hspa 3.5 g

1Mobility: A Double-Edged Swordfor HSPA Networks

A Large-scale Test on Hong Kong Mobile HSPA Networks

Fung Po Tso, Graduate Member, IEEE , Jin Teng, Student Member, IEEE , Weijia Jia, SeniorMember, IEEE , and Dong Xuan, Member, IEEE ,

School of Computing Science, University of Glasgow, UKDepartment of Computer Science, City University of Hong Kong, HKSAR

Department of Computer Science and Engineering , The Ohio-State University, US

AbstractThis paper presents an empirical study on the performance of mobile High Speed Packet Access (a 3.5G cellularstandard usually abbreviated as HSPA) networks in Hong Kong via extensive eld tests. Our study, from the viewpoint of endusers, covers virtually all possible mobile scenarios in urban areas, including subways, trains, off-shore ferries and city buses.We have conrmed that mobility has largely negative impacts on the performance of HSPA networks, as fast-changing wirelessenvironment causes serious service deterioration or even interruption. Meanwhile our eld experiment results have shownunexpected new ndings and thereby exposed new features of the mobile HSPA networks, which contradict commonly heldviews. We surprisingly nd out that mobility can improve fairness of bandwidth sharing among users and trafc ows. Also thetriggering and nal results of handoffs in mobile HSPA networks are unpredictable and often inappropriate, thus calling for fastreacting fallover mechanisms. Moreover, we nd that throughput performance does not monotonically decrease with increasedmobility level. We have conducted in-depth research to furnish detailed analysis and explanations to what we have observed. Weconclude that mobility is a double-edged sword for HSPA networks. To the best of our knowledge, this is the rst public report ona large scale empirical study on the performance of commercial mobile HSPA networks.

Index TermsMobile HSPA Networks, Performance Evaluation, Bandwidth Sharing

1 INTRODUCTIONWireless communication networks have seen a largeincrease in popularity and capacity in real world de-ployment [11]. The recent High Speed Packet Access(HSPA) [2], which is commonly referred to as 3.5Gcellular standard, is conceived as a natural evolutionof the existing Wideband CDMA (WCDMA, knownas 3G). HSPA is a combination of High Speed Down-link Packet Access (HSDPA) and High Speed UplinkPacket Access (HSUPA). It has emerged as one ofthe most dominant beyond-3G technologies and hasbeen serving many regions and countries to supportbroadband applications.

1.1 Motivation

HSPA offers data rates up to 14.4 Mbps in downlinkand 5.76 Mbps in uplink for stationary users. There-fore, stationary users can easily access broadbandapplications in most areas with good signal qualitythrough HSPA [10]. Nevertheless, the most importantadvantage of wireless communication systems is tosupport users mobility. Mobility is crucial in our dailylife as it provides end users with exibility and op-portunities to easily connect with people world wide.

Many companies often adopt mobile data access to ex-pand their business coverage, save time and improveproductivity. The most common forms of mobilityobviously include walking, driving and traveling viapublic transport. Passengers on public transport aremore likely to entertain themselves via forthcomingmobile data access than people who are driving vehi-cles or walking. In Hong Kong, for instance, millionsof people use public transport that traverses the cityeveryday. Many, virtually each, of these passengersown a mobile phone [1]. Among them, more andmore have Smartphones and Netbooks. With theseon-body small devices communicating over 3G andbeyond networks, people can entertain themselveson the move using broadband applications such asYouTube [29]. They can also share live video capturedby mobile phone cameras like Livecast [18], or Veo-dia [25]. All these broadband applications require theupload and/or download of realtime multimedia con-tents, for which there is support available to stationaryusers [10]. But that is not always feasible when usersare on the move.

Despite its importance, the performance of HSPAnetworks in the mobile environments has not beendeeply investigated before in a large-scale operationalnetwork via on-site tests. Extensive efforts have beendevoted to theoretically analyze network performance

Digital Object Indentifier 10.1109/TPDS.2011.289 1045-9219/11/$26.00 2011 IEEE

IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMSThis article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.

2or merely to study general network performance[10] [16] [5]. The empirical study of WCDMA net-works has been performed in stationary environ-ments [24], but this provides no information aboutthe performance for mobile users. The performanceof mobile HSPA remains highly unclear in practice, asradio conditions vary with different speeds, locationsand time in a mostly unpredictable fashion. Hence, anextensive empirical study for mobile HSPA is neededto resolve these performance uncertainties.

1.2 Contributions

We have studied the performance of two largest com-mercial HSPA networks in Hong Kong with extensiveeld tests. Our tests cover virtually all possible mobilescenarios in urban areas - from land scenarios (e.g.buses) to sea scenarios (e.g. ferries) and from groundscenarios (e.g. trains) to underground scenarios (e.g.subways) - totaling over 100 km distance during a 3months period. We investigate their data throughput,round trip time (RTT), and packet loss rate underdifferent network trafc loads and various trafcpatterns. We also examine the behavior of networkresource allocation mechanisms and the fairness ofthe radio-link schedulers in allocating the bandwidthresources to multiple data calls in a HSPA network.Here we present our ndings and conclusions on avery high level:

First, we have conrmed with solid data and de-tailed analysis that mobility has largely negative im-pact on the performance of HSPA networks. Fastchanging physical wireless environment and frequenthandoffs both contribute to the degradation of overallHSPA service provisioning. In particular, we make thefollowing two observations:

Mobile HSPA performance is greatly differentfrom static HSPA performance. We record a largespread of HSPA throughput (e.g., 0 kbps to 3500kbps) and signal quality (-1 dBm to -32 dBm)for mobile User Equipments (UEs). In spite ofthe degradation, our data show that subways notonly have the highest signal quality and coveragebut also give the highest throughput among allmobile scenarios.

We measure and report HSPAs performance intransitional regions. Handoffs, which normallyhappen in the transitional regions, are longthought to have been properly handled since thetime of 2G networks. However they are still amajor problem restricting the trafc volume ofHSPA networks. We show that during a hand-off, the service is virtually interrupted and thethroughput of end users decreases to almost zero.

Second, our eld experiment results have shownseveral unexpected new ndings and thereby ex-posed some misunderstandings commonly held forthe HSPA networks:

Common View: Mobility is irrelevant, if not detrimen-tal, to the fairness in HSPA bandwidth sharing amongusersWe observe that the bandwidth sharing practicein stationary HSPA environments is unfair. Incontrast, mobility surprisingly improves fairnessof bandwidth sharing among users by diminish-ing the chance of a consistent unfair schedulingstrategy in one single cell. The sharing is im-proved with mobility such that all UEs get nearlyequal service.

Common View: Mobility largely degrades networkthroughput and higher mobility level deteriorates moreWe have found out that throughput performancedoes not monotonically decrease with increasedmobility level. By dividing throughputs intosmaller groups with interval of 10 km/h pergroup, we observed increments in throughputperformance for some higher mobility levels. Fur-thermore, our data exhibited that moving speedat or below 100km/h only slightly affects HSPAthroughput performance at end users devices,and the end users can still sustain their commu-nications with reasonable throughput.

