48 IEEE TRANSACTIONS ON MOBILE COMPUTING, …web.iitd.ac.in/~swadesd/pubs/JNL/Handoff-TMC2014.pdf · Long Term Evolution-Advanced (LTE-A) and World-wide Interoperability for Microwave

Class-Based Shared Resource Allocationfor Cell-Edge Users in OFDMA Networks

Chetna Singhal, Student Member, IEEE, Satish Kumar,

Swades De, Member, IEEE, Nitin Panwar, Ravindra Tonde, and Pradipta De

Abstract—In this paper, we present a new resource allocation scheme for cell-edge active users to achieve improved performance in

terms of a higher system capacity and better quality-of-service (QoS) guarantee of the users, where we utilize the two-dimensional

resource allocation flexibility of orthogonal frequency division multiple access (OFDMA) networks. Here, the mobile stations (MSs) at

the cell-edge can maintain parallel connections with more than one base station (BS) when it is in their coverage area. A MS, before

handoff to a new BS, seeks to utilize additional resources from the other BSs if the BS through which its current session is registered is

not able to satisfy its requirements. The handoff procedure is termed as split handoff. The BSs participate in split handoff operation while

guaranteeing that they are able to maintain QoS of the existing connections associated with them. In this study, first, we present the

proposed shared resource allocation architecture and protocol functionalities in split handoff, and give a theoretical proof of concept of

system capacity gain associated with the shared resource allocation approach. Then, we provide a differentiated QoS provisioning

approach that accounts for the MS speed, its channel quality, as well as the loads at different BSs. Via extensive simulations in Qualnet,

the benefits of the proposed class-based split handoff approach is demonstrated. The results also indicate traffic load balancing

property of the proposed scheme in heavy traffic conditions.

Index Terms—Split handoff, shared resource allocation, effective capacity, differentiated QoS

Ç

1 INTRODUCTION

STEADILY increasing data rate support along with the

inherent advantages of wireless access networks, such as

easy scalability and low cost of deployment and main-

tenance, have led to the emergence of broadband wireless

access (BWA) as a popular alternative to the wire-line

access infrastructure. The data rate landmarks in fourth

Generation (4G) wireless broadband access networks, like

Long Term Evolution-Advanced (LTE-A) and World-wide

Interoperability for Microwave Access-Mobile (WiMAX-

Mobile), are set around 1 Gbps in downlink and 300 Mbps

in uplink as per the International Mobile Telecommunica-

tions-Advanced (IMT-Advanced) specifications [1]. To

achieve and maintain these very high rates in a wireless

environment, mobile devices/stations (MSs) are required to

change the base station (BS), if there exists one within the

reach of the MS, with, for example, a better link quality.

This procedure is called handoff. Handoff is performed on

the basis of some metric threshold, which can be chosen as

per the communication system requirements, application

constraints of an individual MS, and speed.In today’s evolving wireless networks, the issues

related to handoff are not treated as an isolated physicallayer problem. The protocol end points that are located inthe BS are needed to be moved from the source BS to thetarget BS. This relocation can be done in two differentways: 1) protocol status transfer from the source BS to thetarget BS; 2) protocol reinitialization after the handoff.In LTE networks, the relocation is done using a hybridapproach, where the downlink protocol status is trans-ferred from the source BS to target BS [2], [3]. A packetforwarding approach from the old BS to the new BS hasalso been considered for the in-flight packets, whoseimpact on the user connection has been studied in [3]. Foruplink traffic, random access procedure is performed atthe target BS. The results in [3] showed that the extra loadcaused by the forwarding is not significant. Also, theimpact of forwarding on the end-user connections canbe reduced by using a scheduling architecture at thetransport layer that is able to differentiate among differentservice classes while allocating the bandwidth.

Maintaining guaranteed quality-of-service (QoS) supportat the cell-edge and dimensioning a mobile wirelessnetwork are the two major challenges, especially whenthe user rate demands are high. Therefore, efficientscheduling of BS resources is essential to ensure fairnessto the users and high network performance. Many schedul-ing techniques have been proposed in the literature whichtake care of user-level fairness and network capacity [4], [5],

48 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 13, NO. 1, JANUARY 2014

. C. Singhal is with the Communication Networks Research Laboratory,Bharti School of Telecom, Indian Institute of Technology Delhi, Hauz Khas,New Delhi 110016, India. E-mail: [email protected].

. S. Kumar is with Qualcomm India Pvt. Ltd., Hyderabad 500081, AndhraPradesh, India. E-mail: [email protected].

. S. De is with the Department of Electrical Engineering, Indian Institute ofTechnology Delhi, Hauz Khas, New Delhi 110016, India.E-mail: [email protected].

. N. Panwar is with Cisco Systems Pvt. Ltd., Bangaluru 560087, Karnataka,India. E-mail: [email protected].

. R. Tonde is with the Samsung India Software Centre, Noida 201301, UttarPradesh, India. E-mail: [email protected].

. P. De is with the Department of Computer Science, State University ofNew York (SUNY) Korea, Yeonsu-Gu, Incheon City, Korea.E-mail: [email protected].

Manuscript received 31 Dec. 2011; revised 5 Aug. 2012; accepted 28 Sept.2012; published online 12 Oct. 2012.For information on obtaining reprints of this article, please send e-mail to:[email protected], and reference IEEECS Log Number TMC-2011-12-0701.Digital Object Identifier no. 10.1109/TMC.2012.210.

1536-1233/14/$31.00 � 2014 IEEE Published by the IEEE CS, CASS, ComSoc, IES, & SPS

[6], [7]. The problem aspect of capacity and fairnessguarantee, however, changes when a MS cannot hold onto the existing connection and decides to perform ahandoff, or when a macrodiversity approach is adoptedto mitigate the disconnection and loss of packets indownlink and uplink directions.

1.1 Related Works

In view of high data rate support, several advanced handoffapproaches have been proposed in the recent literature. Thehandoff process is of two major types: hard handoff andsoft handoff. Different variants of hard handoff and softhandoff are semisoft handoff and fractional handoff inOFDM based systems [8], [9], fast base station switching(FBSS) [10], and macrodiversity handoff (MDHO) [10].Before performing handoff, an appropriate BS candidatemust be chosen and then the handoff procedure should becontinued based on the current technology and the specificapplication constraints of the MS. The exact proceduresvary depending on the used technology, and usually withinthe technology several alternatives are available as well.

A routing-based seamless handoff was proposed in[11], where layer-2 and layer-3 handoff progress simulta-neously to minimize the handoff delay and packets loss. Apacket access router was introduced, which connects toboth serving BS and target BS. After the MS sets up a newconnection with the target BS, the router starts sendingduplicate packets to a target BS. Upon notification of thelast packet sent from the serving BS, the target BS startstransmitting the onward packets to the MS. In [12], a fasthandoff scheme was proposed for real-time applications.Here, the target BS receives QoS parameters and thechannel identity (CID) from the serving BS through thebackbone network. The target BS uses these old CIDs fordownlink transmission until new CIDs are assigned at thetarget BS. The arriving packets during the handoffinterruption time are buffered at the serving BS and sentto the target BS over the backbone. In a handoff schemefor downlink traffic in CDMA (code division multipleaccess) systems, called spatial multiplexed soft handoff[13], each participating base station sends only a subset ofthe main data stream, and the MS reassembles them andrestores the main data stream. The motivation was toimprove the processing gain of the system. The limitationof this approach in CDMA systems is due to the receivepower disparity from the participating BSs.

As suggested in [14], depending on the fading condi-tions, the cell coverage overlap area where a MS can haveaccess to more than one BS, can extend up to 47 percent ofthe total area of the two adjacent cells. Thus, in the handoffzone, the signals from two or more BSs can be exploitedeven if the MS is traveling at a high speed, because thecoverage overlap region is sufficiently large. To exploit thisgenuine coverage overlap, simulation-based studies wereindependently carried out in [15] and [16] for downlink cell-edge users. Their results suggested that, the rate supportedto a cell-edge user with resources from multiple BSs is notalways better than the normal transmission from a singleBS. To this end, we argue that, while evaluating such apossibility, the cell loads and the individual user QoSshould be considered while exploiting the signal from more

than one BS, because the user QoS will define the need formulticell transmission, and the load on the BSs will decidethe feasibility and limits of such transmissions.

