Multi-objective handover in LTE macro/femto-cell networks

578 JOURNAL OF COMMUNICATIONS AND NETWORKS, VOL. 14, NO. 5, OCTOBER 2012

Multi-Objective Handover in LTE Macro/Femto-CellNetworks

Abhishek Roy, Jitae Shin∗, and Navrati Saxena

Abstract: One of the key elements in the emerging, packet-basedlong term evolution (LTE) cellular systems is the deployment ofmultiple femtocells for the improvement of coverage and datarate. However, arbitrary overlaps in the coverage of these femto-cells make the handover operation more complex and challeng-ing. As the existing handover strategy of LTE systems consid-ers only carrier to interference plus noise ratio (CINR), it oftensuffers from resource constraints in the target femtocell, therebyleading to handover failure. In this paper, we propose a new ef-ficient, multi-objective handover solution for LTE cellular sys-tems. The proposed solution considers multiple parameters like sig-nal strength and available bandwidth in the selection of the optimaltarget cell. This results in a significant increase in the handoversuccess rate, thereby reducing the blocking of handover and newsessions. The overall handover process is modeled and analyzedby a three-dimensional Markov chain. The analytical results forthe major performance metrics closely resemble the simulation re-sults. The simulation results show that the proposed multi-objectivehandover offers considerable improvement in the session blockingrates, session queuing delay, handover latency, and goodput duringhandover.

Index Terms: Femtocell, handover, long term evolution (LTE),Markov chain, multi-objective, queuing analysis.

I. INTRODUCTION

As shown in Fig. 1, third generation partnership project(3GPP) long term evolution (LTE) [1] cellular systems employthe presence of multiple femtocells (Home eNodeBs or HeNBs)arbitrarily overlapped with the current macrocells (eNodeBs,abbreviated to eNBs) for the purpose of improving the cover-age and data rates with lower service charges. The existing LTEconsists of eNBs, providing user and control plane protocol ter-minations for mobile user equipment (UE). The eNBs are alsoconnected to the mobility management entity (MME), and areresponsible for radio resource management, IP header compres-sion, and encryption of the user data stream, selection of anMME during UE attachment, routing of the user plane data to-ward the serving gateway, scheduling and transmission of broad-cast and paging messages (originating from the MME), andmeasurement reporting configuration for mobility and schedul-

Manuscript received November 2, 2011; approved for publication by RaoufBoutaba, Division III Editor, August 1, 2012.

This work was supported by a National Research Foundation (NRF) grant,funded by the Korea government (MEST) (2010-0016896).

A. Roy is with Samsung Electronics, e-mail: [email protected]. Shin and N. Saxena are with the College of Information and Communication

Engineering, Sungkyunkwan University, Suwon, 440-746, South Korea, e-mail:jtshin, [email protected].∗J. Shin is the corresponding author.Digital Object Identifier 10.1109/JCN.2012.00016

Fig. 1. An overview of the architecture of LTE femtocell systems.

ing. The MME is responsible for access security control, nodesignaling for mobility between different 3GPP LTE access net-works, paging, tracking area (location) list update, gateway se-lection, roaming, authentication, and data plane establishment.

The increasing demand for indoor wireless multimedia, andongoing trends for mobile convergence, are paving the wayfor the industry-wide deployment of femtocells. These femto-cells can be open access or closed access [2]. Open access al-lows an arbitrary user to use the femtocell, whereas closed ac-cess restricts the use to the users explicitly approved by theowner. While the ultimate goal of femtocells is to improve theefficiency, coverage and services at a reduced operation cost,the possibility of arbitrary handover between the existing eNB,and an indoor, open access HeNB, raises significant new chal-lenges. The major technical challenges associated with the massdeployment of femtocells are interference to/from other femto-cells and macrocells, self-organization/configuration of the fem-tocells, handover control among the femtocells and a macrocell,timing and synchronization, and security and access control [3],[4]. Among these issues, this paper focuses on handover con-trol. For seamless connectivity with precise data rates, the op-timal selection of the target femtocell (or macrocell) for han-dover is of the utmost importance. Unfortunately, the existinghandover solution in LTE systems is far from optimal. This han-dover operation only considers the signal strength of the neigh-boring eNBs/HeNBs (eNBs or HeNBs) and selects the eNB orHeNB with the maximum signal strength for hand over of themobile device. This often results in degradation of system per-formance on or after the handover operation, thereby resultingin blockage in the handover session and admission failure.

In this paper, we design a new, efficient, and multi-objectivesolution for handover from the source eNB to target eNB/HeNBin emerging LTE systems. To utilize the increased spectrum effi-ciency, we explore complete sharing [5], where the entire spec-

1229-2370/12/$10.00 c© 2012 KICS

ROY et al.: MULTI-OBJECTIVE HANDOVER IN LTE MACRO/FEMTO-CELL NETWORKS 579

trum band is shared by femtocells and macrocells. We also con-sider open access femtocells, as only open access femtocells al-low handovers from arbitrary outside users, and network oper-ators prefer an open access deployment as it provides an inex-pensive way to expand their network capabilities [2]. More pre-cisely, our contributions include the following

• While the existing handover operation in LTE systems onlyconsiders signal strength, the basic novelty of our work isin the optimization of the handover operation by exploringmultiple objectives, such as the signal strength and avail-able wireless resources (bandwidth). To this extent, we usea fixed total spectrum consisting of a number of resourceblocks (RB), where a RB of LTE is the smallest time fre-quency resource consisting of 12 subcarriers that can beallocated to a user. For example, a channel bandwidth of20 MHz and 10 MHz contains a total number of 100 and50 RBs [1], respectively.

• We show that such handover from source eNB to tar-get eNB/HeNB in emerging LTE systems can be effi-ciently modeled by a discrete 3-dimensional (3-D) Markovchain. Using this model, the relevant system performancemetrics, such as the new and handover session blockingprobabilities and mean queuing delay of new sessions canbe obtained.

• The simulation results validate the performance model-ing and analysis and demonstrate that the proposed multi-objective handover strategy has the potential to signifi-cantly reduce the handover blocking probability in compar-ison to the existing handover operations in LTE systems.