Common View: Mobility affects all ows equally. AndTCP ows suffer more than UDP onesDue to TCPs rigid control strategy, TCP ows arein great disadvantage competing for bandwidthwith UDP ones. Intuitively TCP ows shouldsee worse service provisioning with more packetlosses and larger delays caused by mobility. How-ever, the analysis of our experiments data revealsthat TCP ows unexpectedly see much betterperformance during mobility than UDP ows. Inmobile scenarios, the round trip time (RTT) ofTCP ows, as well as its deterioration due tomobility, is far less than UDP ones. It compen-sates for TCPs disadvantage in competing withUDP trafc for network access, especially whenthe network trafc is very heavy.

Common View: Handoffs are triggered in the tran-sitional region between cells and always result in abetter wireless connectionWe have found out that the triggering and thenal result of handoffs are often unpredictable.Sudden deep fading can force a UE to makea premature handoff. And in nearly 30% of allhandoffs, selection of a base station with poorersignal quality can be witnessed. This discoverycalls for fast reacting fallover mechanisms, whenUEs in HSPA networks have underwent an un-successful handoff.

In short, we have concluded that mobility has bothfavorable and unfavorable impacts on the perfor-mance of HSPA, just like a double-edged sword. Inorder to gain better understanding and control of thisblade, we have conducted in-depth research to furnish


3analysis and explanations to what we have observedlater in this paper, in hope of exposing the fundamen-tal tradeoffs in HSPA network implementation.

To the best of our knowledge, this is the rstpublic report on a large scale empirical study on theperformance of commercial mobile HSPA networks.Our experimental results are useful for network plan-ners/designers. They can plan better cell coverage toalleviate harmful transition from an already poor toa poorer cell. Moreover, they can improve networkcapability to address users demands while they aremobile. Furthermore, they can change their networkcontrol policies to make themselves more competi-tive on the market. Our ndings are also helpful toapplication developers in designing mobile bufferingalgorithm and uplink ow control to eliminate sideeffects of handoff. Application developers can use amore aggressive algorithm for mobile users to betterutilize network resources based on the fact that mo-bility improves the fairness of bandwidth sharing.

The remainder of this paper is organized as follows.We conduct a survey of related work in Section 2.In Section 3, we give a brief introduction of HSPAtechnologies. In Section 4, we describe our measure-ment methodology. In Section 5, 6 and 7, our experi-mental observations and analysis are presented. AndSection 8 concludes the paper.

2 RELATED WORKMany theoretical studies which are more usefulfor preliminary capacity approximation and networkplanning purposes are proposed in prior works. Theseworks focus on HSDPA scheduling or performanceevaluation by simulations [4] [19] [20] [26] [27]. How-ever in real 3G and HSPA networks, the theoreticalmodel is hard to formulate, especially in mobile en-vironments.

There have also been quite a number of eld mea-surement studies on operational 3G networks, Cano-Garcia [8] and Pentikousis [22] mainly focus on thepacket delay behaviour and goodput of pure data traf-c under lightly-loaded scenarios respectively. AndTan [24] measures the data throughput and latencyof live 3G (WCDMA) networks under saturated con-ditions, using a mixture of data, video and voicetrafc, but which is just in stationary environmentsunder WCDMA. Liu [17] presents an experimentalcharacterization of the physical and MAC layers inCDMA 1xEV-DO and their impact on transport layerperformance.

Derksen [10] presents the results of HSDPA mea-surements made in live, commercial networks sup-plied by Ericsson. But the paper just gives an aver-age downlink throughput in mobile vehicle and notanalyzes the factors impact on the performance. Lit-jens [16] presents a ow level performance evaluationof data transfer in a UMTS/HSDPA network with a

principal focus on the performance impact of terminalmobility. But the experiments are carried out in anexperimental setup and small scale in a cell. Arjona [5]gives an empirical study towards high quality VoIPin 3G and HSPA Networks. Arjona [6] evaluatesthe VoIP performance of the HSDPA netowrk andthe proprietary FLASH-OFDM network in Finland.Kara [15], Ruser [23] and Banitsas [7] evaluate theperformance of video streaming in HSDPA networksin simple scenarios. Yao [28] gives an empirical studyof bandwidth predictability in the mobile environ-ments. The authors repeat trips along a 23km route inSydney under typical driving conditions and measurebandwidth from two independent cellular providersimplementing the popular HSDPA technology in twodifferent peak access rates (1.8 and 3.6 Mbps). Butthey only investigate the bandwidth and examinedownload trafcs.3 HSPA BACKGROUNDHSPA networks are now in operation in many re-gions and countries around the world. It is commonlyreferred to as 3.5G and comprises two components,namely High Speed Downlink Packet Access (HS-DPA) and High Speed Uplink Packet Access (HSUPA).HSDPA was standardized as part of 3GPP Release 5in 2002. HSUPA then was included as part of 3GPPrelease 6 in 2004. HSPA is deployed on top of theWCDMA network so that it shares all the networkelements in the core network and in the radio net-work. The upgrade from WCDMA to HSPA requiresonly new software packages and some new pieces ofhardware in the base station to support higher datarate and capacity.

Different from WCDMA, HSDPA improves down-link performance using adaptive modulation and cod-ing (AMC), fast packet scheduling at the base station,and fast retransmissions from the base station. A cellequipped with HSDPA can transmit up to 15 parallelHigh Speed Downlink Shared Channels (HS-DSCHs).All HS-DSCHs operate with a xed spreading factor

Fig. 1: Measurement traces in Hong Kong


4Fig. 2: Measurement Setup

of 16. Modulation scheme and error protection over-head are adapted such that in good radio conditionsthe useful bit rate is very close to the physical layerrates for each HS-DSCH at 480 kbps using QPSK and960 kbps using 16QAM. The theoretical maximum 14Mbps HSDPA bit rate is derived from the simulta-neous and continuous allocation of all 15 channelsto one User Equipment operating under ideal radioconditions [3].

Although their names are similar, HSDPA andHSUPA operate in different way. HSDPA is based onthe rapid statistical multiplexing of a set of xed-rate downlink common channels on a cell, whereasHSUPA, more correctly called enhanced dedicatedchannel (E-DCH), is based on the enhancement ofdedicated channels rather than of common chan-nels (as is the case for HSDPA). Nevertheless, thereare similarities between the systems. Like HSDPA,HSUPA also achieves higher data rates, primarilythrough rapid scheduling of resources, fast retrans-missions and channel adaptation, but adds the pos-sibility of using lower spreading factors. HSUPA the-oretically has uplink speeds up to 5.76 Mbps underideal radio conditions [3].

4 MEASUREMENT METHODOLOGYWe have performed our measurements over two ma-jor commercial HSPA networks in Hong Kong. In thissection, we introduce the measurement routes, setupas well as metrics in details.