Interference management in OFDMA networks withdynamic fractional frequency reuse (FFR) and interferenceaware power control was approached in [17], where thehandoff scheme employs disjoint links and beacon. Thisapproach, however, cannot reduce the delay in handoff.Also, FFR limits the overall achievable system capacity. FFR,interference coordination, and interference cancellationschemes are also discussed in [18], [19] with a systemcapacity viewpoint. As the new services with diversified QoSrequirements are evolving, we need to address individualuser QoS along with the other preexisting constraints.

To enhance user QoS and network capacity performance,besides simple data rate guarantee traffic classification isalso important. Explicit differentiated QoS support [20] wasproposed for Internet applications to address traffic class-specific resource allocation, where the traffic was categor-ized into the three specified DiffServ model classes, namely,expedited forwarding (EF), assured forwarding (AF), andbest effort (BE). There have been several recent studies onexplicit QoS support to the BWA users. A set of works, forexample, the studies in [21], [22], [23], [24], [25], addressedQoS support, and service differentiation to the wirelessusers, which focused on single BS access scenario. In [23],physical signal strength and mobility information wereexploited for a better support to MDHO and FBSS. Acomparative protocol-level study on vertical handoff acrossdifferent wireless technologies was presented and differ-entiated QoS support was discussed in [24]. These studiesprovide a less detailed account of service differentiationapproach. Also, they lack on addressing the schedulingissues when resources are available from multiple BSs.

On class-based dynamic resource allocation there havebeen some recent studies in Ethernet passive opticalnetworks. The dynamic bandwidth allocation strategy in[26] assigns a fixed bandwidth to the EF (e.g., voice) traffic,irrespective of the immediate requirements. The leftoverbandwidth is allocated to the AF (e.g., video) traffic first,and then to the BE (data) traffic. The strategy proposed in[27] limits the allocation to EF and AF traffic to theirrespective service level agreements (SLAs), while assigningthe remaining bandwidth to BE traffic. To achieve fairnessamong all classes, the bandwidth allocation in [28] is donein three stages: first allocate resource proportional to thequeue length of all classes, then prune the allocatedresource if it exceeds the respective SLA high value orSLA low value, and finally allocate the excess bandwidthproportionally to all queues. To guarantee strict priorityand optimal resource usage, a further modification wassuggested in [29] which also addressed traffic burstinessdependent optimal prediction of future traffic. We note that,while the studies in optical access domain provide somebasis for differentiated resource allocation, they do not dealwith channel rate degradation at the cell-edge and thepossibility of shared resource usage from multiple cells.

1.2 Motivation and Contribution

Intuitively, resource sharing among the BSs can bebeneficial to the cell-edge users to mitigate packet loss,

SINGHAL ET AL.: CLASS-BASED SHARED RESOURCE ALLOCATION FOR CELL-EDGE USERS IN OFDMA NETWORKS 49

delay, as well as radio link failure during handoff.However, the prior studies did not simultaneously accountfor the user level parameters, such as QoS and speed, andthe network level parameters, such as load distributionamong the BSs. To increase the capacity while improvingthe QoS performance of the active cell-edge users, in thispaper we present a new handoff scheme, which we call splithandoff, in an orthogonal frequency division multiple access(OFDMA) cellular network. The proposed strategy isapplicable to the cellular systems with universal frequencyreuse (UFR) plan (i.e., with frequency reuse factor ¼ 1) [30],as prevalent in LTE and WiMAX systems, as well as withpartial frequency reuse plan (with frequency reuse factor<1). The two-dimensional flexibility (in frequency and timedomains) of resource allocation of OFDMA systems isutilized to allocate resource to cell-edge users from morethan one BS. This allocation from multiple BSs is calledsharing of resources.

Note that, in contrast with the conventional soft handoffschemes, such as MDHO in WiMAX or the soft handoffvariants in CDMA standards, the proposed splithandoff scheme does resource sharing with differentprinciples. 1) Instead of replicating the packets via theparticipating BSs to the mobile user, in split handoff thepacket stream is appropriately divided among the participat-ing BSs to the mobile user. 2) The packet stream splittingprinciple in split handoff takes care of the link qualities,which allows the participating BSs to choose/adapt therespective modulation and coding schemes independently.3) The split handoff further considers the BS loads whiledividing the traffic stream across the participating BSs,leading to a more effective traffic load balancing. It mayalso be noted that MDHO is not generally considered as anefficient alternative, as it is known to have a highbandwidth resource overhead [31, Ch. 15] due to packetreplication and diversity combining from multiple BSs forthe MSs in the overlapping coverage region, resulting in apoorer network capacity.

The proposed traffic splitting implementation approachas well as its objectives are different from those in [13].Unlike in a CDMA-based system, traffic splitting via morethan one BS to one MS in an OFDMA-based cellular accessnetworks, having subcarrier-based frequency and slot-based time resource allocation, is a distinctive challenge,especially due to intercell interference. Also, in contrastwith the goal of increased signal distance in [13], we aim atnetwork capacity gain and user QoS improvement, whereload sharing among the BSs is a function of individual cellload and QoS requirement of the individual users.

One of the main contributions of this work is the systemarchitecture for shared resource usage along with thefunctional details to enable split handoff design. Thechallenge is to construct a DL-map (downlink subcarrierfrequency and time slot allocation map) at the controller,which manages access between BSs and MSs. Typically, aDL-map contains information associated with a single BS.However, in split handoff design, a MS can communicatewith multiple BSs simultaneously. Therefore, the DL-mapconstruction in split handoff must take this into account.

We also present an analytical model to capture capacitygain using the proposed split handoff scheme. To maximize

the performance at the cell-edge, the effective capacityconcept in a single cell scenario [32] is extended to theproposed shared resource allocation policy. The analyti-cally predicted capacity gain performance is verified viarigorous simulations.

Finally, we provide a framework for shared resourceallocation to differentiated service classes (voice, video, anddata), which is implemented in Qualnet simulator usingWiMAX and the other necessary toolboxes. Performance ofthe proposed scheme is compared with the standardnonshared resource allocation policy (hard handoff, with apriori association and authentication with the target BS, i.e., withFBSS option) as well as with MDHO. We show that theproposed resource sharing is quite beneficial in terms ofcapacity gain at the cell edge, especially for the users withstringent QoS requirements. We also show the impact ofnetwork load and user speed on the users’ QoS perfor-mance. It may be highlighted that the gain in the proposedscheme is achieved without requiring extra networkresources, such as power and bandwidth.

1.3 Paper Organization

The rest of the paper is presented as follows: In the nextsection, the proposed system model and split handoffactivities are described. Section 3 deals with analysis of userQoS-related capacity gain using the proposed split handofftechnique. Section 4 presents a practically implementableframework for shared resource allocation to class-basedusers. Section 5 provides the numerical and simulationresults. The paper is concluded in Section 6.

2 SYSTEM MODEL

We consider a BWA system built on OFDMA-based physicallayer technologies, as in LTE or WiMAX standards. Theadjacent BSs with overlapping coverage areas can operate onthe same frequency band (i.e., with frequency reuse factor¼1), as in LTE and WiMAX systems, or they can havedifferent frequency bands. The resources can be shared by amobile user from multiple BSs based on the user require-ments, network resource availability, and coverage condi-tions. This resource sharing eventually results in handoffwhen the condition of coverage and network resourceavailability for sharing are not met, or when the user doesnot need the sharing of resources. The resulting handoffs canbe called in-handoff or out-handoff. In-handoff occurs when auser starts sharing from multiple BSs but later comes back toits parent BS. Similarly, the out-handoff takes place when theuser starts sharing and later it goes to the other BS before thecall ends. Before either of these handoffs, since the totalresource to a user is “split” between the neighboring BSs(based on certain metrics, such as load of the respective BSs,the user’s QoS requirement, etc.), we call these handoffscombinedly as split handoff.