The rest of the paper is organized as follows. Section II reviewsrelated works on mobility management. The proposed multi-objective handover solution is discussed in Section III. Subse-quently, the overall performance of the handover operation ismodeled and analyzed in Section IV. The simulation resultsin Section V corroborate the analysis and demonstrate the ef-ficiency of the proposed solution in reducing the handover, aswell as for new session blocking, while improving the new ses-sion queuing delay. Section VI concludes this paper with someindications on future work.

II. RELATED WORK

Given the wide variety of research proposals pertaining tothe handover process, we discuss the major works related tonext generation wireless systems in general and LTE systemsin particular. The on-going standardization activities in IEEE,IETF, and 3GPP toward seamless homogeneous and heteroge-neous handover support are discussed in [6]. An end-to-endquality of service (QoS) architecture and a mobility-aware reser-vation signaling protocol for seamless handover support in next-generation, IP-based wireless networks are proposed in [7]. Onthe other hand, the effectiveness of the IEEE 802.21 frameworkand media-independent pre-authentication technique in improv-ing the handover performance is shown in [8]. The effectivenessof definitive layer-2 (L2) triggers in reducing latency and packetloss, associated with fast handovers for mobile IPv6 (FMIPv6),is discussed in [9]. Recent research activities have pointed outthe necessity and efficiency of vertical handover [10] in het-

erogeneous wireless networks (e.g., UMTS and WLAN) withproactive, congestion-aware [11] strategies. An overview of theLTE handover process is provided in [12] and [13], with asubsequent study of the impact of connection-forwarding, out-of-order delivery and HARQ/ARQ state discard. An inter-cellinterference mitigation scheme using a combination of partialreuse and soft inter-sector handover is discussed in [14]. Recentsimulation experiments [15], [16] have pointed out the optimalhandover parameters required for achieving a suitable compro-mise between the average number of handovers and averageuplink carrier to interference plus noise ratio (CINR) for cer-tain user speeds. A cross layer framework for handover predic-tion [17] was also proposed to improve the LTE handover perfor-mance. The handover decision depends on various parametersor handover metrics including the available bandwidth, delay,jitter, access cost, transmit power, current battery status of themobile device, and even the user’s preferences. In [18], a mo-bility management scheme using a profile database (PDB) [18]is developed to best serve the application and user require-ments. PDB-based handover collects dynamic status informa-tion from all protocol layers for cross-layer techniques, and thePDB maintains both the static and the dynamic information nec-essary for handover-related decisions on a per-application ba-sis. However, this per-application mobility management can bein conflict with different requirements among diverse applica-tions, and the information needed from higher protocols requiresmuch processing time, so these kinds of complex platforms maybe impractical to deploy.

A close look at the existing LTE handover procedure [13],[16] and associated research reveals that the handover processsuffers from several drawbacks which increase in importancewith increasing numbers of eNBs and HeNBs. The reason forthis is that an HeNB or an eNB lying in close proximity to theUEs can offer a very high CINR. Based on the existing han-dover decision process, the source eNB/HeNB can select any ofthese eNBs or HeNBs as the possible target eNB/HeNB for han-dover. However, the particular target eNB/HeNB can be over-loaded by a number of UEs, thereby resulting in resource un-availability, which leads to service degradation and blockage af-ter the handover process. Thus, with the emerging LTE specifi-cations, the existing handover procedure needs to be enhancedto consider other necessary parameters other than the currentCINR ratio. This motivates us to investigate the handover prob-lem in LTE systems with multiple objectives in mind.

III. PROPOSED MULTI-OBJECTIVE HANDOVER

Many real world problems require the simultaneous opti-mization of multiple objectives, which is quite different froma single-objective optimization. There may not even be a singlebest solution with respect to all of the objectives under consid-eration. Instead, there might exist a set of solutions superior tothe rest in the entire search space. Such a solution set is referredto as Pareto-optimal [19] and non-dominated. Since none of thesolutions in this set is absolutely better than any other, there isfreedom to choose the best possible solution that conforms tothe application specific requirements.

580 JOURNAL OF COMMUNICATIONS AND NETWORKS, VOL. 14, NO. 5, OCTOBER 2012

Dominated

Non-dominated

Global optima

solutions

solutions

X (minimize)

Y (

min

imiz

e)

Global optima

Fig. 2. Pareto-optimal solutions.

• Definition: A point x∗ is Pareto-optimal if, for every x, ei-ther ∩i(fi(x) = fi(x

∗)) or there is at least one i such thatfi(x) > fi(x

∗), ∀i ∈ I (set of integers), where fi(x) is thefitness function. In other words, x∗ is Pareto-optimal if thereexists no feasible vector x which would decrease some crite-rion without causing a simultaneous increase in at least oneother criterion.As shown in Fig. 2, Pareto optimality can be visualized in a

scatter-plot of solutions, where each criteria (or objective func-tion) is plotted on a separate axis. In a problem with two cri-teria, both of which are to be minimized, Pareto-optimal (non-dominated) solutions are those in the scatter-plot with no pointsdown and to the left of them. We argue that, for efficient han-dover in LTE systems, Pareto optimal target eNBs need to beobtained.

A. Handover Decision Problem Formulation

The handover process in existing LTE systems is shown inFig. 3 [1]. Upon instruction from the source eNB/HeNB, theUE periodically sends the measurement report (MR) in the uplink to the source eNB/HeNB. If the MR contains any targeteNB/HeNB with a CINR better than a pre-defined thresholdin comparison to the CINR of source eNB/HeNB, the sourceeNB/HeNB sends a “handover request” message to that targeteNB/HeNB. The target eNB/HeNB responds with the “handoverack” message. In the handover decision process, the sourceeNB/HeNB now selects the best target eNB/HeNB and informsit to the UE by downlink “handover command” message. TheUE can now switch its communication to the selected tar-get eNB/HeNB and sends the uplink “handover confirm” mes-sage. After switching the data path from source eNB/HeNBto target eNB/HeNB, the source eNB/HeNB releases the re-sources. We use the same message exchange, but propose thatthe network load and available radio RBs be considered, as wellas the existing CINR, in selecting the target eNB/HeNB. Thehandover ack message can be easily modified to include theavailable RBs of the corresponding target eNB/HeNB. In the restof this subsection, we discuss how to probabilistically model thehandover decision parameters.