4.1 Measurement Routes

In our measurements, we have covered virtually allpossible mobile scenarios in urban areas spreadingfrom land to sea and surface ground to underground,

TABLE 1: Adopted bit rate and packet size

Packet Size Bit Rate

256Bytes 64 kbps 512 kbps512Bytes 128 kbps 1024 kbps1024Bytes 256 kbps 2048 kbps

including subways, trains, city buses, ferries and self-driving vehicles. The trace map is shown in Fig. 1.These kinds of mobile transport systems have differ-ent velocities, directions, traces and mostly cover allthe popular places in Hong Kong. The characteristicsof different types of transportation are as follow:

Trains - They are on surface ground. Stations aresparsely distributed so that sometimes the trains aremoving with speed up to 100kmh. They have averagemoving speed of 40kmh. We select East Line forcarrying out the experiments is simply because it isthe busiest line among train lines.

Subways - They are in underground. Stations aredensely distributed so that the average moving speedis about 30kmh and the highest speed is 80kmh. Wemost frequently use Tsuen Wan line because it consistsof both surface ground and underground stations, alsoa long harbor crossing tunnel.

Self-driving Vehicles & Buses - They are on surfaceground too and move around open urban areas. Theaverage moving speeds are about 50kmh and 30kmhrespectively. We choose a bus route that cover urbanroads and a highway, and then we follow the sameroute using a self-driving vehicle.

Ferries - They are of course in the sea and it has thehighest average moving speed of 80kmh among fourtypes of listed transportation. The ferry line we selectis a regular transportation connecting Hong Kong andMacau.

Our main data come from train and subway lines,since they feature stable speed and constant traces.Measurements acquired on buses, ferries are used ascomplementary data and hence the reported resultsare combination of above mentioned transport modes.The speeds of railway (subways & trains) , buses andoff-shore ferries average respectively 40kmh, 30kmhand 70kmh.

4.2 Measurement SetupOur measurement setup is shown in Fig. 2. We useda set of client-and-server programs to conduct suchmobility experiments. The servers are PCs runningLinux with Kernel 2.6.20. Two identical servers areseparately hosted in City University of Hong Kongand a commercial data center to ensure the resultswill not be affected or biased by anything specicto one server and its path. We have also developeda set of data collection programs in C++ and Java,and data analyzing tools in C++ and MATLAB script.Our UEs include four laptops (Linux Kernel 2.6.20)equipped with four MC950D HSPA modems, twoHTC G1 Smartphones, and two iPhone 3Gs.


50 500 1000 1500 2000 2500 3000 35000

0.2

0.4

0.6

0.8

1

Throughput (kbps)

CD

F

StationarySubwayFerryTrainBus

(a) Downlink of Operator A

0 200 400 600 800 1000 1200 14000

0.2

0.4

0.6

0.8

1

CD

F

Throughput (kbps)


(b) Uplink of Operator A

0 500 1000 1500 2000 2500 3000 3500 40000

0.2

0.4

0.6

0.8

1

Throughput (kbps)

CD

F


(c) Downlink of Operator B

0 200 400 600 800 1000 1200 14000

0.2

0.4

0.6

0.8

1

Throughput (kbps)

CD

F


(d) Uplink of Operator B

Fig. 3: CDF of HSPA Network Throughput

We have conducted three types of evaluations:download only, upload only and simultaneous down-load & upload. These three types can almost coverall kinds of our frequently used applications. Down-loading communications typically occur with FTPtransmissions and website browsing, while uploadingcommunications are closely associated with le shar-ing or P2P services. For simultaneous download &upload communications, bidirectional sessions, suchas video calling/conferencing and online gaming, arethe most typical. These tests were done by sendingbursts of data packets back-to-back from UEs to theservers via HSPA links and measuring throughputbased on the inter-arrival times between packets ina burst. Table 1 illustrates the bit rates and packetsizes adopted in our experiments using both TCP andUDP. Servers and laptops use CUBIC variant of TCPimplementation by default and HTC G1 Smartphonesuse TCP reno. However, TCP implementation foriPhone 3G is unknown to us but this should not affectour measurement result in general. We purposelychoose different bit rates for emulating different typesof forthcoming services including realtime light andbulky UDP data calls and non realtime light andbulky TCP data calls. Different packets are also usedin order to nd out the network behaviors due tovarious packet sizes. For example, 2048 kbps through-put has been intentionally used to maximize the cellsthroughput so that we can determine the behavior ofbandwidth sharing under saturated network.

We used throughput, round trip time (RTT) and lossrate of UDP trafc as primary metrics for compari-son throughout this study because they are directlyrelated with QoS of trafc sessions. Throughput isusually the rst users impression for perceiving thenetwork performance since HSPA is usually claimed,but always fall short of, as mobile broadband access.But is it a sufcient indicator of the actual networkperformance? UDPs connectionless mechanism al-lows constant bit rate regardless of actual networkcondition so that its throughput is highly biasedin comparison with that of TCP. When network isexperiencing congestion, TCPs reliable transmissionmechanism will refrain the terminal from sending toomuch to the congested network. In comparison, UDPkeeps sending at original speed leading to potentiallymany packet losses. Therefore, when users want tolearn more, RTT, which is referred to as applicationlayer level round trip time, is often used to estimatethe degree of network congestion. Less congested(smaller RTT) network path is capable of providingbetter QoS for certain types of services, e.g. VoIP.Along with the metrics, we have also logged the CellID, as well as the Ec/Io of the UEs, queried via ATcommands, while the measurements were being con-ducted. Thus we can establish relationship betweenthe HSPA performance data with the physical locationand the cell information. Through cell information,many impacts of mobility on performance, such ashandoffs between cell towers and signal uctuation,


6Very Good Good Medium Poor Very Poor No Signal0

0.1

0.2

0.3

0.4

0.5

0.6

0.7P

DF

SubwayTrainFerryBus

(a) Ec/Io of Operator A

Very Good Good Medium Poor Very Poor No Signal0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

PD

F

SubwayTrainFerryBus

(b) Ec/Io of Operator B

Fig. 4: PDF of Ec/Io for Mobile HSPA Network

can be observed more clearly. So it is possible for us toevaluate the impact level of different mobility-relatedfeatures. The above method is not only applied tonetwork saturation measurements, but also appliedto throughput measurements under different trafcmodes.

At the beginning of measurement, we calibrated ourGPS receiver for logging the moving speed and tracksof the transport. We then checked the associated CellID of all UEs to make sure they are connected to thesame cell. During the measurement, all UEs sat nextto one another to ensure they have identical, at leastvery similar, test environment. For each test journey,half of UEs ran TCP tests and the other half ran UDPtests at the same packet size and bit rate setup. Alltests have been carried out at off-peak hours so as toalleviate the impact of background trafc to the testsand vice versa.

5 GENERAL IMPACT OF MOBILITYMobile HSPA performance is very different from staticHSPA. As was stated in Section 1.2, we record a largespread of HSDPA bit rates (Fig. 3) and signal quality(Fig. 4) because the HSPA UEs pass through a varietyof dynamic and unpredictable radio conditions alongthe different mobility routes.