2.1 The Proposed System Architecture

Keeping in mind the flexibility of the proposed scheme wepresent a new transport layer queuing system model, wherea two-level queuing is employed, as depicted in Fig. 1, forall active users to reduce the impact of user movement (in/out-handoff) on the user connection. Here, the nodearchitecture and its interaction with other network entities


is presented. Queuing is applied at the data link control(DLC) layer as well as the transport control (TCP) layer, todistribute the traffic to the BSs which are participating indata transmission of the users in the shared region.

In this paper, scheduling discussions are restricted to thedownlink traffic only, although the scheduling principleapplies to uplink traffic as well with some modifications inthe control message exchange. Also, in Fig. 1 as well as inthe following discussions, we have considered the examplesof resource sharing between two BSs, which is extensible tothree or more BSs.

Two BSs, BSi and BSj are shown in the figure with anoverlapping coverage. There are Ni users that are servedonly by BSi, Nj users served only by BSj, and Nij usersserved by both BSi and BSj in shared mode. BSi maintainsqueues for Ni þNij users and serves them using its Sdownlinkscheduler. Likewise, BSj maintain queues for Nj þNij usersand serves them using Sdownlink. Here, we assume that oneuser has only one class of service at a time. If a single usermaintains multiple parallel connections with different QoSrequirements, then the scheduling can be easily handledby additional queues, called “priority queues,” at a BS.A controller directs the flows from the classifier according

to the routing table maintained therein. Controller andclassifier are the two logical entities that can be physicallycolocated. Based on the feedback from the BSs, the classifieris used to distinguish the incoming/outgoing flows if theyare of a shared user—served by two BSs, or a nonshareduser—served by only one BS. The controller also maintainsthe queues for all users which can be served by both BSs.Flow scheduling at the controller is according to the ruleprovided by Scontroller. The parameters considered forsplitting of traffic are fed back to the controller usingfeedback links.

Some of the advantages of the proposed architecture are:

1. centralized routing information maintenance for thesubscribers to create multiple parallel connectionswhen necessary;

2. avoidance of packet duplication, by distributingpackets for a cell-edge user across the BSs, therebyminimizing resource wastage;

3. rule-based splitting of traffic by using schedulerScontroller; and

4. possibility of resource allocation based on trafficclassification.

2.2 System Functionalities

The controller in the proposed split handoff is connected tothe BSs via high-speed wireline or wireless links. Beyondsignal transmission-reception over the radio links, the BSshave a very little role to play. Functionality-wise, acontroller will perform some extended tasks beyond aconventional base station controller (BSC) or a radionetwork controller (RNC). The specific activities of acontroller in split handoff are: 1) construction of universalDL-map and broadcasting to all BSs, and 2) schedulingand traffic load balancing by accounting the carrier-to-interference-and-noise ratio (CINR) at the MS from theconnected BS and the neighboring BSs, available resourcesof the neighboring BSs, and QoS requirements of subscri-bers. The participating BSs are assumed synchronizedthrough the controller.

We have used the following terms and assumptions inthis paper. Primary BS (PBS) is the BS with which a MSexchanges the management messages as well as data.Secondary BS (SBS) is a BS with which the MS exchangesonly data. Following the WiMAX standard notations forchannel usage, the downlink interval usage code (DIUC)used by a MS with the PBS is denoted as DIUC1, and theDIUC used by a MS with the SBS is denoted as DIUC2. Asindicated in the proposed system architecture (Section 2.1),traffic splitting is done at the transport layer. The controllerstores the BS IDs and their associated loads. With respect toa particular MS, it stores the MS ID, its MAC address, PBSID, DIUC1, priority calculated based on the service flowQoS parameters, and SBS ID and DIUC2—in case the MS isin contact with two BSs.

The timing diagram of a MS session and a split handoffprocess is shown in Fig. 2, in which only downlink datatraffic is considered. The initialization procedure is similar asin a standard service flow set up. The MAC information andDIUC1 of the MS is passed to the controller at the networkentry phase. During data and management messageexchange with the PBS, the MS sends scanning request


Fig. 1. System architecture, with different downlink queue structure forthe shared and nonshared users.

(MOB_SCN-REQ) to the PBS when the CINR from the PBSfalls below Threshold-1 (which corresponds to a higher-than-the-lowest modulation and coding rate). If the re-sponse (MOB_SCN-RSP) from the PBS is positive, the MSstarts scanning for a SBS and synchronizes with one,initiating the split handoff set up phase. Otherwise, itcontinues its connection with the PBS only. Duringscanning, the MS sends (via the PBS) the SBS information(BS ID, DIUC2, traffic load) to the controller. Subsequently,a new CID for data connection with the SBS is created bythe controller, which is forwarded to the MS. With theongoing CID for the PBS and the new CID for the SBS, theuser-level split handoff procedure starts. The timing informa-tion of the bursts (slots) for connecting to the PBS and SBS isnotified to the MS by the controller via a universal DL-map.To address user QoS and cell load imbalance, the controlleraccounts for the QoS priority, buffer status at the MS, andthe BS traffic load at the time of burst scheduling. The bursttimings for the PBS and SBS are separated within a framesuch that the subcarrier frequency reassignment latency ofthe MS is sufficiently accommodated. Note that, withfrequency reuse factor ¼1, it may be possible that the MSis connected to the BSs at different time slots over the sameassignment of subcarriers, in which case no reassignment isnecessary. On the other hand, in case the adjacent BSsoperate at different carrier frequencies, i.e., with frequencyreuse factor <1, split handoff involves subcarrier reassign-ment latency. Finally, when the CINR from the PBS fallsbelow Threshold-2 (corresponding to the lowest allowabledata rate), a PBS change request is sent to the controller, andat that point the SBS assumes the responsibility of PBS forthe MS. This process marks the end of split handoff.

3 ANALYTICAL MODEL FOR SHARED RESOURCE

ALLOCATION

In this section, we present an analytical framework todevelop an intuition of capacity gain via shared resourceusage by the cell-edge users. Since hard handoff is thedefault scheme in LTE/WiMAX standards, we consider thisscheme as the baseline. Later, in Section 5 we willdemonstrate via numerical and rigorous system simulationstudies the capacity gain as a function of user QoS, wherethe relative performance with respect to soft handoff as wellis considered.

3.1 QoS and Effective Capacity

Theoretically, QoS is defined by the maximum tolerabledelay Dmax for a user (traffic type) beyond which the delayviolation probability exceeds a predefined threshold �[32], i.e.,

supt

PrfDðtÞ � Dmaxg � �:

It was also shown in [32] that, for a dynamic queuingsystem, where the arrival and service processes arestationary and ergodic, the probability that the delay attime t, DðtÞ exceeds the threshold Dmax can be accuratelygiven as

supt

PrfDðtÞ � Dmaxg � �ð�Þe��ð�ÞDmax ; ð1Þ

where �ð�Þ is a function of constant source rate �, Dmax is asufficiently large quantity, and �ð�Þ is the probability thatthe delay of a particular packet is nonzero, i.e., �ð�Þ ¼PrfDðtÞ > 0g at a randomly chosen time instant t. Here,�ð�Þ > 0 in (1) is a parameter describing the exponentialdecay rate of probability of QoS violation. �ð�Þ is referred asthe QoS exponent. A large value of �ð�Þ corresponds toa fast decaying rate, i.e., a stringent QoS requirement, and asmall value of �ð�Þ corresponds to a slow decay rate, i.e., aloose QoS requirement.