1. Like existing systems, the handover in LTE should con-sider the CINR. The UE measures the reference signalreceived power (RSRP), which includes the pathloss, an-tenna gain, log-normal shadowing and fast fading, aver-aged over all the reference symbols within the measure-ment bandwidth. The downlink-received RSRP and the

Fig. 3. Handover process in LTE.

CINR from the kth cell are respectively estimated as fol-lows

RSRPk = P∑

j∈all symbols

Gk,j

fk(P ,G) = CINRk =RSRPk

∑

i∈C;i6=k

RSRPi +N0(1)

where N0, C, P , and Gk,j represent the thermal noise, setof all cells, downlink received power and estimated chan-nel gain for the jth symbol of the kth eNB, respectively.

2. The second optimization parameter is the availableradio resources or wireless bandwidth in the targeteNB/HeNB. As mentioned before, this is estimated bythe RBs available in the target eNB. Generally, the targeteNB/HeNB capable of offering the maximum available re-sources, (i.e., the target eNB/HeNB with maximum avail-able RBs) should be selected as the best choice. The totalnumber of available RBs in the target eNB/HeNB after al-locating RBs for all of the UEs currently under its cover-age is given by (φ−

∑ηi=1 βi), where φ is the total number

of RBs of the target eNB/HeNB, βi is the RBs consumedby the ith UE, and η is the number of active UEs in thekth target eNB/HeNB. Thus, the fraction of the total RBsavailable is given by

gk(η, β) =

φ−η∑

i=1

βi

φ. (2)

The distribution of this RBs (β =∑η

i=1 βi) varies de-pending on the application. While wireless voice calls areoften modeled using a memory-less Poisson distribution,data traffic is modeled using a heavy-tailed Pareto distri-bution. Note that, for a particular channel bandwidth thetotal number of RBs (φ) is fixed for every eNB or HeNBacross the entire frequency. Depending on it’s current cell-load and channel conditions, every eNB or HeNB allocatesa portion of these RBs amongst different users and can in-dependently estimate gk(η, β).

It is now clear that the optimal selection of the target eNBneeds to satisfy both of the above-mentioned parameters and

ROY et al.: MULTI-OBJECTIVE HANDOVER IN LTE MACRO/FEMTO-CELL NETWORKS 581

Fig. 4. Flowchart of proposed handover strategy.

constraints. Hence, the optimization problem can be formallystated as

OPT : max∀k∈Ω

[fk(P ,G), gk(η, β)] (3)

where Ω represents the set of all eNBs/HeNBs.

B. Solving the Handover Decision Problem

As mentioned in [20], the most intuitive approach to solveany multi-objective optimization problem is to construct a singleaggregate objective function (AOF). Instead of using complexheuristics to solve this multi-objective optimization problem, weprefer the AOF based approach, for its easy implementation inreal LTE systems, without significant computational complex-ity and storage overhead. The basic idea is to combine all ofthe objectives into a single functional form using the weightedlinear sum. Thus, the optimization problem OPT can be solvedby assigning two weights (coefficients) ζ1 and ζ2 for objectivefunctions fk(P ,G) and gk(η, β), respectively. By varying ζ1 andζ2, different Pareto optimal solutions can be obtained. Now, asboth the weights ζ1 and ζ2 are constants, the problem can bereduced as another maximization problem with weight coeffi-cients 1 and ξ = ζ1/ζ2 without loss of generality. Thus, wecan reduce the optimization problem into a weighted single-objective maximization problem as shown below

max∀k∈Ω

[fk(P ,G) + ξgk(η, β)] . (4)

The weighting coefficient represents the weight assigned by aservice provider or network operator to the available bandwidthcompared to the traditional signal strength CINR. A larger orsmaller value of ξ indicates that the service provider prefersthe bandwidth availability or signal strength CINR, respec-tively. Those solutions offering the maximum CINR and maxi-mum available bandwidth can be obtained by calculating respec-tive partial differentiations (if they exist), i.e., ∂fk(P ,G)/∂P∂Gand ∂gk(η, β)/∂η∂β. Apart from the HeNB/eNB offering max-imum CINR and available bandwidth, the Pareto optimal set

(Ψ) could include some other HeNBs/eNBs offering compro-mise solutions. HeNBs/eNBs providing compromise solutionsare not worse than the HeNB/eNB offering maximum CINRand maximum available bandwidth, when both the objectivesare considered. For example, there could be some HeNB/eNBthat offers a better available bandwidth than the HeNB/eNB pro-viding maximum CINR. The network operator can select thefinal target eNB/HeNB from the Pareto optimal set by tuningthe parameter ξ. Hence the computation complexity of the algo-rithm is O(|ξ|χ), where χ represents the total number of eNBsand HeNBs and |ξ| is the cardinality of ξ. With a huge numberof femtocells, as χ ≫ |ξ|, the complexity approaches towardsO(χ), i.e., almost linear with the number of eNBs/HeNBs. Inorder to confirm the convergence of our algorithm to optimality,we refer to an important theorem in [21], which states that so-lution of the weighted criteria method problem is sufficient forPareto optimality. Fig. 4 shows the overall flow of the proposedstrategy, which is executed during the handover decision processshown in Fig. 3 in source eNB/HeNB.

IV. MODELING AND PERFORMANCE ANALYSIS

Once the source eNB selects the target eNB/HeNB, we eval-uate and analyze the handover performance metrics. In this sec-tion, we first use a 3-D Markov chain to model the handoverprocess with the selected target eNB/HeNB. Subsequently, weanalyze different performance metrics related to the handoverprocess.

A. System Model

An HeNB (femtocell) is assumed to be on the boundary be-tween the source eNBs (macrocell) under consideration and itsneighboring eNBs/HeNBs. It covers a fraction of the handoverregion of the source eNB and shares the handover traffic be-tween the source eNB and its neighbors. As shown in Fig. 5,from the viewpoint of the source eNB, incoming and outgo-ing handovers refer to the ongoing sessions entering and leav-ing the source eNB from/to any of its neighboring cells, re-spectively. Blocked outgoing handovers will not be counted inthe following analysis of the handover blocking probability ofthe cell of interest. Note that, as the source eNB consideredis macro eNB, new sessions are always attached to this macro(source) eNB. Hence, the complementary problem involvingfemto HeNB as the handover source is not included in this anal-ysis. For convenience, we refer to the data channels of the HeNBand eNB as the femto data channel (FDCH) and macro datachannel (MDCH), respectively. As mentioned before in Sec-tion I, to achieve better cellular coverage, higher data rates andlower service charges, the HeNBs are preferred over the macroeNBs to be selected as the target eNB, provided coverage anddata channels are available. In other words, incoming and out-going handovers can use the bandwidth of the HeNB, as longas the handover region is covered by the HeNB, and an avail-able FDCH can be found in this HeNB. When a UE undergoesa handover operation from the source eNB:

1. It will be assigned an FDCH if it passes through theHeNB-covered handover region and if any FDCH is avail-able; otherwise,

582 JOURNAL OF COMMUNICATIONS AND NETWORKS, VOL. 14, NO. 5, OCTOBER 2012

Fig. 5. LTE system model for handover analysis.

