A. Humblet, “Information capacity andpower control in single-cell multiuser communications,” inProceedings of IEEE International Conference on Communica-tions (ICC ’95), vol. 1, pp. 331–335, Seattle, Wash, USA, June1995.

[12] A. Jalali, R. Padovani, and R. Pankaj, “Data throughputof CDMA-HDR: a high efficiency-high data rate personalcommunication wireless system,” in Proceedings of the 51stIEEE Vehicular Technology Conference (VTC ’00), vol. 3, pp.1854–1858, Tokyo, Japan, May 2000.

[13] P. Ameigeiras, J. Wigard, and P. Mogensen, “Performance ofthe M-LWDF scheduling algorithm for streaming services inHSDPA,” in Proceedings of the 60th IEEE Vehicular TechnologyConference (VTC ’04), vol. 2, pp. 999–1003, Los Angeles, Calif,USA, September 2004.

[14] M. Lundevall, B. Olin, J. Olsson, et al., “Streaming applicationsover HSDPA in mixed service scenarios,” in Proceedings of the60th IEEE Vehicular Technology Conference (VTC ’04), vol. 2,pp. 841–845, Los Angeles, Calif, USA, September 2004.

[15] A. K. F. Khattab and K. M. F. Elsayed, “Channel-qualitydependent earliest deadline due fair scheduling schemes forwireless multimedia networks,” in Proceedings of the 7th ACMInternational Symposium onModeling, Analysis and Simulationof Wireless and Mobile Systems (MSWiM ’04), pp. 31–38,Venice, Italy, October 2004.

[16] M. C. Necker, “A comparison of scheduling mechanismsfor service class differentiation in HSDPA networks,” AEU -International Journal of Electronics and Communications, vol.60, no. 2, pp. 136–141, 2006.

[17] T. E. Kolding, “QoS-aware proportional fair packet schedulingwith required activity detection,” in Proceedings of the 64thIEEE Vehicular Technology Conference (VTC ’06), pp. 1–5,Montreal, Canada, September 2006.

[18] P. Ameigeiras, J. Wigard, and P. Mogensen, “Performance ofpacket scheduling methods with different degree of fairness inHSDPA,” in Proceedings of the 60th IEEE Vehicular TechnologyConference (VTC ’04), vol. 2, pp. 860–864, Los Angeles, Calif,USA, September 2004.

[19] T. E. Kolding, F. Frederiksen, and P. E. Mogensen, “Perfor-mance aspects of WCDMA systems with high speed downlinkpacket access (HSDPA),” in Proceedings of the 56th IEEEVehicular Technology Conference (VTC ’02), vol. 1, pp. 477–481, Vancouver, Canada, September 2002.

[20] D. Staehle and A. Mader, “A model for time-efficient HSDPAsimulations,” in Proceedings of the 66th IEEE Vehicular Technol-ogy Conference (VTC ’07), pp. 819–823, Baltimore, Md, USA,September-October 2007.

[21] A. Mader and D. Staehle, “A flow-level simulation frameworkfor HSDPA-enabled UMTS networks,” in Proceedings of the


10th ACM Symposium on Modeling, Analysis, and Simulationof Wireless and Mobile Systems (MSWiM ’07), pp. 269–278,Chania, Greece, October 2007.

[22] “Medium Access Control (MAC) protocol specification,” Tech.Rep. TS 25.321 V6.6.0, 3GPP, Valbonne, France, September2005.

[23] H. Holma and A. Toskala, Eds., HSDPA/HSUPA for UMTS:High Speed Radio Access for Mobile Communications, JohnWiley & Sons, New York, NY, USA, 1st edition, 2006.

[24] A. Mader, D. Staehle, and M. Spahn, “Impact of HSDPA radioresource allocation schemes on the system performance ofUMTS networks,” in Proceedings of the 66th IEEE VehicularTechnology Conference (VTC ’07), pp. 315–319, Baltimore, Md,USA, September-October 2007.