The effective capacity for a given QoS exponent �specifies the maximum constant arrival rate that can besupported by the system at the link layer.1 Effectivecapacity concept in [32] can be applied to wireless channelswith arbitrary physical-layer characteristics. For a discrete-time stationary and ergodic service process with rate �ðnÞand channel service rate RðmÞ ¼

Pmn¼0 �ðnÞ, the effective

capacity is defined as

ECð�Þ ¼def � limm!1

1

�mlnEfe��RðmÞg; ð2Þ

where m is the block length. For uncorrelated block fadingchannels, where the service process f�ðnÞ; n ¼ 1; 2; . . .g isalso uncorrelated, the expression (2) is reduced toECð�Þ ¼ � 1

� lnEfe��ðnÞg, given any n ¼ 1; 2; . . . We furthernormalize the effective capacity with respect to the framelength Tf and system bandwidth B as

ECð�Þ ¼ �1

�TfBlnEfe��g; ð3Þ


Fig. 2. Timing diagram of a MS session and split handoff process, withonly the downlink traffic considered.

1. Here, dependency of � on the source rate � is not shown for simplicity.In rest of the paper, we will follow the same convention unless otherwisementioned.

where the product TfB is the total time-frequency resourcesavailable in one frame S ¼ TfB.

3.2 Scheduling of Shared Users for MaximizingCapacity

An interesting saturation condition is where the totalresource demand is more than the available resources atthe BS such that

PNi

u¼1 Su ¼ Si, where Si is the total resource

available per frame in BSi, and ð0 � Su � SiÞ is the resourceallocated to the user u from BSi to maintain its QoSdemand. For simplicity without loss of generality weassume, the total resource available at each BS in a clusterwith UFR plan are equal, i.e., Si ¼ S 8 i 2 Nc.

From (3), for a user u scheduled from BSi, the effectivecapacity can be expressed as

EuC;ið�uÞ ¼ �

1

�uSlnE

�e��

u�ui�; ð4Þ

where �u is the QoS exponent of user u and �ui ¼ rui Su is therate provided to user u from BSi. Here, rui is the modulationindex of user u from BSi.

Now, if the same user is scheduled from two BSs, forexample, BSi and BSj, then the total effective capacity,which we term as joint effective capacity Eu

C;jointð�uÞ, isdefined as

EuC;jointð�uÞ ¼ �

1

�uSlnE

�e��u�u

ið1Þ�� 1

�uSlnE

�e��u�u

jð2Þ�

¼ � 1

�uSln�E�e��u�u

ið1Þ�E�e��u�u

jð2Þ��;

ð5Þ

subject to the condition that the joint resources from BSiand BSj are the same as in (4) and the CINR is above theacceptable threshold �th. Stated mathematically: Su ¼Suið1Þ þ Sujð2Þ with the conditions f�i; �jg > �th. Here, �uiðkÞindicates the kth part of the rate achievable from BSi foruser u. Similarly, SuiðkÞ indicates the kth part of the resourcesallocated from BSi to user u.

Let us denote pi ¼ Prf�ui � �thg and pj ¼ Prf�uj � �thg.Then,

E�e��u�u

ið1Þ�¼ e��

u�uið1Þ ð1� piÞ þ pi;

E�e��u�u

jð2Þ�¼ e��

u�ujð2Þ ð1� pjÞ þ pj:

For simplicity of the subsequent expressions, the modulationindices for the user u from the two BSs are assumed equal.That is, rui ¼ ruj � r. Hence, the expression (5) reduces to

EuC;jointð�uÞ

¼ � 1

�uSln��e��urSu

ið1Þ ð1� piÞ þ pi��e��urSu

jð2Þ ð1� pjÞ þ pj��;

s:t: Suið1Þ þ Sujð2Þ ¼ Su and Su > Suið1Þ; Sujð2Þ > 0:

ð6Þ

Note that, (4) is the effective capacity when a user isscheduled from only one BS, i.e., from BSi, whereas (6) isthe effective capacity when a user is scheduled from twoBSs, i.e., from BSi and BSj. To maximize Eu

C;joint in (6), firstwe substitute Suið1Þ with Su � Sujð2Þ and denote

�e��urðSu�Su

jð2ÞÞð1� piÞþ pi��e��urSu

jð2Þ ð1� pjÞþ pj�¼ �

�Sujð2Þ

�:

ð7Þ

By differentiation of (7) with respect to Sujð2Þ and equating

it to zero, we have

Sujð2Þ ¼Su

2þ 1

2�urlnð1� pjÞ=pjð1� piÞ=pi

� ; ð8aÞ

Suið1Þ ¼Su

2þ 1

2�urlnð1� piÞ=pið1� pjÞ=pj

� : ð8bÞ

By double differentiation of (7), it can easily be proved

that the obtained values of Suið1Þ and Sujð2Þ maximize the joint

effective capacity in (6).When the MS is in the coverage region of both the BSs,

(8a) and (8b) give the resource allocation from them, so that

the total effective capacity can be increased.Numerical and simulation results on capacity gain with

shared resource usage in a single class traffic environment

will be discussed in Section 5.

4 CLASS-BASED SHARED RESOURCE ALLOCATION

To implement a class based resource allocation policy, we

learn from the dynamic bandwidth resource allocation

policies in optical access networks [26], [27], [28], [29].

However, the proposed allocation policy in BWA is

additionally influenced by the variability of available

bandwidth, shared BS resource usage by the cell-edge

users, and dynamic cell load conditions.Although the current standard service differentiation

approaches suggest to divide the user traffic into five

service classes (e.g., in [10]), for a proof of concept service

differentiated shared resource usage, we categorize the user

traffic into three classes: P0 (voice packets), P1 (video

traffic), and P2 (data traffic) [20]. P0 traffic is the most delay

sensitive, requiring a guaranteed channel bandwidth, P1

has a higher delay flexibility but requires a minimum

bandwidth (rate) guarantee. P2 traffic has neither delay nor

bandwidth guarantee constraints. The proposed priority

allocation approach can be easily extended to more number

of classes.We consider time frames of fixed length Tf seconds for

downlink resource scheduling. Let, at any instant there beNi users within the coverage region of BSi only, Nj userswhich are within the coverage region of BSj only, and Nij

users in the overlapping coverage region of BSi and BSjboth. P0 and P1 traffic have their respective service levelagreements (SLAs), which are the respective upper limits ofresource that can be allocated to them. Since P0 traffic ismost delay sensitive, first resources are allocated to thistraffic class. Then, the resources are allocated to P1 type oftraffic in the first phase. Remaining resource at each basestation are calculated as excess resource, which is allocatedto P1 traffic in second phase and to the P2 traffic.

For resource allocation to the mobile users in a class Pc,

c 2 f0; 1; 2g, we successively select the user with maximum

scheduling function u [33], where, u ¼ max f�u�ug, �u is

current bit rate of the user u (based on its channel

conditions), �u is the user throughput to ensure fairness.

To maximize throughput, �u will ensure that the users with

best channel conditions is selected while �u will ensure that


no user experiences starvation. The users with high �u and/or low �u will be selected, hence ensuring fairness.

4.1 Predicted Resource Allocation

Time-frequency resource allocation for the mobile userswith different service classes is influenced by the statisticalcharacteristics of the traffic arrival and the channel behavioras well as the usage of the backlog history. To predict theresource required due to the incoming traffic over a frameinterval Tf , we adopt a linear predictor [29], [34]:

~Suð�ÞPcðnþ 1Þ ¼

XLPc�1

l¼0

uPc;lðnÞSuð�ÞPcðn� lÞ;

where c 2 f0; 1; 2g and LPc is the prediction order—afunction of traffic type Pc. ~S

uð�ÞPc

is predicted resourcerequirement for user u and traffic type Pc due to newarrivals over the interval Tf . uPc;l is the parameter indicatingthe impact of the actual resource requirement S

uð�ÞPcðn� lÞ

due to new arrivals in frame ðn� lÞ on the predictedresource requirement ~S

uð�ÞPc

for user u and priority type Pc.uPc;l is updated by standard least mean square (LMS)algorithm as [34]

uPc;lðnþ 1Þ ¼ uPc;lðnÞ þ uPcðnÞ

"uPcðnÞSuð�ÞPcðnÞ

;

where "uPcðnÞ is the prediction error in the nth frame, definedas: "uPcðnÞ ¼ S

uð�ÞPcðnÞ � ~S

uð�ÞPcðnÞ, and uPcðnÞ is defined as:

uPcðnÞ ¼LPcPLPc�1

l¼0

�Suð�ÞPcðn� lÞ

�2 :With the predicted new arrivals, the requested resource

SuðrÞPcðnþ 1Þ for frame ðnþ 1Þ and Pc type traffic is

SuðrÞPcðnþ 1Þ ¼ SuðqÞPc

ðnÞ þ ~Suð�ÞPcðnÞ;

where the superscript q indicates the resource required dueto the queued traffic.