2. the UE will be assigned an MDCH if there is any MDCHavailable in the macrocell; otherwise,

3. this ongoing handover session will be blocked.On the other hand, upon a new session arrival in the source eNB:

1. An MDCH will be assigned, if any MDCH is available;otherwise,

2. the session request will be waiting in the queue, if thequeue is not full; otherwise,

3. the new incoming session will be blocked.Let H out of a total M available MDCHs of macro eNB be

reserved exclusively for handover sessions. Upon the arrival ofa new session, if the total number of MDCHs being used is lessthan M −H , an MDCH will be assigned; otherwise, the sessionrequest will be entered into the queue or blocked depending onwhether or not the queue is full. It is assumed that a UE witha new session request waiting in the queue does not move fromone eNB to another. Both new sessions and session handover areassumed to be generated according to Poisson distributions withaverage rates of λn and λo, respectively. It is assumed that thelocation of the UE’s handover is uniformly distributed over theentire handover region. Thus, a handover session is generatedwithin the HeNB’s handover region with probability c.

The session duration is assumed to be exponentially dis-tributed with mean Tc = 1/µ, where µ is the session servicerate. The session dwell times in a macro eNB and in the femtoHeNB are also assumed to be exponentially distributed withmean Tdu and Tdw, respectively. In the worst case, assumingthat the UE’s movement is random at any time instant, a UEwith an ongoing session using an MDCH moves away from theeNB at rate µu = 1/Tdu, and a UE with an ongoing sessionusing an FDCH of the HeNB moves out of the HeNB coveragearea at rate µw = 1/Tdw.

B. Markov Chain Representation

The behavior of the LTE cellular system described above canbe characterized by a 3-D discrete Markov chain. The major pa-rameters and symbols used in our modeling and analysis arelisted in Table 1. A state in the 3-D Markov chain is definedas

(m,n, b), 0 ≤ m ≤ M, 0 ≤ n ≤ N, 0 ≤ b ≤ B (5)

where m and n are the numbers of MDCHs and FDCHs beingused, respectively, and b is the number of new session requestswaiting in the queue. Since a new session request will not wait

Fig. 6. An example of a 3-D Markov chain handover model (M = 3,N = 3, H = 1, and B = 2).

Table 1. Major modeling parameters and symbols.Symbol Explanationλn New session arrival rateλo Handover session rateµ Session service rateµw Handover service rate from HeNBµu Handover service rate from eNBM Total no. of MDCHH No. of MDCH reserved for handoverc Ratio of handover regions of HeNB and eNB

Pr[Bnew] New session blocking probPr[BHO] Handover session-blocking prob

in the queue when the number of free MDCHs is greater thanthose reserved for handover, states (m,n, b) with m < (M−H)and b > 0 do not exist. Thus, the total number of states in theMarkov chain, Snum = [M + 1 + B(H + 1)]·(N + 1). An ex-ample with M = 3, N = 3, H = 1, and B = 2 is shownin Fig. 6. Four different types of events can cause state transi-tions: (i) New session arrival, (ii) session completion, (iii) out-going session handover, and (iv) incoming session handover. Bydefining the traffic intensity as the ratio of the session arrivalrate and the service rate, we can estimate the normalized traf-fic intensities of new and handover sessions, Tn and To, asTn = λn/µ and To = λo/µ, respectively. Similarly, we definethe normalized handover rates in eNB and HeNB as αu = µu/µand αw = µw/µ, respectively. Now, when the system traf-fic is statistically stable, the traffic intensity of the incominghandovers is equal to that of the outgoing handover calls, i.e.,To = αuTn. Fig. 7 shows the expressions of the state transi-tion probabilities as functions of the traffic intensities. The statetransitions from states (mi, ni, bi) to (mj , nj, bj) are explainedbelow

1. When the number of MDCHs used is less than those re-served for handover, i.e., when mi < H .(a) Upon the completion of a session using an MDCH or

an FDCH, an MDCH or FDCH is released, i.e., nj =ni − 1 or mj = mi − 1.

(b) Upon an incoming handover outside the HeNB cover-age, an MDCH is assigned, i.e., nj = ni and mj =mi + 1.

(c) Upon an incoming handover session inside the HeNBcoverage area, an available FDCH (if found) is as-signed, i.e., nj = ni + 1 and mj = mi; otherwise,

ROY et al.: MULTI-OBJECTIVE HANDOVER IN LTE MACRO/FEMTO-CELL NETWORKS 583

Fig. 7. State transition probabilities (For the cases in Fig. 5).

an MDCH is assigned, i.e., nj = ni and mj = mi+1.(d) Upon an outgoing handover session using an MDCH,

the MDCH is released, i.e., mj = mi − 1; and anavailable FDCH (if found) is assigned, if the handoverpasses through the HeNB covered handover region,i.e., nj = ni + 1.

(e) Upon a session handover using an FDCH, the FDCH isreleased, i.e., nj = ni − 1; and an MDCH is assignedif the corresponding UE moves into the intended eNB,i.e., mj = mi + 1.

2. When the number of MDCHs used is greater than that re-served for handover, i.e., when H ≤ mi ≤ M .(a) Upon the arrival of a new session, the session request

needs to wait in the queue if the queue is not full, i.e.,bj = bi + 1.

(b) Upon the completion of a session using an FDCH, theFDCH is released, i.e., nj = ni − 1.

(c) Upon the completion of a session using an MDCH, theMDCH is released. If the number of available MDCHsis not larger than H , an MDCH is assigned to a newsession waiting in the queue, i.e., if mi − 1 ≤ H , bj =bi − 1, and mj = mi; otherwise, bj = bi and mj =mi − 1.