Fig. 3 depicts CDF of mobile HSPA throughputunder various mobile scenarios for both OperatorA and Operator B. It manifests that the maximum

achievable bit rate for downlink is 3500 kbps and1400 kbps for uplink. It also shows that stationaryUEs have throughput falling into a certain short rangewhile mobile UEs have wide spread of throughput.Nonetheless, the recorded throughput for mobile UEsmainly falls into less than 2000 kbps in downlink andless than 600 kbps in uplink, which are far lowerthan those of static environments. Particularly, thereare incredibly high, about 10% to 20% of the time,UEs would experience no throughput at all in allmobile environments compared with less than 4% instationary environments.

The signal quality for different transport means areshown in Fig. 4. We nd that UEs in all mobilescenarios have signal quality spreading from -1 dBm(very good) to -32 dBm (no signal) but rarely havesignal quality spreading range more than 5 dBm instationary environments.

One point to mention is that we are aware thatnormally the signal quality is often measured by aver-age HS-DSCH signal-to-interference-plus-noise ratio(SINR) [12] as shown in Equation 1:

SINR = SF16PHSDSCH

(1 )Ior + Ioc +N0 (1)

where SF16 is the HS PDSCH spreading factorof 16, PHSDSCH is the received power summingover all active physical shared channel codes, Ior isthe received own-cell interference, is the downlinkorthogonality factor, Ioc is the received other cell inter-ference, and N0 is the received noise power. Neverthe-less, we employ another commonly used metric, theratio of received pilot energy over-the-air (It is oftenused to check coverage levels. It should be highestnear the tower, declining to a minimum level at thehandoff point.), Ec, to total received energy or thetotal power spectral density, Io, which is essentiallyequal to Ior + Ioc +N0. So, as an alternative to SINRwe employ this metric in our paper to indicate thequality of the received signal in the cellular network.

Interestingly, CDF of mobile HSPA throughput(Fig. 3) and PDF of signal quality (Fig. 4) turn out toconverge on one common phenomenon. Both guresshow that subways conspicuously have the best per-formance among all selected mobile scenarios. Mean-while, all other open area mobilities surprisingly havevery close (i.e. similar performance) CDF of mobileHSPA throughput and PDF of signal quality curves.Fig. 3 illustrates that overall throughput for UEs insubways apparently outperform those in other mo-bility situations as subways has the closest CDF withstationary one. On the other hand, despite the fact thatopen space mobilities such as trains, buses and ferrieshave larger differences with stationary one, they sur-prisingly have nearly overlapped CDF of throughputin both operators network. More interestingly, PDFof mobility signal quality matches strongly with CDF


70 110 1120 2130 3140 4150 5160 6170 7180 8190911000

200

400

600

800

1000

Speed (km/h)

Ave

rage

Thr

ough

put (

kbps

)

Operator AOperator B

(a) Mobility Level Impact on Downlink Throughput

0 110 1120 2130 3140 4150 5160 6170 7180 8190911000

100

200

300

400

Speed (km/h)

Ave

rage

Thr

ough

put (

kbps

)


(b) Mobility Level Impact on Uplink Throughput

Fig. 5: Average Throughput for Different Mobility Velocities

of mobile HSPA throughput. Among them, as shownin Fig. 4, subways has the highest good signal qualitycoverage of 92% of the journey compared with only76%, 68% and 58% (fairly close) on trains, buses andferries respectively for Operator A. Likewise in Oper-ator B, subways also have 85% good signal coveragecompared with 67%, 72% as well as 63% (very close)on trains, buses and ferries respectively.

The similarity among CDF of mobile HSPAthroughput and PDF signal quality evidently providesus a hint that there exists a fundamental difference be-tween of underground and surface ground network.The distinct difference is due to different transmissionmedia in use. Open areas are mainly covered bynormal cell radio as electromagnetic (EM) wave canpropagate efciently from cell tower to end user UEsat sufciently low cost. Nevertheless, all subwaysstations are covered by leaky cables in tubes becauseEM wave can not be efciently transmitted throughtunnels. Thus, the leaky cables, though the cost ishigher, are purposely used to guide and radiate EMwave from station into the tubes to avoid under-ground communication blind spot and also to providea highly optimized underground radio coverage.

It is commonly expected that cellular throughputperformance decreases with increased mobility be-cause cellular throughput performance is affected bycell reselection and routing area update procedures,particularly for fast-moving UEs. Increasing mobil-ity levels will lead to a higher number of cell re-selections and routing area updates, and will alsoincrease the amount of packet outage time. Our dataappear to align with this common view rstly becausethe data has demonstrated that mobility deterioratesHSPA throughput performance on average. However,in above context our data have also pointed outdegradation of signal quality, Ec/Io, is the majorcause of throughput performance degradation, mean-ing that mobility degrades signal quality and thusthe throughput performance. Therefore, there is in-adequate evidence for telling whether mobility level,i.e. velocity, plays a signicant role in such throughput

degradation in HSPA networks.To justify the mobility levels impact on the

throughput performance at user application level, wehave torn down the throughput of one typical trainstrace, for into different mobility levels with interval of10 km/h for each level, as it is shown in Fig. 5. Weseparated throughput at static scenario, i.e. when trainstops at stations, for comparison. The gure clearlyshows that, in terms of end users mobile devices,mobility level does not affect average throughputperformance because the averaged throughput doesnot monotonically decrease with increased mobilitylevel. For example in Fig. 5a, downlink throughputat 40 km/h is apparently better than that at 30 km/h(similar phenomenon can also be observed in Fig. 5b).We have also noticed in the gure that throughputat 100 km/h for Operator B is comparable to whentrain stays static. During the whole journey, trains canonly reach at 100 km/h or above when they travel inbetween of two distantly separated stations (Fan Lin Tai Wo and Tai Po Market University) whichresults in fewer samples being collected for movingspeed over 100 km/h. Moreover, while train is com-muting in between of these stations, UEs generallyhave better signal quality as the radio condition ismore outstanding and stable in open areas. These arethe reasons for seeing comparably good throughputperformance at 100 km/h group.

Although increased mobility level leads to in-creased number of cell reselections, network design-ers have mitigated this problem using HierarchicalCell Structure (HCS). In this structure, large cellsguarantee continuous coverage, while small cells arenecessary to achieve good spectrum efciency. Smallrange cells are used by low mobility UEs, whilehigh range cells serve high mobility UEs. Hence thisscheme effectively reduces service terminations ratecaused by frequent handoffs during UEs mobility.Meanwhile, since HSPA bit rate can adapt quicklyto the different radio conditions (Transmission TimeInterval, TTI 2ms), even in mobility the rapid changeof environment leads to instantaneous poorer radio


8

(a) Static HSPA bandwidth sharing for Operator A

(b) Static HSPA bandwidth sharing for Operator B

(c) Mobile HSPA bandwidth sharing for Operator A

(d) Mobile HSPA bandwidth sharing for Operator B

Fig. 6: Mobility leads to fairer bandwidth sharing.

condition, the throughput rebounds immediately innext scheduling cycle if the UE gets scheduled andit sees better signal quality. To this extent, data rateduring mobility is only slightly affected by mobilitylevel but is dominated by signal quality those UEsexperienced at the moment.