Note that, it is important to choose an optimum numberof taps LPc for a given traffic type Pc that would maximizethe quality of prediction ~S

uð�ÞPc

using the historical prediction.As studied in [29], while a higher value of LPc wouldincrease the prediction accuracy by sharply tracking thetraffic burstiness, it causes a higher lag in predictedtraffic—which can also be detrimental to the predictionquality. On the other hand, a very small value of LPc may notclosely track the burstiness, although it tracks the traffic fast.

4.2 Allocation of Resources: PP 0 Traffic

P0 traffic has a strict delay constraint. Accordingly, for the

users connected to BSi only, the granted resource in frame

ðnþ 1Þ is: SuðgÞP0;iðnþ 1Þ ¼ minfSuðrÞP0;i

ðnþ 1Þ; SLAP0g. The allo-

cation to the users in BSj only is similarly done. For the

users in shared region, the resource allocations from BSi

and BSj are: SuðgÞP0;ið1Þðnþ1Þ¼ Lj

LiþLj �minfSuðrÞP0;i;jðnþ 1Þ; SLAP0

gand S

uðgÞP0;jð2Þðnþ 1Þ ¼ Li

LiþLj �minfSuðrÞP0;i;jðnþ 1Þ; SLAP0

g, where

Li and Lj are the respective loads on BSi and BSj. This

strategy ensures that more resources are allocated from the

lightly loaded BS.

4.3 Allocation of Resources: PP 1 Traffic (Phase 1)

For the users associated with only BSi, ðu ¼ 1 to NiÞ,

SuðgÞP1;iðnþ 1ÞjI ¼

SuðrÞP1;iðnþ 1Þ; if S

uðrÞP1;iðnþ 1Þ � SLAP1

;SLAP1

; otherwise:

A similar approach is taken for the users associated with

only BSj. For the users in the shared region (for

u ¼ 1 to Nij), the allocated resource SuðgÞP1;i;jðnþ 1ÞjI ¼

SuðgÞP1;ið1Þðnþ 1ÞjI þ S

uðgÞP1;jð2Þðnþ 1ÞjI is shared between BSi

and BSj based on their loads, as done for P0 traffic.After allocation of resources to P0 traffic and P1 traffic

in the first phase, the remaining resources SðeÞi ; S

ðeÞj are

calculated:

SðeÞi ¼ Si �

XNi

u¼1

�SuðgÞP0;iþ SuðgÞP1;i

��I

��XNij

u¼1

�SuðgÞP0;ið1Þ þ S

uðgÞP1;ið1Þ

��I

�;

SðeÞj ¼ Sj �

XNj

u¼1

�SuðgÞP0;jþ SuðgÞP1;j

��I

��XNij

u¼1

�SuðgÞP0;jð2Þ þ S

uðgÞP1;jð2Þ

��I

�:

4.4 Allocation of Resources: PP 1 Traffic (Phase 2)and P2 Traffic

For the users associated with only BSi, ðu ¼ 1 to NiÞ,

SuðgÞP1;iðnþ 1ÞjII ¼

SuðrÞP1;i

; ifSuðrÞP1;i�SLAP1

;

SLAP1þ

SðeÞi �S

uðrÞP1 ;iPNi

u¼1

�SuðrÞP1 ;iþSuðrÞ

P2 ;i

�þPNij

u¼1

�SuðrÞP1 ;i;jþSuðrÞ

P2 ;i;j

�Lj

LiþLj

; otherwise;

8><>:and

SuðgÞP2;iðnþ 1Þ ¼ min

SuðrÞP2;i

;SðeÞi � S

uðrÞP2;iPNi

u¼1

�SuðrÞP1;iþ SuðrÞP2;i

�þPNij

u¼1

�SuðrÞP1;i;jþ SuðrÞP2;i;j

� LjLiþLj

8<:

9=;:

Similarly, the resources are allocated to the users in BSj.

For the users in the shared region ðu ¼ 1 to NijÞ, the

resources are allocated in phase II from the two BSs based

on their traffic loads. Thus, the fractional allocation fromBSi,

for example, are: SuðgÞP1;ið1Þðnþ1ÞjII¼

LjLiþLj � S

uðgÞP1;i;jðnþ1ÞjII , and

SuðgÞP2;ið1Þðnþ 1Þ ¼ Lj

Li þ Lj� SuðgÞP2;i;j

ðnþ 1Þ:

5 NUMERICAL AND SIMULATION RESULTS

In light of the practical deployment scenarios, we have

considered three different cases: The first scenario deals

with two adjacent BSs and a straight line movement

trajectory of the MS through them, to justify the gain of the

proposed scheme over a typical hard handoff as well as a

soft handoff approach (MDHO), verify via simulations the

numerical results the system capacity gain with respect

to hard handoff, and study the effect of traffic loads in the

BSs. The second scenario is with three BSs in sequence, to

further generalize the impact of split handoff on different

service classes. In the third scenario, we have taken a


generalized 3-tier (19 cell) network with random move-ment pattern of the MSs, to study the average networkperformance.

For verification of numerical results as well as forgeneralized network performance, the simulations werecarried out in Qualnet 5.2 with mobile WiMAX toolbox foremulating broadband wireless access. Urban propagationlibrary was used for channel modeling. Hard handoff wasimplemented with FBSS option. For implementing sharedresource allocation in split handoff and replicated resourceallocation in MDHO, the necessary control message mod-ifications in uplink and downlink frames were done in UL-map and DL-map to allow downlink data communicationto the MS via more than one BS and control messageexchange via the PBS only. In all simulations other than forverification of numerical results, adaptive modulation andcoding scheme was enabled so that a MS can select asuitable rate depending on its channel condition.

Following the typical cases of wireless systems withadaptive modulation, the downlink power budget for eachuser is kept constant. The signals from the center cell as wellas from the neighboring cells are considered with path lossand shadow fading. Unless otherwise specified, the defaultsystem settings and parameters in the simulations arementioned in Table 1.

The class-based traffic parameters are taken as follows:P0: VoIP traffic—exponentially distributed with averageON time 1.34 s and OFF time 1.67 s; P1: video traffic—Parissequence with H.264 variable bit rate encoded at 30 framesper second with 352� 288 pixels/frame; P2 is the besteffort data traffic, which is taken as web traffic with meanpacket size 1,500 Bytes and Poisson distributed packetinterarrival times with mean 0.133 s. In our simulationstudies, we have determined the respective optimumvalues of prediction order LPc , for c 2 f0; 1; 2g, as 1, 4,and 2. The ratio of P0, P1, and P2 type users were taken as10 : 45 : 45. The performance measures were: packet drop

rate for P0, normalized throughput for P1, and packetdelay for P2.

5.1 Scenario 1: Two BS Case

5.1.1 Enhanced Data Rate during Handoff

A simple simulation scenario with two BSs and one MSis depicted in Fig. 3. The speed of the MS is 5 m/s. Forthis specific scenario, a CBR connection with Packet size1,024 Bytes.

With the two BSs equally loaded, supported data ratewith the two handoff schemes at different position of theMS is shown in Fig. 4. In the handoff (coverage overlap)region the throughput achieved with the proposed handoffis much higher than hard handoff as well as MDHO.Specifically, our simulation results show that the averagethroughput maintained by split handoff in the coverageoverlap region of the two BSs (spanning about 200 mdistance) is 3.69 Mbps in comparison to 2.22 Mbps via hardhandoff and 3.07 Mbps via MDHO. Note that, when the MSis at either of the cell boundaries, the performance of allthree handoff schemes are equally poor, as there are nosharable network resource for the split handoff in this two-BS scenario.