(d) Upon an incoming handover outside the HeNB-covered handover area, an available MDCH (if found)is assigned, i.e., mj = mi + 1.

(e) Upon an incoming handover inside the HeNB-coveredhandover area, an available FDCH (if found) is as-signed, i.e., nj = ni + 1 if ni < N ; otherwise, anMDCH is assigned, i.e., mj = mi + 1.

(f) Upon an outgoing handover using an MDCH, theMDCH is released and an available FDCH (if found)is assigned if the handover is inside the HeNB-coveredhandover region, i.e., nj = ni + 1. If the number ofavailable MDCHs is not larger than H , a MDCH isassigned to a new session waiting in the queue, i.e.,bj = bi − 1 and mj = mi if mi − 1 ≤ H ; otherwise,bj = bi and mj = mi − 1.

(g) Upon a handover from HeNB, an FDCH is released,i.e., nj = ni − 1. An available MDCH (if found) isassigned if the handover moves from the HeNB to theintended eNB, i.e., mj = mi − 1.

Let Pr(m,n, b) be the probability that the system status re-sides in state (m,n, b). The flow equilibrium equation for state

(m,n, b) can be formed by equating the flux out of this stateto the flux into this state. There are a total of Snum − 1 lin-early independent equilibrium equations. The conservation rela-tion among all state probabilities is given as follows

M∑

m=0

N∑

n=0

B∑

b=0

Pr(m,n, b) = 1. (6)

We collect all the state probabilities into a column vector P

through mapping between the state index (m,n, b) and vectorindex i, i = 1, 2, · · ·, Snum, so that the ith entry of P is theprobability of state (mi, ni, bi), i.e., Pi = Pr(mi, ni, bi). Col-lecting the equilibrium equations for all the states and writingthem in vector-matrix form, we obtain the linear equation

(X−Y)P = 0 (7)

where X and Y are the output and input transition probabilitymatrices, respectively; such that every element, x(i, i), of the di-agonal matrix X represents the rate flux out of state (mi, ni, bi),and every element, y(i, j), of matrix Y provides the transitionrate from state (mj , nj, bj) to (mi, ni, bi). The Markov chaincan be solved to obtain P constrained by (6). The individualstate probabilities and corresponding performances can now beestimated. Note that the above model is general and can be usedto obtain the state probabilities employing different LTE han-dover schemes. Only the state transition probabilities need to bemodified accordingly.

C. Performance Analysis

After obtaining the individual state probabilities of the 3-DMarkov chain, we analyze the relevant system performance.

C.1 New Session Blocking Probability

A new session is blocked if the number of available MDCHsin the eNB of interest is less than that reserved for session han-dover, i.e., m ≥ M −H , and the queue is full. By obtaining thestate probabilities, the blocking probability of a new session,Pr[Bnew], can be estimated as

Pr[Bnew] =

M∑

m=M−H

N∑

n=0

Pr(m,n,B). (8)

C.2 Handover Blocking Probability

The handover session blocking probability is analyzed fromthe viewpoint of the eNB of interest. When a handover sessionis generated outside the HeNB’s coverage area with probability1−c, it is blocked if there is no available MDCH in the intendedmacro eNB. Hence, the blocking probability of such a handoversession, Pr[BHO1

], is given by

Pr[BHO1] =

N∑

n=0

B∑

b=0

Pr(M,n, b). (9)

On the other hand, when an ongoing session enters the eNBof interest through the handover region covered by an HeNBwith probability c, this handover session is assigned an FDCH if

584 JOURNAL OF COMMUNICATIONS AND NETWORKS, VOL. 14, NO. 5, OCTOBER 2012

any available FDCH can be found with probability Pr[FDCH]otherwise, a MDCH of the intended macrocell is assigned (ifavailable); otherwise this handover is blocked with probabilityPr[BHeNB]. An incoming handover being assigned an FDCH isblocked when the corresponding UE moves out of the HeNB’scoverage area to the intended macro eNB, with probabilityPeNB, and no MDCHs of the intended macro eNB are availablewith probability Pr[MDCH]. Assume that the session stays inthe HeNB for a sufficient period of time such that the states inwhich the system resides when it enters and leaves the HeNBare independent. Hence, the blocking probability of such an in-coming handover session, Pr[BHO2

], can be obtained as

Pr[BHO2] = Pr[BHeNB] + Pr[FDCH]Pr[eNB]Pr[MDCH]

(10)where Pr[BHeNB], Pr[FDCH], Pr[eNB], and Pr[MDCH] canbe estimated as

Pr[BHeNB] =B∑

b=0

Pr(M,N, b), (11)

Pr[FDCH] =M∑

m=0

N−1∑

n=0

B∑

b=0

Pr(m,n, b), (12)

Pr[eNB] =αw

αw + 1, (13)

Pr[MDCH] =

∑Nn=0

∑Bb=0 Pr(M,n, b)

∑Mm=0

∑Nn=0

∑Bb=0 Pr(m,n, b)

. (14)

Hence, the blocking probability of a handover session fromthe point of view of the eNB of interest, Pr[BHO], is now esti-mated as

Pr[BHO] = (1− c)Pr[BHO1] + cPr[BHO2

]. (15)

C.3 Mean Session Queuing Delay

We assume that, when the concerned new session is gener-ated, b new session requests are waiting in the queue and m MD-CHs are being used, i.e., the system state in which this new ses-sion enters upon its arrival is (m,n, b), n ≤ N . The position inthe queue of this new session can be moved forward from (b+1)to b only when the number of available MDCHs is less than(M−H). Thus, this new session can not be served during time tif the number of session completions exceeds the number of han-dover session arrivals by no more than m−(M−H−1)+b. LetKC(t) and KA(t) be the number of sessions completed and thenumber of handover sessions accepted during time t, respec-tively. The probability of the queuing delay of this new sessionrequest, Tm,b, being larger than t, Pr[Tm,b > t], can be ex-pressed as

Pr[Tm,b > t]

= Pr[KC(t)−KA(t) ≤ m− (M −H − 1) + b]

=

∞∑

Ka=0

Pr[KA(t) = ka]

·

ka+m−(M−H−1)+b∑

kc=0

Pr[KC(t) = kc]. (16)