[25] D. Staehle and A. Mader, “An analytic model for deriving thenode-B transmit power in heterogeneous UMTS networks,” inProceedings of the 59th IEEE Vehicular Technology Conference(VTC ’04), vol. 4, pp. 2399–2403, Milan, Italy, May 2004.

[26] F. Brouwer, I. de Bruin, J. C. Silva, N. Souto, F. Cercas, andA. Correia, “Usage of link-level performance indicators forHSDPA network-level simulations in E-UMTS,” in Proceedingsof the 8th IEEE International Symposium on Spread SpectrumTechniques and Applications (ISSSTA ’04), pp. 844–848, Syd-ney, Australia, August-September 2004.

[27] H. J. Kushner and P. A. Whiting, “Convergence ofproportional-fair sharing algorithms under generalconditions,” IEEE Transactions on Wireless Communications,vol. 3, no. 4, pp. 1250–1259, 2004.

[28] D. N. Tse, “Optimal power allocation over parallel Gaussianbroadcast channels,” in Proceedings of IEEE InternationalSymposium on Information Theory (ISIT ’97), p. 27, Ulm,Germany, June-July 1997.

[29] L. Li and A. J. Goldsmith, “Capacity and optimal resourceallocation for fading broadcast channels—I. Ergodic capacity,”IEEE Transactions on Information Theory, vol. 47, no. 3, pp.1083–1102, 2001.

[30] S. Borst, “User-level performance of channel-aware schedulingalgorithms in wireless data networks,” IEEE/ACMTransactionson Networking, vol. 13, no. 3, pp. 636–647, 2005.

[31] U. Turke, M. Koonert, R. Schelb, and C. Gorg, “HSDPAperformance analysis in UMTS radio network planning sim-ulations,” in Proceedings of the 59th IEEE Vehicular TechnologyConference (VTC ’04), vol. 5, pp. 2555–2559, Milan, Italy, May2004.

[32] J. M. Holtzman, “CDMA forward link waterfilling powercontrol,” in Proceedings of the 51st IEEE Vehicular TechnologyConference (VTC ’00), vol. 3, pp. 1663–1667, Tokyo, Japan,May 2000.

[33] A. Furuskar, S. Parkvall, M. Persson, and M. Samuelsson, “Per-formance of WCDMA high speed packet data,” in Proceedingsof the 55th IEEE Vehicular Technology Conference (VTC ’02),vol. 3, pp. 1116–1120, Birmingham, Ala, USA, May 2002.

[34] A. Haider, R. Harris, and H. Sirisena, “Simulation-basedperformance analysis of HSDPA for UMTS networks,” inProceedings of the Australian Telecommunication Networks andApplications Conference (ATNAC ’06), Melbourne, Australia,December 2006.

[35] U. Turke, M. Koonert, R. Schelb, and C. Gorg, “HSDPAperformance analysis in UMTS radio network planning sim-ulations,” in Proceedings of the 59th IEEE Vehicular TechnologyConference (VTC ’04), vol. 5, pp. 2555–2559, Milan, Italy, May2004.

[36] J. M. Holtzman, “CDMA forward link waterfilling powercontrol,” in Proceedings of the 51st IEEE Vehicular Technology

Conference (VTC ’00), vol. 3, pp. 1663–1667, Tokyo, Japan,May 2000.

[37] A. Furuskar, S. Parkvall, M. Persson, and M. Samuelsson, “Per-formance of WCDMA high speed packet data,” in Proceedingsof the 55th IEEE Vehicular Technology Conference (VTC ’02),vol. 3, pp. 1116–1120, Birmingham, Ala, USA, May 2002.

[38] A. Haider, R. Harris, and H. Sirisena, “Simulation-basedperformance analysis of HSDPA for UMTS networks,” inProceedings of the Australian Telecommunication Networks andApplications Conference (ATNAC ’06), Melbourne, Australia,December 2006.

spatialandtemporalfairnessinheterogeneoushsdpa-enabled ...overview of the current literature. in...

Documents