In most cases, stationary users deliberately choosespots where their devices can retain good receptionsignal. That means stationary UEs have better averagesignal quality which leads to higher average through-put because they are less vulnerable to sudden degra-dation of signal quality as compared with mobilityscenarios.

6 MOBILITY IMPACT ON BANDWIDTH SHAR-INGMobility hurts the general performance of HSPA net-works. However, our experiment data reveal that insome cases, mobility does help some aspects of thenetworking functionalities in HSPA networks. In thissection, we explore the useful side of this two-edgedsword.

6.1 Bandwidth Sharing among UsersBandwidth sharing is more discussed together withscheduling or call admission. It is not straightforwardto link bandwidth sharing directly with mobility. Peo-ple may think that the dynamic and unpredictable

mobile environment will further aggravate the unfair-ness of bandwidth allocation among users. However,this is not true. In fact, we have observed that thebandwidth sharing is not as fair as expected in sta-tionary HSPA environment, while mobility actuallyimproves the fairness of bandwidth sharing amongusers, though it also introduces a plenty of side effectson link performance.

In stationary environment, we observed that a sin-gle UE is more likely to occupy most of availablebandwidth and leave only a little portion for otherUEs. For example, in Fig. 6a, UE2 has throughputat about 2Mbps during 60% of all the time, butUE1, within the same time period, has throughputaveragely distributed from 1.3Mbps to 2Mbps. In thestatic scenario of Fig. 6, the bandwidth allocation isnot fair both in the uplink and downlink directions.The behavior somehow violates HSPA bandwidthallocation mechanism, because if UEs were placedin a similar environment, each of them should havehad the equal chance to get served by Node B. Onthe contrary, The improvement can be seen from theFig. 6b that all UEs have no apparent advantages overone another when moving, though sometime one ofthem would have no granted bandwidth, but otherUEs had equal chance to have the highest bandwidthgranted.

There are some scheduling algorithms can be ap-


90 200 400 600 800 1000 1200 1400 1600 1800 20000

2

4

6

8

10

12

14x 104


(a) Cross Correlation of Ec/Io for static scenario

0 200 400 600 800 1000 1200 1400 1600 1800 20000

1

2

3

4

5

6

7

8

9x 104


(b) Cross Correlation of Ec/Io for mobile scenario

Fig. 7: Cross Correlation of Ec/Io.

plied to HSPA (both downlink and uplink) accordingto design requirements.

Scheduling the user with the best instantaneousradio link conditions is often referred to as max-C/Ischeduling [9]. This scheduling algorithm maximizesystem capacity because the scheduled one will typ-ically have high data rate. Mathematically, the max-C/I scheduler can be expressed as scheduling user kgiven by

k = argmaxnRn (2)

where Ri is the instantaneous data rate for user n.Obviously, max-C/I scheduling is benecial for cellthroughput but it in turn results in unfair schedulingamong UEs. Assuming there is a UE locating faraway from the base station or bad radio environ-ments, apparently the UE can hardly get scheduledas it could never have best instantaneous C/I amongpeers, starving for getting scheduled.

Round-robin schedules users turn by turn withoutconsidering the instantaneous channel conditions. Al-though round-robin strikes perfect fairness among allusers, it also biases UEs seeing good channel qualityleading to lower overall system throughput.

Therefore, a scheduler which takes advantages ofboth round-robin and max-C/I is more desirable. Inother words, the scheduler should utilize fast vari-ations in channel conditions as much as possiblewhile still satisfying some degree of fairness [9].Consequently, a proportional-fair scheduler is initiallyproposed in [13] [14]. Mathematically, the user isscheduled according to

k = argmaxnRn

Rn(3)

where Rn is the instantaneous bit rate for user n andRn is the average bit rate for user n. The average iscalculated over a certain averaging period TPF whichis typically set to be in the order of one second.Simply put, this scheduler selects users according to

their instantaneous channel quality relative to aver-age channel conditions. As a result, users with bestinstantaneous are not necessarily privileged.

So which scheduler is being used in examinedHSPA networks? Our data initially point to max-C/Ias we have seen large degree of throughput unfair-ness among static testing UEs. Nevertheless, max-C/Istarves UEs with poor C/I, and this is not the casebecause we have witness that UEs having poor C/Ican also get served by Node-B, of course, with poorthroughput as well. Hence, it is highly possible thatproportional-fair is in use as it compromises systemthroughput and fairness.

Assuming PF is in services, the numerator Rn ofEquation 3 changes rapidly by chance in mobilitywhich means each UE should have equal chance tohave high or low Rn. As we know that randomness re-sults in fair allocation of network resources only whentime limitations are not considered. In PF schedulingalgorithm TPF is typically set to be in the order of onesecond which is remarkably long compared with 2msTTI of HSPA networks. Therefore it is highly possiblethat fairer level of bandwidth sharing can be archivedin mobility. Moreover, the scheduling algorithm is ofbase station to base station basis, so that the decisionsmade by one base station would not propagate to nextcell for the cell change due to handoff, that meansdecision maker changes from time to time and thuseventually results in fairer decisions.

We have calculated the auto-correlations and cross-correlations of the signal quality for all UEs in aperiod of over 20 minutes to see similarity of sig-nal quality between UEs. Mathematically, the cross-correlation, also called covariance, is dened as:

(f g)[n] =

m=f[m] g[n+m]. (4)

Cross-correlation is a measure of similarity betweentwo signal sequences. It is a function of time. If asimilar pattern occurs in the two sequences, a spike


10

0 0.5 1 1.5 2 2.50

0.2

0.4

0.6

0.8

1

RTT (s)

CD

F

StaticMobile

(a) TCP RTT for Operator A

0 0.5 1 1.5 2 2.5 3 3.5 40

0.2

0.4

0.6

0.8

1

RTT (s)

CD

F

StaticMobile

(b) TCP RTT for Operator B

0 10 20 30 40 50 60 70 800

0.2

0.4

0.6

0.8

1

RTT (s)

CD

F

StaticMobile

(c) UDP RTT for Operator A

0 1 2 3 4 5 60

0.2

0.4

0.6

0.8

1

RTT (s)

CD

F

StaticMobile

(d) UDP RTT for Operator B

Fig. 8: Static Roundtrip time for TCP and UDP connections in two operators networks

will appear in the cross-correlation function. For twoconstant signal sequences, their cross-correlation willturn out to be a triangle. The farther the cross-correlation shape deviates from the perfect triangle,the larger difference and irregularity the two signalsequences display.

Autocorrelation is a special case of the cross-correlation. It correlates the signal sequence with it-self. If a repeated pattern keeps occurring, a spikewill appear at the time point of period, as well asits multiples, in the auto-correlation function.

R(s, t) =E[(Xt t)(Xs s)]

ts(5)

Eq. 5 is a probabilistic version of normalizedauto-correlation. As we do not know the exact formof the signal, we can only correlate it probabilistically,and the variances in the numerator are the normaliz-ing factor.