5.1.2 System Capacity Gain

The numerical results from the analysis developed inSection 3 on capacity gain of split handoff with respect to


TABLE 1Default System Parameters for Simulations

Fig. 3. Two BS scenario with straight line trajectory of MS from one BSto the other.

Fig. 4. Comparison of data rate in the three handoff schemes. Mobilespeed 5 m/s.

hard handoff are compared with the simulation results,where two different service classes, P0 and P1, wereindividually considered. The capacity gain is computed as

Eu;maxC;gainð�uÞ¼

def Eu;maxC;jointð�uÞ �Eu

Cð�uÞEuCð�uÞ

� 100%: ð9Þ

Here, in each direction of movement of the MS, the numberof BSs that can share the resources are limited to two.Beyond the required parameters in Table 1, the parametervalues considered for the numerical results are as follows:Path loss factor ¼3; modulation scheme is 4-QAM (corre-spondingly, bits per symbol is r ¼ 2); distance between thetwo adjacent BSs is 700 m.

Fig. 5 shows the numerical results for two different valuesof QoS exponent � on capacity gain, which are plottedagainst distance of the MS from the center of one BS to theother. As the distance between the two BSs is 700 m, theresults are presented for 0 up to 350 m. The MS is consideredtraveling at a constant speed of 5 m/s. It can be noted that,the gain is higher for the traffic with a stricter QoS constraint(i.e., with a higher value of �). The plots also show that, ascompared to a user with a loose QoS requirement, the onewith a stringent QoS starts benefiting earlier in its trajectoryto the adjacent BS. For example, at 200 m distance from thecenter of the left-side BS (Fig. 3), the analytically predictedcapacity gain of P0 traffic with split handoff is about26 percent, whereas that of the P1 class is still 0. Intuitively,the capacity gain reaches the maximum value in both trafficclasses when the MS is equidistant from the two BSs.

Fig. 5 also presents the comparison of analysis andsimulation results for the two different classes of traffic,which correspond to two different values of �. Here, P0

(VoIP) traffic corresponds to the numerically generated plotwith � ¼ 1:5, and P1 (video) traffic corresponds to thenumerical result with � ¼ 0:8. The numerical and simula-tion results are reasonably well matched. A little differencecan be attributed to the fact that the constant arrival rateassumption, which is a baseline for the effective capacitydefinition, does not strictly hold in simulations withpractical parameters of the two traffic classes.

5.1.3 Traffic Load Balancing

Fig. 6 shows the impact of BS loading on sharing ofresources from the two BSs. The mobile speed considered

here is 20 m/s. The two curves show the data packets

handled by the two sharing BSs. It can be noticed in Fig. 6a,

when the BS1 is carrying more load (with utilization factor

�1 ¼ 0:67) as compared to BS2 (�2 ¼ 0:33), the sharing

window is positioned closer to the BS1 center. This is

because, for load balancing the BS2 starts sharing the traffic

from the handoff region quite early. Fig. 6b shows the same

impact on sharing window position when the load on BS1

is �1 ¼ 0:33 and on BS2 it is �2 ¼ 0:67.Fig. 7 presents the impact of the proposed scheme

compared to the hard handoff and MDHO on packet drop

rate. When the MS is close to either of the BSs, i.e., at the

positions X ¼ 400 m and X ¼ 1100 m (see Fig. 3), the data

packet drop rate is lower. But as the MS goes away from

either of the BSs, the packet drop rate increases. In split

handoff scheme, the drop rate is significantly less than the

hard handoff in the coverage overlap region because of its

resource sharing capability. The replicated resource usage

in MDHO helps stem the packet drop rate compared to that

in hard handoff, but its performance is poorer than split

handoff because, by its principle of macrodiversity for all

cell-edge users the available resource from the neighboring

BSs for the cell-edge users is reduced. In particular, the

three curves in Fig. 7 indicate that the highest packet drop

probability during handover is reduced to 20 percent

in split handoff from 35 percent in hard handoff and

27 percent in MDHO. In other parts of the trajectory, where

there is no sharing possibility, both curves are overlapping,


Fig. 5. Comparison of effective capacity gain of split handoff with respectto hard handoff, as defined in (9), ref. experiment scenario in Fig. 8.�th ¼ 3 dB, � ¼ 0:9, and MS speed 5 m/s.

Fig. 6. Distribution of number of data packets handled with movementtime elapsed (ref. mobility scenario in Fig. 3) and position of sharingwindow at different cell loads. Mobile speed 20 m/s.

Fig. 7. Comparison of packet drop probability in the three handoffschemes. Mobile speed 20 m/s.

which is intuitive as the unique features of the split handoffscheme is not exploited there.

The key observations on split handoff from this studycan be summed up as follows: 1) More data rate can beachieved while sharing the resources from more than oneBS. 2) The sharing window shifts toward the BS which ismore loaded, demonstrating cooperation and load sharingamong the neighboring BSs. 3) The packet drop rate is alsoless in case of shared resource allocation strategy, as theproposed strategy exploits the additional resource from theneighboring BSs in a more efficient way than in MDHO.

5.2 Scenario 2: Three BS Case

A more generalized scenario was considered next withthree BSs and the MS trajectory from the center of the firstBS to the center of the last BS, as depicted in Fig. 8. Thesimulation results are presented against movement time,starting at the MS position in BSi and stopping in BSk.There are two handoff points in the full trajectory. Thenetwork carries all three classes of traffic: P0, P1, and P2.P0 packet drop rate and P2 packet delay are captured in

the three handoff strategies, as shown in Fig. 9. As the MSmoves, the number of packets dropped in Fig. 9a are nearlythe same for all the handoff schemes, except inside thecoverage overlap region where the handoff takes place.During hard handoff all stored packets are dropped, and sodrop rate during that time is very high. In MDHO, due toreplicated resource usage from the two BSs, the performanceimproves to some extent. In split handoff, on the otherhand, as more than one downlink path is available throughsharing, the packets can be successfully delivered by takingthe diversity advantage. The average delay performance in

Fig. 9b can be similarly explained. Unlike in split handoff,in hard handoff as well as in MDHO the packetsexperiencing bandwidth resource limitations are droppedduring the handoff. They have to be either retransmitted, ortunneled from the first BS to the second, which introducesdelay. The added fluctuation in delay performance can beattributed to the best effort nature of P2 traffic, which isserved after guaranteeing the P0 and P1 SLAs.

5.3 Scenario 3: Three Tier Case (a Cluster of19 BSs)

To study the effect of the proposed handoff strategy ata broad level in a typical cellular scenario, we have taken a3-tier cellular structure with 19 BSs in a cluster, and the usertraffic is broadly categorized into three classes: P0 (VoIP), P1

(packet video), and P2 (data).By the principle of split handoff the resource sharing for

a particular MS can be from two or more BSs such that thegain in capacity is maximized. However, the number ofconnection threads required with the increased number ofshared BSs as well as the increased number of MSs and theconsequent increase in runtime in our machine (DellOptiplex 990) are noted to be very high. Therefore, werestrict our simulation scenario and resource sharingbetween only two neighboring BSs (PBS and SBS). EachBS has up to 38 MSs that are uniformly distributed over thecoverage area of a BS. Mobile users are considered movingwith a random mobility.

Fig. 10a shows the comparison of average packet droprate of P0 traffic versus speed during the handoff phase. Asshown earlier in Fig. 9, the packet drop rate during handofffor a constant speed are higher in hard handoff and MDHOas compared to the proposed scheme. It can be observedfrom Fig. 10a that as the speed of the users increases thepacket drop rate in the proposed scheme increases at ahigher rate. With the increased speed, the performance gainof proposed scheme tends to level off. The reason is that witha higher mobility, the effective region for resource sharing isdecreased, resulting in a reduced performance benefit.

Fig. 10b shows the throughput comparison of P1 (video)traffic versus speed of the mobile users. The throughput hasbeen normalized with respect to the highest achievablethroughput. The first observation is that the throughputdecreases in all handoff schemes as the mobile speedincreases. This is because the packet drop rate increaseswith speed. Second, as the speed increases the throughput


Fig. 8. Three BS scenario with straight line MS movement. Center-to-center distance ¼ 700 m.