Since the exact close-form of the queuing delay is difficultto derive, the upper bound can be readily estimated by the fol-lowing suitable approximations. We approximate the handoversession arrival rate seen by the eNB of interest, λeNB, as λo be-cause, according to the state diagram shown in Fig. 5, the exactinstantaneous value of λeNB is given by

λeNB =

(1− c)λo + nαw/2 if n < N,

λ0 if n = N.(17)

Clearly, the estimation of the above expression depends on thevalue of n. When the queue is nonempty, m, i.e., the number ofMDCHs being used instantaneously, varies in the range betweenM −H and M . We approximatem as M −H when M >> H .Since the handover blocking probability is small, the number ofhandover calls being served can be approximated well by thenumber of handover session arrivals. According to the aboveapproximations, the probabilities of kC session completions andkA handover session arrivals during time t, i.e., Pr[KC(t) =kC ] and Pr[KA(t) = kA], respectively, are obtained as follows

Pr[KC(t) = kC ] =(λot)

kC

kC !exp(−λot),

P r[KA(t) = kA] =

[

(M −H)(1 + αu)µt]kA

kA!

·exp[

(M −H)(1 + αu)µt]

.

With the knowledge of Pr[KC(t) = kC ] and Pr[KA(t) =kA] and then Pr(Tm,b > t), the mean queuing delay of the newsession request, Tm,b, can be obtained as

Tm,b =

∫ ∞

0

1− Pr[Tm,b > t]dt. (18)

The summation over all m, M −H ≤ m ≤ M and b, b < B,yields the mean queuing delay of a session waiting in the queue,i.e.,

T =

M∑

m=M−H

B−1∑

b=0

Qm,bTm,b

PQn

(19)

where PQn=

∑B−1b=0

∑Nn=0

[∑M

m=M−H Pr(m,n, b)]

is theprobability that a new session request enters the queue upon itsarrival, and Qm,b =

∑Nn=0 Pr(m,n, b) is the probability that

m MDCHs are used and b session requests are waiting in queue.

V. PERFORMANCE EVALUATION

We now present the simulation results obtained using themodules developed based on an optimized network engineer-ing tools (OPNET) simulation environment [22] for the pur-pose of studying the newly proposed handover process in LTEsystems. An LTE cellular system with different macro eNBsand femto HeNBs constitutes the heart of the simulation en-vironment. While mobility is instantiated by idle and moveevents, communication is managed by session-start and session-terminate events. Before discussing the simulation results, wedescribe the various parameters used in our study.

ROY et al.: MULTI-OBJECTIVE HANDOVER IN LTE MACRO/FEMTO-CELL NETWORKS 585

A. Simulation Parameters

A typical integrated LTE system with 19 macro eNBs havinga hexagonal layout and 1000 HeNBs with arbitrary coverage issimulated. The session arrival process is Markov with a Poissonarrival rate of 0.3 calls/h, and exponentially distributed holdingtimes with mean 1/µ (= 10 min) and variance σ (= 3 min). The“COST231 Hata urban propagation model” [23] is assumed forpathloss with 10 dB log-normal shadow fading. On the otherhand, for the femtocells, we used the “ITU indoor propagationmodel,” also known as the “ITU model for indoor attenuation.”This is a radio propagation model that estimates the path lossinside a room or a closed area inside a building delimited bywalls of any form and is suitable for appliances designed forindoor use. In order to simulate mobility, we used the randomdirection model [24], a widely used mobility model to simulateuser mobility. The time spent by the user in every cell is ex-ponentially distributed. As the down-link (DL) CINR generallylies between [0, 30] dBm or [1, 1000] mWatts, fj and gj in (4)lie in the ranges of [1, 1000] and [0, 1], respectively. Hence, inorder to get a fair measure of (4), the value of ξ is varied inthe range [1, 1000]. Simulation experiments are carried over aperiod of 10 weeks with a total 50, 000 handover samples andthe average results with 98% confidence interval are reportedhere. The other system parameters are assumed to be M = 12,c = 0.3, and αu = 0.3. We considered three different perfor-mance metrics to compare the proposed handover strategy withthe existing handover schemes. These three performance metricsare the: (a) Handover latency (b) goodput, and (c) session block-ing during handover. While the latency is one of the establishedperformance metrics for QoS traffic such as voise over internetprotocol (VoIP), the goodput demonstrates the performance ofbest effort (BE) TCP flows during handover. Using these threemetrics, the comparative performances of the proposed strategywith PDB-based handover [18] and existing handover in the LTEsystems [13], [16], having handover based only on the CINR,are presented.

B. Simulation Results

Fig. 8 demonstrates the set of five Pareto-optimal targeteNBs/HeNBs (Ψ) that are generated. For different values of ξ,the handover decision criteria fj + ξgj are varied for differenttarget eNBs/HeNBs. This provides the network operator ampleflexibility to choose the near-optimal target eNB/HeNB by con-trolling ξ. By tuning ξ, the network operator can obtain the mostsuitable target eNB/HeNB from the set Ψ and provide it to theUE for handover. Alternatively, the operator can also allow theUE to receive the entire set of target eNBs/HeNBs (Ψ) and letthe UE select the final target eNB/HeNB from this set. Figs. 9and 10 show the session blocking probability of the new andhandover sessions, respectively, for different values of N , H ,and B. With the additional bandwidth offered by the HeNB, thesession handover blocking probability is reduced. Meanwhile,the total traffic intensity including handover and new sessiontraffic intensities seen by the eNB is reduced. Hence, the newsession blocking probability is improved. The improvement ofthe new session blocking probability is, however, less than thatof the handover blocking probability. For instance, as shown in

Fig. 8. Dynamics of Pareto-optimal target eNBs/HeNBs.

Fig. 9. New session blocking probabilities with N (total FDCHs), H (re-served MDCH for handover), and B (total new sessions waiting inqueue).

Fig. 10. Handover session blocking probabilities with N , H, and B.

Figs. 9 and 10, the blocking probabilities of a new session anda handover are reduced by 15% and 39% respectively, when thesession arrival rate is 10 calls/h, H = 2 and B = 2. In addi-tion, as can be observed in Figs. 9 and 10, with increasing H ,the blocking probability of a handover session can be reduced atthe cost of a slight increase in the blocking probability of a newsession.