The results illustrated in Fig. 7 show that whenstatic, both UEs see almost the same signal qualityover the time span and the cross-correlations betweenUE1 and UE2 is a nearly perfect triangle, whichmeans that their relative value does not change.However, in mobility, there are much noises forall auto-correlations and cross-correlations gures.Auto-correlations normally see a 5% 20% uctua-tion compared to their static counterpart, and cross-correlations a 10%30% uctuation. The results imply

that the radio environment is changing very fast, soan UE can hardly keep its dominancy in a cell dueto unstable radio signals and ensuing failed transmis-sions.

6.2 Bandwidth Sharing among Trafc FlowsThere are different types of trafc ows in HSPAnetworks, e.g., TCP and UDP ows. As mobility intro-duces dynamic changes to the fundamental physicalenvironment of end users, all these ows are expectedto suffer deterioration caused by mobility. But do allthe ows experience the same inuence?

At rst sight, mobilitys impact will not differen-tiate between high level trafc, e.g., transport layerows. However, our data point in another direction.Through studying the RTT distribution of TCP/UDPows, we have found unexpectedly that TCP owssee better performance during mobility. It compen-sates for TCPs disadvantage in competing with UDPtrafc, especially in heavy load scenarios. From Fig.8, we can see all mobile CDF curves lie to the leftof static CDF curves. Though the RTT improvementis marginal, yet it still contradicts the common sensethat unstable mobile wireless environment will slowdown all ow trafc. In order to account for this, weneed to look at the other major category of ows inthe network, the UDP ows.

First point to observe regarding UDP ows is thatthe UDP packets return much slower than TCP pack-ets. From Fig. 8, we can see that almost all TCP


11

packets can return within 2.5 seconds. But for UDPpackets, it takes at least 5 seconds for 95% of themto return. In extreme mobile cases, the RTT of somepackets can even reach 50 seconds. This is not thecase in traditional static networks, as TCP imposesstricter acknowledgement and ow control strategy.We gure out that, besides scheduling difference, theow control functionality of TCP protocol helps toprevent too much trafc pouring in the network,thus avoiding relatively long delays, as experienced inUDP trafc. In Fig. 9, we plot the average TCP trafcand UDP trafc pumped into the wireless channel.The gure shows TCP trafc is much constrainedand adaptive to the channel condition, while UDPtrafc keeps pumping almost the same amount of dataregardless of the channel condition, resulting in muchpoorer performance of UDF trafc than TCP trafc.Our second observation regarding UDP ows is thatthe UDP RTT degrades signicantly in the mobilescenarios. In Operator As network, RTT of over 15%UDP packets exceeds 40 seconds. In Operator Bsnetwork, the average mobile RTT is nearly 1 secondslonger than their static counterparts.

In a wireless network, TCP and UDP trafc coexist.But due to the unconstrained nature of UDP, UDPows normally eat up the majority of the bandwidth,if the total volume is high. So TCP ows are in agreat disadvantage when competing for bandwidth.However, in mobile scenarios, the efciency of UDPows dramatically deteriorates, leaving more spacefor TCP ows to get in. This enhances the fairnessin bandwidth sharing between TCP and UDP owsin HSPA networks. We also notice that TCP trafcis normally related to webpage browsing and corre-sponds to the Interactive QoS class, where the trafcamount is not large and accurate transmission is vital.UDP trafc is more tailored for real time services,such as streaming and conversational, where certainpercentage of packet loss can be well tolerated. Sogiven limited bandwidth in HSPA networks, it isreasonable to guarantee the prioritized handling ofTCP ows.

Finally, it is also worth mentioning that the RTTgraphs disclose the scheduling strategies employedby the operators regarding TCP and UDP trafc.With mobility, the RTTs vary more dramatically, andexhausting all possible scheduling strategies appliedto different round trip lag time. We made our conjec-tures on the Node-B and provider scheduling strategybased on the following two points: 1, we deliberatelyselect off-peak time to carry out our measurement, soit is reasonable to assume that there is no other trafcfor our connected base station in most of the time; 2,we believe that the RTT statistics, which are based ona large amount of samples, can fairly represent thebuffer time at the base station.

In general, there are four types of QoS classes inHSPA networks, i.e., conversational, streaming, inter-

0 500 1000 1500 2000 2500 3000 35000

0.2

0.4

0.6

0.8

1

Bit Rate

CD

F

TCP Bit RateSTD of TCPUDP Bit RateSTD of UDP

Fig. 9: CDF of TCP & UDP sending bit rate in one test.TCP is more constrained and adaptive, while UDPbit rate is almost constant. STD stands for StandardDeviation.

0 50 100 150 2000

500

1000

1500

2000

2500

3000

Time (s)

Thro

ughp

ut (K

bps)

(a) Instantaneous Sample of TCP Throughput

7LPHV

(F,RG%P

(b) Instantaneous Sample of Ec/Io

Fig. 10: Impact of handoff on mobile HSDPA through-put.

active and background. Different scheduling strategyis applied to these classes. In each scheduling strategy,a utility function Un(rn) is dened, where n denotes aparticular HSDPA user and rn is the average through-put for the nth user. The objective of the schedulingalgorithm is to nd a selection n, such that

n = argmaxn{Mn}, where Mn = dn Un(rn)rn

(6)

Here, Mn denotes metric, dn is the instantaneousdata rate that HSDPA user number n can supportin the next TTI. The RTT CDF of TCP trafc inFig. 8 complies with the features of proportional fairqueue scheduling [21]. We can write the metric andthe utility functions as Un = log(rn), Mn = dn/rn.


12

0 200 400 600 800 1000 1200 1400 1600 1800 20000

0.2

0.4

0.6

0.8

1

Average TCP throughput (kbps)

CD

F

Without HandoffsWith Handoffs

Fig. 11: CDF of TCP throughput reduction for hand-offs

The curve grows gradually without abrupt changes.However, for UDP trafc, the service providers applydifferent scheduling algorithms to meet their ownneeds. Operator B is more tolerant towards UDPtrafc, for it schedules these trafc with relativelyhigh efciency. But Operator A differentiates betweendifferent types of UDP trafc. 80% of the trafc canpass within 20 seconds. However, some of the low-priority packets do not get scheduled until after a longtime. But around the 50th second, a modied largestweighted delay rst (M-LWDF) scheduling is likely tobe used. The M-LWDF aims to achieve

Pr(Di > Ti) i (7)where Di indicates the Head of Line (HOL) packet

delay of user i, Ti represents the delay bound, whichis set to around 50 seconds, and i is the allowedpercentage of discarded packets, which is set to a verylow value near zero.

7 MOBILITY IMPACT IN TRANSITIONAL RE-GIONS

Frequent handoffs and rapid change of radio environ-ments, which often happen in transitional regions andalways result in poor data throughput to UEs, usuallyexist in the event of mobility. Handoff is an naturalprocedure when the UE is moving away from the areacovered by one cell and entering the area covered byanother cell, the call is transferred to the second cell inorder to avoid call termination when the phone getsoutside the range of the rst cell.