Fig. 9. Comparison of class-based performance: (a) packet drop rate forP0 traffic; (b) average packet delay for P2 traffic. Mobile speed 20 m/s.

gain in the proposed scheme with respect to the hardhandoff and MDHO diminishes. The reason is same as inFig. 10a; as the speed increases the sharing window sizeshrinks and hence the proposed scheme does not performas good as in lower speeds. The delay performance of P2

traffic is also plotted in Fig. 10c. Although the plots look alittle random, the variations in magnitude are ratherminuscule. Yet, there is a clear trend of increased delayperformance at higher speeds in all three handoff schemes,which has the same intuitive reasoning as in Figs. 10a and10b. The randomness in the delay plots could be because ofthe least-priority handling of the P2 traffic, which is servedonly after serving P0 class and P1 class. The bursty naturesof P0 and P1 traffic entail that the resource available for P2

can vary widely from frame to frame, leading to a poorconvergence to its average performance. Also, the increasein velocity has very little impact on the delay variation,resulting in a bloated representation of the minor variationin delay.

Fig. 11 shows variation of the same performanceparameters as in Fig. 10 for P0, P1, and P2 traffic withtraffic load per BS. In Fig. 11a, the packet drop rate for P0

traffic is drawn. The drop rate in the proposed scheme isless compared to those in hard handoff and MDHO. But itis noticeable that as the average load per BS increases, thepacket drop rates in hard handoff and MDHO increase athigher rate as compared to the proposed scheme. Thus, theproposed scheme is more beneficial in high load conditionsfor high QoS applications. Fig. 11b shows P1 throughputperformance of the proposed scheme in comparison with

hard handoff and MDHO. The throughput of the proposedscheme is quite higher. Moreover, compared to the othertwo handoff schemes, the throughput has less impact of thenumber of users per BS on the proposed scheme. As thenumber of users increases, the throughput tend to decrease.But, in split handoff more opportunities are created toshare resource for active handoff users, and as a result therate of decay in throughput is less. Average delay for P2

traffic (during handoff) versus traffic load per BS is shownin Fig. 11c. Again, the delay of the proposed scheme is lessas we have seen in the two BS case. The increment in delayin the proposed scheme is less, which is attributed to thesame reason of resource sharing. It may be noted here that,in contrast with the impact of speed variation on P2 delay,the impact of cell load is much significant, and as a result,the trends of delay variation versus cell load is smooth.

We also present single BS statistics in a two-cellenvironment in terms of percentage of shared resourcesand position of sharing window with the proposed scheme.In general, the number of users per BS basically representsits load. We recall that static users are kept constant becausethey do not take part in handoff. So, as the number ofmobile users increases, the load on BS also increases. Fig. 12shows the effect of load in a BS on the sharing windowposition and resources shared from the more loaded BS. Asthe load increases on a BS, it extracts the benefit of sharingby shifting more of its load to the neighboring BS. Withrespect to the impact on load sharing window, the sharingwindow position means the location of midpoint of sharingzone between two BSs. The center of sharing window shiftstoward the BS, which is heavily loaded to distribute itstraffic to the nearby BS in form of shared users. So, in thisway the split handoff inherently takes care of the loadbalancing property of a cellular network.

6 CONCLUSION

We have presented a QoS aware handoff strategy inOFDMA broadband cellular networks, called split handoff,that allows sharing of channel resource from more than oneBS when the user is in their joint coverage area. We haveproposed a system architecture that supports the proposedscheme, outlined the system interactions involved, andprovided an analytical framework to quantify the capacity


Fig. 10. Class-based performance in a multicell handoff scenario atdifferent speed of mobile users. (a) P0 packet drop rate; (b) P1 packetthroughput; (c) P2 packet delay. Average number of MS per BS is 38.

Fig. 11. Class based performance in a multi-cell handoff scenario atdifferent traffic loads. (a) P0 packet drop rate; (b) P1 packet throughput;(c) P2 packet delay. Speed of mobile users is 25 m/s.

Fig. 12. Load sharing performance versus load in a cell, with the otherBS having a fixed load of 100 users. Speed of mobile users is 30 m/s.

gain by shared resource usage. The capacity enhancementanalysis has been validated by network simulation resultsusing Qualnet. Further, we have provided a heuristic class-based shared resource allocation policy for the cell-edgeusers that aims at maximizing the QoS support to differentservice classes while maximizing the resource utilization.We have conducted rigorous network simulations inQualnet simulator, where mobile WiMAX has been con-sidered as an example broadband wireless access networkstandard. Our results demonstrated that the sharedresource allocation in the proposed split handoff schemesignificantly improves the network performance withrespect to the hard handoff as well as the macrodiversityhandoff in terms of system capacity, QoS guarantee, as wellas traffic load balancing, without incurring additionalnetwork operation cost, such as power consumption andbandwidth usage. The proposed strategy and system modelcan be customized to scenario-specific requirements.

ACKNOWLEDGMENTS

This work was supported by the US Department of Scienceand Technology (DST) under grant no. SR/S3/EECE/0122/2010. The authors are thankful to the anonymous reviewersfor their constructive criticism, insightful comments, andvaluable suggestions, which have significantly improvedthe quality of the paper. P. De was previously with IBM-India Research Lab, New Delhi, India.

REFERENCES

[1] E. Dahlman, S. Parkvall, and J. Skold, 3G Evolution: HSPA and LTEfor Mobile Broadband, second ed. Press, 2008.

[2] “3GPP Long-Term Evolution (LTE),” http://www.3gpp.org/Highlights/LTE/LTE.htm, 2013.

[3] L. Bajzik, P. Horvath, L. Korossy, and C. Vulkan, “Impact of Intra-LTE Handover with Forwarding on the User Connections,” Proc.IEEE Mobile and Wireless Comm. Summit, July 2007.

[4] B.G. Lee, D. Park, and H. Seo, Wireless Communications ResourceManagement, first ed. Wiley-IEEE, Dec. 2008.

[5] J. Tang and X. Zhang, “Cross-Layer-Model Based AdaptiveResource Allocation for Statistical QoS Guarantees in MobileWireless Networks,” Proc. ACM Int’l Conf. Quality of Service inHeterogeneous Wired/Wireless Networks, Aug. 2006.

[6] T. Ali-Yahiya, A.-L. Beylot, and G. Pujolle, “An Adaptive Cross-Layer Design for Multiservice Scheduling in OFDMA BasedMobile WiMAX Systems,” Elsevier Computer Comm., vol. 32,pp. 531-539, Feb. 2009.

[7] Z. Kong, Y.-K. Kwok, and J. Wang, “A Low-Complexity QoS-Aware Proportional Fair Multicarrier Scheduling Algorithm forOFDM Systems,” IEEE Trans. Vehicular Technology, vol. 58, no. 5,pp. 2225-2235, June 2009.

[8] H. Lee, H. Son, and S. Lee, “Semisoft Handover Gain Analysisover OFDM-Based Broadband Systems,” IEEE Trans. VehicularTechnology, vol. 58, no. 3, pp. 1443-1453, Mar. 2009.

[9] J. Chang, Y. Li, S. Feng, H. Wang, C. Sun, and P. Zhang, “AFractional Soft Handover Scheme for 3GPP LTE-AdvancedSystem,” Proc. IEEE Int’l Conf. Comm., June 2009.

[10] IEEE Std. 802.16e, Standard for Local and Metropolitan Area NetworksPart 16: Air Interface for Fixed and Mobile Broadband Wireless AccessSystems, IEEE, 2005.

[11] P. Li, X. Yi, and Y. Pan, “A Seamless Handover Mechanism forIEEE 802.16e Systems,” Proc. Int’l Conf. Comm. Technology, Nov.2006.

[12] W. Jiao, P. Jiang, and Y. Ma, “Fast Handover Scheme for Real-Time Applications in Mobile WiMAX,” Proc. IEEE Int’l Conf.Comm. (ICC), June 2007.