Fig. 11 shows the mean queuing delay of a new session re-quest waiting in the queue for each of these three handoverstrategies. It shows that, in comparison to the existing LTE han-

586 JOURNAL OF COMMUNICATIONS AND NETWORKS, VOL. 14, NO. 5, OCTOBER 2012

0 2 4 6 8 10 120

10

20

30

40

50

60

70

80

Session arrival rate

Qu

eu

ing

de

lay

(m

s)

Simulation results

Analytical results

PDB−based approach

Existing LTE systems

Fig. 11. Queuing delay dynamics.

Fig. 12. Comparison of session blocking probabilities.

dover and PDB-based handover processes, the proposed han-dover process offers improvements of almost 40% and 25%, re-spectively, in terms of the queuing delay for new sessions. Alsothe analytical result of the proposed one is added to show itsgood tracking to the simulation result.

Fig. 12 demonstrates the comparative session blocking proba-bilities for the proposed multi-objective handover process, PDB-based handover scheme and existing LTE handover procedure. Itcan be seen that, with increasing session arrival rates (callsper hour), the proposed handover process reduces the block-ing probability to about 1/3 that of the existing handover pro-cess and about 1/2 that of the PDB-based handover strategy, forboth new and handover sessions. This is a direct consequenceof the improved handover process based on multiple objectives(signal strength as well as available bandwidth). The optimiza-tion of the available bandwidth, together with the CINR, effec-tively ensures the availability of sufficient resources in the targeteNB/HeNB, thereby guaranteeing handover success.

The handover latency is one of the key performance metricsfor QoS-based VoIP sessions. In order to compare the handoverlatency, we demonstrate the relative performances of the pro-posed scheme in Fig. 13, PDB-based strategy and existing LTEhandover with increasing UE speed. It is quite clear that, withincreasing speed, the handover latency initially increases andthen becomes saturated for all of the handover strategies. How-ever, the proposed handover process significantly outperforms

Fig. 13. Comparison of handover latency.

Fig. 14. Comparative TCP traffic goodput during handover.

both the PDB-based handover scheme and the existing LTE han-dover scheme. While the PDB based handover scheme results ina handover latency of 2.5 s for a speed greater than 50 km/h,the proposed scheme offers a handover latency bounded onlyby 1.5 s. The existing LTE handover scheme results in a muchhigher latency of 3.5 s. From the set of non-dominated tar-get eNBs/HeNBs offered by the proposed strategy, the serviceprovider can select the target eNB/HeNB offering the best han-dover latency by tuning the attribute ζ. This is the major reasonfor the gain achieved by the proposed handover solution.

In order to compare the performance of BE transmission con-trol protocol (TCP) traffic, Fig. 14 demonstrates the comparativegoodputs of a single downlink TCP session for the different han-dover schemes. This figure shows that, during the handover pro-cess, with increasing speed, the TCP goodput is initially reducedand then becomes saturated. However, it is clear that the multi-objective handover scheme offers a TCP goodput of 400 Kbps,while the goodputs offered by the PDB-based handover and ex-isting LTE systems are 350 Kbps and 250 Kbps, respectively. Ina similar manner to the handover latency, from the set of non-dominated target eNBs/HeNBs offered by the proposed strategy,the service provider can select the target eNB/HeNB offering thebest TCP goodput by tuning the attribute ζ. As in the case of ses-sion blocking, handover latency and goodput, the improvementin queuing delay for the new sessions is also attributed to themulti-objective optimization achieved by tuning ζ.

ROY et al.: MULTI-OBJECTIVE HANDOVER IN LTE MACRO/FEMTO-CELL NETWORKS 587

VI. CONCLUSION AND FUTURE WORK

In this paper, we propose a multi-objective and optimal han-dover solution for emerging LTE systems. The proposed solu-tion considers the signal strength, as well as the available band-width, to select the optimal target eNB/HeNB. By consideringthe available resources (bandwidth), the proposed solution re-duces the blocking probabilities of both new and handover ses-sions and can provide shorter delays and better goodputs. Theefficiency of the overall solution is modeled using a 3-D Markovchain, and suitable performance analyses for the blocking ofhandover and new sessions and for the queuing delay are per-formed and evaluated with the OPNET simulation. The simula-tion results demonstrate that the proposed strategy can achieveconsiderable improvements in the session blocking and queuingdelay in comparison to the existing schemes. Clearly, significantincrease in user’s speed might reduce its stay in the femtocellsto a few seconds and trigger a series of continuous handoversbetween different femtocells with increasing overhead and se-vere degradation of TCP goodput. Thus, our future interest liesin reducing such un-necessary handovers to improve the perfor-mance even further.

REFERENCES[1] 3rd Generation Partnership Project (3GPP), “Evolved universal terrestrial

radio access (E-UTRA) and evolved universal terrestrial radio access net-work (E-UTRAN); overall description, technical specification, stage 2 (re-lease 8),” TS 36.300 V8.4.0, 2008.

[2] P. Xia, V. Chandrasekhar, and J. G. Andrews, “Open vs. closed accessfemtocells in the uplink,” IEEE Trans. Wireless Commun., vol. 9, no. 12,pp.3798–3809, Dec. 2010.

[3] V. Chandrasekhar, J. G. Andrews, and A. Gatherer, “Femtocell networks:A survey,” IEEE Commun. Mag., vol. 46, pp. 59–67, Sept. 2008.

[4] R. Y. Kim, J. S. Kwak, and K. Etemad, “WiMAX femtocell: Require-ments, challenges, and solutions,” IEEE Commun. Mag., vol. 47, pp. 84–91, Sept. 2009.

[5] M. Andrews, V. Capdevielle, A. Feki, and P. Gupta, “Autonomous spec-trum sharing for mixed LTE femto and macro cells deployments,” in Proc.IEEE INFOCOM, Mar. 2010, pp. 1–5.

[6] M. Emmelmann, S. Wiethoelter, A. Koepsel, C. Kapler, and A. Wolisz,“Moving toward seamless mobility: State of the art and emerging aspectsin standardization bodies,” Springer Wireless Personal Commun., vol. 43,no. 3, pp. 803–816, 2007.

[7] R. Bless, J. Hillebrand, C. Prehofer, and M. Zitterbart, “A quality-of-service signaling architecture for seamless handover support in next gener-ation, IP-based mobile networks,” Springer Wireless Personal Commun.,vol. 43, no. 3, pp. 817–835, Nov. 2007.