Although much effort has been paid for guarantee-ing service continuity for handoff since the time ofintroduction of 2G cellular network, the same featureseems not truly available for mobile HSPA data trafc.Fig. 10a shows a instantaneous sample of HSDPA TCPthroughput for a mobile UE. Fig. 10b is about theEc/Io value before and after the changes of cell ID.Despite the fact that handoff can be easily identiedwhen the UE reports the change of cell ID, it cannot be readily seen in Fig. 10b. We understand thatEc/Io values have to be always smaller or equal to0, in order to visualize it more clearly, we show the

0 200 400 600 800 1000 120040

20

0

20

40

Time (s)

EcIo

(dBm

)

(a) Time sample 1

0 200 400 600 800 1000 120040

20

0

20

40

Time (s)

EcIo

(dBm

)(b) Time sample 2

Fig. 12: Typical Ec/Io during handoffs. All Ec/Iovalues are originally negative. For clarity, we ip thevalue to opposite, whenever a handoff happens.

handoff as oscillations between positive and nega-tive values. That means, changing from +ve regionto -ve region (or vice versa) is an indication thatUE left the original cell with a certain Ec/Io valueand now entered another cell with new Ec/Io value(i.e. handoff). Hence, from Fig. 10, we see that TCPthroughput drops sharply when handoff happens. Byanalyzing all logged handoff events from combinationof subway, train, ferry and bus, we also show inFig. 11 the CDF of average TCP throughput withand without handoffs. Fig. 11 intuitively depicts thathandoffs generally induces 50% reduction in through-put since while UEs are experiencing handoff thethroughput often drops and rebounds quickly. How-ever, the peak throughput is only about 30% of thatwithout handoff. At the moment handoff is beingconducted, the link is temporarily lost. Consequentlysome packets are lost or the returning acknowledg-ment packets are delayed. Either case will cause TCPmechanism to prolong its retransmission timer anddecrease its transmission windows size resulting inreduced throughput for longer interval. This situationpersists until the logical link is reestablished andsignal quality rebound to acceptable level. As for UDP,we also observed longer RTT and larger loss, both areat least doubled, during handoff period because UDPdo not employ reliable ow control mechanism likeTCP.

We also observe that UEs undergoing ping-pongeffect have even worse performance than in handoff.The end users may not be aware of the temporaryservice unavailability for delay-tolerant services like


13

!

Fig. 13: CDF of Ec/Io of the original and new bases-tations during handoffs

web browsing but it is apparently noticeable anddestructive to real-time services such as VoIP, videoconferencing.

From our experiments, we nd that the handoffshappening on fast-moving transport vehicles can alsoresult from sudden strong fading or interference. Aswe can see from Fig. 12 that almost every handoffcomes with a large slump of Ec/Io. It is worth notingthat a large proportion, namely around 30% in Fig. 12,of the new base stations after handoffs also have lowEc/Io. It implies that the general wireless condition atthat time is very bad. In fact, none of the base stationcan satisfy the QoS requirement of the UE. However,the handoff still takes place, which does not help a lotunder these situations. So it is recommended that thehandoff functionality be turned off when the UE isvirtually screened or shadowed, until there is a solidchance that it may be connected to a base station withsignal quality good enough.

After handoffs, UE normally connects to a basestation with substantially better signal quality, i.e.,with higher Ec/Io value. However, it is not alwaysthe case. Fig. 13 shows the Ec/Io immediately beforeand after each handoff. From Fig. 13, we can see theEc/Io of the new base stations are statistically betterthan the original base stations by 10dBm, which isa remarkable improvement for better QoS. But as isshown in Fig. 14, almost 30% of all the handoffs donot end up with a better base stations. This contradictsthe basic principle of handoff. We nd two reasonsto justify this phenomena. First, the wireless signalsvaries quickly over time, so that the UE made awrong handoff decision based on outdated Ec/Ioinformation. Second, the handoff decision during highmobility does not depend only on the reading ofEc/Io. Because a fast moving UE may frequently enteror exit different cells with varying trafc volume.So the handoff decision must also take into accountthe consideration of packet scheduling and admissioncontrol. A good base station candidate may reject a UEdue to a bursty high trafc, so the UE must connectwith another poorer base station, as it has left theoriginal cell.

Another interesting nding worth mentioning isthat in underground subways we have far less hand-

EcIo Differences (dBm)

CD

F

Fig. 14: CDF of Ec/Io difference between the originaland new basestations during handoffs

offs than on surface ground trains due to the benetof leaky cable coverage. Our data shows that, for thesame distance journey, number of handoffs recordedin subways is about half of that of on trains. As aresult, end users can have more satisfactory realtimeservice experience on subways.

8 CONCLUSIONWe have designed and performed a large-scale empir-ical study on the mobility performance of two largestcommercial HSPA networks in Hong Kong. We reportour data and discuss the unexpected ndings in HSPAnetwork and their corresponding causes with details.Interestingly, mobility turns out to be a double-edgedsword. Although it generally degrades HSPA services,mobility surprisingly improves some aspects of net-working performance, such as fairness in bandwidthallocation among users and trafc ows. We alsonote that the communication characteristics in HSPAtransitional regions are very complicated, so moreintelligent handoff algorithms are needed for seamlessservice provisioning. As our future work, we planto research into the performance of real networkingapplications on 3G Smartphones. As Smartphones,which are seamlessly bound to 3G wireless networks,are gaining more and more popularity, a lot of net-working problems emerge and call for further studies.

9 ACKNOWLEDGMENTSAbove work was supported in part by grants fromthe Research Grants Council of the Hong Kong SAR,China No. (CityU 114609) and CityU Applied R &D Funding (ARD) Nos. 9681001, 6351006 and CityUStrategic Research Grant No. 7008110; CityU AppliedResearch Grant (ARG) No. 9667052; ShenZhen-HKInnovation Cycle Grant No. ZYB200907080078A andNSF (China) No. 61070222/F020802. The NationalScience Foundation of CHina (NSFC) under grantNo. 61070221, China 973 Project 2011CB302800, theUS National Science Foundation (NSF) under GrantNo. CNS0916584, CAREER Award CCF0546668, andCNS1065136. Any opinions, ndings, conclusions,and recommendations in this paper are those of theauthors and do not necessarily reect the views of thefunding agencies.


14

REFERENCES[1] http://www.gov.hk/en/about/abouthk/factsheets

/docs/telecommunications.pdf.[2] http://www.gsmworld.com/technology/hspa.htm.[3] HSPA technical update. http://www.wraycastleusa.com/news.php?

id=10, 9, Sep 2008.[4] K. Aho, J. Aijanen, and T. Ristaniemi. Impact of mobility to the

VoIP over HSUPA system level performance. pages 20912095,May 2008.

[5] A. Arjona, J. Kerttula, and A. Yla-Jaaski. Live network perfor-mance challenge: FLASH-OFDM vs. HSDPA. pages 918925,March 2008.

[6] A. Arjona, C. Westphal, A. Yla-Jaaski, and M. Kristensson.Towards high quality VoIP in 3G networks - an empiricalstudy. pages 143150, June 2008.

[7] K. Banitsas, S. Tachakra, E. Stefanidis, and K. Boletis. UsingHSPA to improve the telemedical links on a moving ambu-lance. pages 739742, Aug. 2008.