[13] S.W. Kim, “Spatial-Multiplexed Soft Handoff,” Proc. IEEE Conf.Wireless Comm. and Networking, Mar. 2007.

[14] L.G. de R. Guedes and M.D. Yacoub, “Overlapping Cell Area inDifferent Fading Conditions,” Proc. IEEE Vehicular TechnologyConf., July 1995.

[15] M.R.R. Kumar, S. Bhashyam, and D. Jalihal, “ThroughputImprovement for Cell-Edge Users Using Selective Cooperationin Cellular Networks,” Proc. IEEE/IFIP Conf. Wireless and OpticalComm. Networks, May 2008.

[16] L. Xu, K. Yamamoto, H. Murata, and S. Yoshida, “Adaptive BaseStation Cooperation and Subchannel Reallocation at Cell Edge inCellular Networks with Fractional Frequency Reuse,” Proc. IEEEConf. Personal Indoor and Mobile Radio Comm., Sept. 2009.

[17] M.M. Wang, T. Ji, J. Borran, and T. Richardson, “InterferenceManagement and Handoff Techniques in Ultra Mobile BroadbandCommunication Systems,” Proc. IEEE Int’l Symp. Spread SpectrumTechniques and Applications, pp. 166-172, Aug. 2008.

[18] M.C. Necker, “Interference Coordination in Cellular OFDMANetworks,” IEEE Network Magazine, vol. 22, no. 6, pp. 12-19, Nov./Dec. 2008.

[19] G. Boudreau, J. Panicker, N. Guo, R. Chang, N. Wang, and S.Vrzic, “Interference Coordination and Cancellation for 4G Net-works,” IEEE Comm. Magazine, vol. 47, no. 4, pp. 74-81, Apr. 2009.

[20] S. Shenker, C. Partridge, and R. Guerin, Specification of GuaranteedQuality of Service, IETF RFC 2212, http://www.ietf.org/rfc/rfc2212.txt, 1997.

[21] J. Chen, W. Jiao, and Q. Guo, “Providing Integrated QoS Controlfor IEEE 802.16 Broadband Wireless Access Systems,” Proc. IEEEVehicular Technology Conf. - Fall, Sept. 2005.

[22] M. Ma and B.C. Ng, “Supporting Differentiated Services inWireless Access Networks,” Proc. Int’l Conf. Comm. Systems, Oct.2006.

[23] H.-J. Yao and G.-S. Kuo, “An Integrated QoS-Aware MobilityArchitecture for Seamless Handover in IEEE 802.16e Mobile BWANetworks,” Proc. IEEE Military Comm. Conf. (MOLCOM), Oct.2006.

[24] D. Wright, “Maintaining QoS during Handover among MultipleWireless Access Technologies,” Proc. Int’l Conf. Management ofMobile Business, July 2007.

[25] H. Zhou and Z. Zhang, “Differentiated Statistical QoS Guaranteesfor Real-Time CBR Services in Broadband Wireless AccessNetworks,” Proc. Int’l Conf. Wireless Comm., Networking, and MobileComputing, Sept. 2010.

[26] S.-I. Choi and J.-D. Huh, “Dynamic Bandwidth AllocationAlgorithm for Multimedia Services over Ethernet PONs,” ETRIJ., vol. 24, no. 6, pp. 465-468, Dec. 2002.

[27] Y. Luo and N. Ansari, “Bandwidth Allocation for Multi-ServiceAccess on EPONs,” IEEE Optical Comm. Magazine, vol. 43, no. 2,pp. S16-S21, Feb. 2005.

[28] D. Nowak, J. Murphy, and P. Perry, “Bandwidth Allocation inDiffServ Enabled Ethernet Passive Optical Networks,” IET Comm.J., vol. 3, no. 3, pp. 391-401, Mar. 2009.

[29] S. De, V. Singh, H.M. Gupta, N. Saxena, and A. Roy, “A NewPredictive Dynamic Priority Scheduling in Ethernet PassiveOptical Networks,” Elsevier J. Optical Switching and Networking,vol. 7, no. 4, pp. 215-223, 2010.

[30] S. Sezginer and H. Sari, “Full Frequency Reuse in OFDMA-BasedWireless Networks with Sectored Cells,” Proc. IEEE WirelessComm. and Networking Conf. (WCNC), Apr. 2009.

[31] WiMAX Evolution: Emerging Technologies and Applications, M.D.Katz and F.H.P. Fitzek, eds. John Wiley, 2009.

[32] D. Wu and R. Negi, “Effective Capacity: A Wireless Link Modelfor Support of Quality of Service,” IEEE Trans. Wireless Comm.,vol. 2, no. 4, pp. 630-643, July 2003.

[33] M. Kazmi and N. Wiberg, “Scheduling Algorithm for HS-DSCH ina WCDMA Mixed Traffic Scenario,” Proc. IEEE Personal Indoor,Mobile Radio Comm., Sept. 2003.

[34] S. Haykin, Adaptive Filter Theory, fourth ed. Prentice Hall, 2001.


Chetna Singhal received the MTech degree incomputer technology from the Electrical Engi-neering Department, IIT Delhi, in 2010. She iscurrently working toward the PhD degree at theBharti School of Telecom, IIT Delhi. From June2010 to July 2011, she was with the IBMSoftware Lab, Gurgaon, as a software engineer.Her research interests include handoff studiesand cross-layer optimization in wireless net-works, multimedia multicast, and broadcast

technology. She is a student member of the IEEE.

Satish Kumar received the MS degree from theBharti School of Telecom, IIT Delhi, in 2011. Heis currently associated with Qualcomm India Pvt.Ltd. in the field of wireless technology. Hisresearch interests include cooperative wirelesscommunications, scheduling techniques inWCDMA and OFDMA networks, and LTE/WCDMA small cell optimization.

Swades De received the PhD degree inelectrical engineering from the State Universityof New York at Buffalo in 2004. He is currentlyan associate professor in the Department ofElectrical Engineering at IIT Delhi. He was anassistant professor in the Electrical and Com-puter Engineering Department at the NewJersey Institute of Technology from 2004-2007.He was a postdoctoral researcher at ISTI-CNR,Pisa, Italy (2004) and has nearly five years of

industry experience in India in telecom hardware and softwaredevelopment. His research interests include performance studies,resource efficiency in multihop wireless and high-speed networks,broadband wireless access, and communication and systems issues inoptical networks. He currently serves as an associate editor of the IEEECommunications Letters and Springer Photonic Network Communica-tions journal. He is a member of the IEEE.

Nitin Panwar received the MTech degree fromthe Bharti School of Telecom, IIT Delhi, in 2011.He is currently associated with Cisco SystemsIndia Pvt. Ltd. in the field of networking. Hisresearch interests include networking, handoffschemes in wireless networks, scheduling tech-niques in OFDMA based networks.

Ravindra Tonde received the BE degree inelectronics and telecommunications from PuneUniversity in 2008 and the MTech degree incomputer technology from the Electrical En-gineering Department, IIT Delhi, in 2010. He iscurrently with Samsung India Software Center,Noida, as a senior software engineer. Hisresearch interests include wireless communica-tion technologies, cloud and convergencetechnologies.

Pradipta De received the PhD degree fromStony Brook University. Currently, he is anassistant professor in the Department of Com-puter Science at the State University of NewYork—Korea. He was previously a research staffmember at IBM Research, India, New Delhi, from2005-2012, where he was associated with theTelecom and Mobile Research group dealingwith projects related to mobile-enabled financialservices and mobile cloud computing. He has

also worked on projects related to various aspects of data centermanagement and service delivery.

. For more information on this or any other computing topic,please visit our Digital Library at www.computer.org/publications/dlib.


Documents

48 IEEE TRANSACTIONS ON MOBILE COMPUTING, …web.iitd.ac.in/~swadesd/pubs/JNL/Handoff-TMC2014.pdf · Long Term Evolution-Advanced (LTE-A) and World-wide Interoperability for Microwave