[8] A. Dutta, S. Das, D. Famolari, Y. Ohba, K. Taniuchi, V. Fajardo, R. M.Lopez, T. Kodama, and H. Schulzrinne, “Seamless proactive handoveracross heterogeneous access networks,” Springer Wireless Personal Com-mun., vol. 43, no. 3, pp. 837–855, Nov. 2007.

[9] Y-S. Kim, D-H. Kwon, and Y-J. Suh, “Seamless handover supportover heterogeneous networks using FMIPv6 with definitive L2 triggers,”Springer Wireless Personal Commun., vol. 43, no. 3, pp. 919–932, Nov.2007.

[10] W.-K. Liao and Y.-C. Chen, “Supporting vertical handover between uni-versal mobile telecommunications system and wireless LAN for real-timeservices,” IET Commun., vol. 2, no. 1, pp. 75–81, Jan. 2008.

[11] Y. Wu, K. Yang, L. Zhao, and X. Cheng, “Congestion-aware proactivevertical handoff algorithm in heterogeneous wireless networks,” IET Com-mun., vol. 3, no. 7, pp. 1103–1114, July 2009.

[12] A. Racz, A. Temesvary, and N. Reider, “Handover performance in3GPP long term evolution (LTE) systems,” in Proc. ISTMWC, July 2007,pp. 1–5.

[13] L. Bajzik, P. Horvath, L. Korossy, and C. Vulkan, “Impact of intra-LTEhandover with forwarding on the user connections,” in Proc. ISTMWC,July 2007, pp. 1–5.

[14] C-S. Chiu and C-C. Huang, “Improving inter-sector handover userthroughput by using partial reuse and softer handover in 3GPP LTE down-

link,” in Proc. Int. Conf. Advanced Commun. Technol., vol. 1, 2008,pp. 463–467.

[15] M. Anas, F. D. Calabrese, P. E. Ostling, K. I. Pedersen, and P. E. Mo-gensen, “Performance analysis of handover measurements and layer 3 fil-tering for UTRAN LTE,” in Proc. IEEE PIMRC, Sept. 2007, pp. 1–5.

[16] M. Anas, F. D. Calabrese, P.E. Mogensen, C. Rosa, and K. I. Pedersen,“Performance evaluation of received signal strength based hard handoverfor UTRAN LTE,” in Proc. IEEE VTC, Apr. 2007, pp. 1046–1050.

[17] T-H. Kim, Q. Yang, J-H. Lee, S-G. Park, and Y-S. Shin, “A mobility man-agement technique with simple handover prediction for 3G LTE systems,”in Proc. IEEE VTC, Oct. 2007, pp. 259–263.

[18] M. Chang, H. Lee, and M. Lee, “A per-application mobility managementplatform for application-specific handover decision in overlay networks,”Elsevier Comput. Netw., vol. 53, no. 11, pp. 1846–1858, July 2009.

[19] A. Chinchuluun, P. M. Pardalos, A. Migdalas, and L. Pitsoulis, Eds.,Pareto Optimality, Game Theory, and Equilibria, Springer Optimizationand Its Applications, vol. 17, Springer, 2007.

[20] J. L. Cohon, Multiobjective Programming and Planning, Mathematics inScience and Engineering, Academic Press, 1978.

[21] T. W. Athan and P. Y. Papalambros, “A note on weighted criteria methodsfor compromise solutions on multiobjective optimization,” EngineeringOptimization, vol. 27, no. 2, pp. 155–176, 1996.

[22] OPNET Modeler, OPNET Technologies Inc. [Online]. Available: http://www.opnet.com/

[23] C. Smith and D. Collins, “3G Wireless Networks,” 5th ed. McGraw-HillProfessional, 2001.

[24] C. Bettstetter, “Mobility modeling in wireless networks: Categorization,smooth movement, border effects,” ACM Mobile Comput. Commun. Rev.,vol. 5, no. 3, pp. 55–67, July 2001.

Abhishek Roy is currently working as a Senior Engi-neer (Manager) in the System Lab of Telecommunica-tions Systems Division, Samsung Electronics, SouthKorea. He received his Ph.D. degree in 2010 fromSungkyunkwan University in the College of Informa-tion and Communication Eng., M.S. degree in 2002from the University of Texas at Arlington USA, andB.E. degree in 2000 from Jadavpur University, In-dia, all in Computer Science and Engineering. Hisresearch interests include different mobility and re-source management aspects of 3G and 4G wireless

systems. He served as the Guest Editor of Springer EURASIP Journal of Wire-less Communications and Networking. He has published more than 20 Interna-tional Journals and more than 20 International Conferences. He is the Recipientof Best Masters Award from the University of Texas, Arlington in 2002 andEmployee of Excellence Award from Samsung Electronics in 2011.

Jitae Shin received his B.S. degree from Seoul Na-tional University in 1986, M.S. degree from the Ko-rea Advanced Institute of Science and Technology(KAIST) in 1988, and M.S. and Ph.D. degrees in Elec-trical Engineering from the University of SouthernCalifornia, Los Angeles, USA in 1998 and 2001, re-spectively. He is an Associate Professor in the Col-lege of Information and Communication Engineer-ing of Sungkyunkwan University, Suwon, Korea. Hisresearch interests include video signal processing,transmission over next generation Internet and wire-

less/mobile networks, focusing on QoS/QoE, 4G communication systems, andmultimedia network control/protocol issues. He is a Member of IEEE and IE-ICE.

Navrati Saxena is currently working as an AssociateProfessor in the College of Information and Com-munication Eng. of Sungkyunkwan University, SouthKorea. During 2007–2011, she worked as AssistantProfessor in the same University. Prior to that, sheworked as an Assistant Professor in Amity Univer-sity, India and as a Visiting Researcher in the Com-puter Science and Engineering Department of theUniversity of Texas at Arlington. She completed herPh.D. degree from the Department of Information andTelecommunication, University of Trento, Italy. Her

prime research interests involve 3G/4G wireless and Ubiquitous/Smart Envi-ronments. She is serving as the Guest Editor of Springer EURASIP Journal ofWireless Communications and Networking. She has published more than 20International Journals and more than 20 International Conferences. She is theRecipient of Best Masters Award from Tagra University, India.