[8] J. M. Cano-Garcia, E. Gonzalez-Parada, and E. Casilari. Exper-imental analysis and characterization of packet delay in UMTSnetworks. volume 4003 of Lecture Notes in Computer Science,pages 396407. Springer, 2006.

[9] E. Dahlman, S. Parkvall, J. Skold, and P. Beming. 3G EvolutionHSPA and LTE for Mobile Broadband. Academic Press, 2007.

[10] J. Derksen, R. Jansen, M. Maijala, and E. Westerberg. HSDPAperformance and evolution. Ericsson Review, (3):117120, 2006.

[11] F. Fitzek, A. Kopsel, A. Wolisz, M. Krishnam, and M. Reisslein.Providing application-level QoS in 3G/4G wireless systems: Acomprehensive framework based on multirate CDMA. IEEEWireless Communications, 9(2):4247, April 2002.

[12] H. Holma and A. Toskala. HSDPA/HSUPA for UMTS. JohnWiley & Sons Ltd, 2006.

[13] J. Holtzman. Cdma forward link waterlling power control.volume 3, pages 16631667, Sept 2000.

[14] J. Holtzman. Asymptotic analysis of proportional fair algo-rithm. volume 2, pages F33 F37, Sep/Oct 2001.

[15] N. Kara and V. Planat. Performance analysis of IP multimediaservices over HSDPA mobile networks. pages 15, Dec. 2007.

[16] R. Litjens. HSDPA ow level performance and the impact ofterminal mobility. volume 3, pages 16571663 Vol. 3, March2005.

[17] X. Liu, A. Sridharan, S. Machiraju, M. Seshadri, and H. Zang.Experiences in a 3G network: interplay between the wirelesschannel and applications. In MOBICOM, pages 211222, 2008.

[18] Livecast. http://www.livecast.com/corporate/index.html.[19] M. Lopez-Bentez, F. Bernardo, N. Vucevic, and A. Umbert.

Real-time HSPA emulator for end-to-edge QoS evaluationin all-ip beyond 3G heterogeneous wireless networks. InSimuTools, page 8, 2008.

[20] P. Lunden, J. Aijanen, K. Aho, and T. Ristaniemi. Performanceof VoIP over HSDPA in mobility scenarios. pages 20462050,May 2008.

[21] M. Malkowski, A. Kemper, and X. Wang. Performance ofscheduling algorithms for HSDPA. pages 1052 1056, Aug2007.

[22] K. Pentikousis, M. Palola, M. Jurvansuu, and P. Perala. Activegoodput measurements from a public 3G/UMTS network.Communications Letters, IEEE, 9(9):802804, Sep 2005.

[23] H. Ruser, P. Halvorsen, C. Griwodz, and B. Hestnes. Perfor-mance measurements and evaluation of video streaming inHSDPA networks with 16QAM modulation. pages 489492,23 2008-April 26 2008.

[24] W. L. Tan, F. Lam, and W. C. Lau. An empirical study on thecapacity and performance of 3G networks. IEEE Transactionson Mobile Computing, 7(6):737750, June 2008.

[25] Veodia. http://veedia.com/index.php.[26] V. Vukadinovic and G. Karlsson. Video streaming in 3.5G: On

throughput-delay performance of proportional fair schedul-ing. pages 393400, Sept. 2006.

[27] X. Yan, J. Khan, and B. Jones. Impact of a radio access networkcapacity on the HSDPA link performance. pages 220224, 302007-Oct. 3 2007.

[28] J. Yao, S. S. Kanhere, and M. Hassan. An empirical study ofbandwidth predictability in mobile computing. In WINTECH,pages 1118, 2008.

[29] YouTube. http://www.youtube.com.

Fung Po Tso is currently a SICSA re-search fellow in the School of Computing Sci-ence, University of Glasgow, UK. He receivedhis BEng degree in Computer Engineering,MPhil and Ph.D degrees in Computer Sci-ence from City University of Hong Kong(CityU HK), in 2005, 2007 and 2011 respec-tively. His research interests include clouddata center networks, mobile computing, dis-tributed computing and cyber-physical sys-tem.

Jin Teng is now in the third year of his PhDprogram with the Department of ComputerScience and Engineering at the Ohio StateUniversity. He received his B.S and M.S de-grees in Electronic Engineering at Shang-hai Jiao Tong University, China in 2006 and2009. And he served as RA at City Universityof Hong Kong from July, 2008 to August,2009. His research interests mainly includewireless communication architecture, QoS ofwireless networks, network coverage in WSN

and cyber space security.

Weijia Jia is currently a full Professor inthe Department of Computer Science andthe Director of Future Networking Center,ShenZhen Research Institute of City Univer-sity of Hong Kong (CityU), leading severallarge R&D projects on next-generation mo-bile phone and multimedia software and de-vices. He received BSc and MSc from CenterSouth University , China in 1982 and 1984and Master of Applied Sci. and PhD fromPolytechnic Faculty of Mons, Belgium in 1992

and 1993 respectively, all in Computer Science. He joined GermanNational Research Center for Information Science (GMD) in Bonn(St. Augustine) from 1993 to 1995 as a research fellow. In 1995,he joined Department of Computer Science, CityU as an assistantprofessor.

Prof. Jias research interests include next generation wirelesscommunication, protocols and heterogeneous networks; distributedsystems, multicast and anycast QoS routing protocols. In theseelds, he has a number of publications in the prestige interna-tional journals (IEEE Transactions, e.g., TPDS, TN, TMC, TC etc.),books/chapters and refereed international conference proceedings(e.g. ACM CCS, WiSec, MobiHoc, IEEE ICDCS, INFOCOM etc.). In2005 and 2008, he has been awarded total HK$22 millions from theInnovation & Technology Fund of the HKSAR Government for twoprojects with intentions of design and implementation of cyber cross-platform secure communications to integrate the Internet with 3G,WiFi, WiMAX, ad-hoc network, Sensor and networks for real-timemultimedia communications and mobile video surveillance.

Prof. Jia is the Chair Professor of Central South University,Changsha, China, Guest Professor of Shanghai Jiao Tong University,University of Science and Technology of China, Beijing Jiao TongUniversity and Jinan University, Guangzhou, China. He has servedas area editor for prestige international journals (IEEE TPDS andComCom) and chair and PC member/keynote speaker for variousprestige international conferences. He is the Senior Member of IEEEand the Member of ACM.

Dong Xuan received his B.S. and M.S. de-grees in Electronic Engineering from Shang-hai Jiao Tong University (SJTU), China, in1990 and 1993, and Ph.D. degree in Com-puter Engineering from Texas A&M Univer-sity in 2001. Currently, he is an associateprofessor in the Department of ComputerScience and Engineering, The Ohio StateUniversity (OSU). He was on the faculty ofElectronic Engineering at SJTU from 1993to 1997. His research interests include dis-

tributed computing, computer networks and cyberspace security. Hereceived the NSF CAREER Award in 2005 and the Lumley ResearchAward from the College of Engineering, OSU in 2009.


hspa 3.5 g

Documents