Efficient Resource Allocation for Device-to-DeviceCommunication Underlaying LTE NetworkMohammad Zulhasnine

Systems and Computer EngineeringCarleton University, Ottawa, Canada

Email: mzulhasn@sce.carleton.ca

Changcheng HuangSystems and Computer EngineeringCarleton University, Ottawa, Canada

Email: huang@sce.carleton.ca

Anand SrinivasanEION Inc., Ottawa, Canada

Email: anand@eion.com

Abstract—Device-to-device (D2D) communication as an under-laying cellular network empowers user-driven rich multimediaapplications and also has proven to be network efficient offloadingeNodeB traffic. However, D2D transmitters may cause significantamount of interference to the primary cellular network whenradio resources are shared between them. During the downlink(DL) phase, primary cell UE (user equipment) may suffer frominterference by the D2D transmitter. On the other hand, theimmobile eNodeB is the victim of interference by the D2Dtransmitter during the uplink (UL) phase when radio resourcesare allocated randomly. Such interference can be avoided oth-erwise diminish if radio resource allocated intelligently with thecoordination from the eNodeB. In this paper, we formulate theproblem of radio resource allocation to the D2D communicationsas a mixed integer nonlinear programming (MINLP). Such anoptimization problem is notoriously hard to solve within fastscheduling period of the Long Term Evolution (LTE) network. Wetherefore propose an alternative greedy heuristic algorithm thatcan lessen interference to the primary cellular network utilizingchannel gain information. We also perform extensive simulationto prove the efficacy of the proposed algorithm.

I. INTRODUCTION

Due to the emergence of fourth-generation (4G) mobiletechnology, Cisco estimates 131 percent compound annualgrowth rate of the mobile traffic between 2008 and 2013 [1].Third Generation Partnership Project (3GPP) is working on anew air interface known as Long Term Evolution (LTE). Withmore compelling user-driven applications, LTE is expected toachieve high data rates, low latency and packet optimized radioaccess technology (RAT). Recently, device-to-device (D2D)communication as an underlay to cellular networks has beenintroduced as a technology component to the LTE-Advanced[2], [3]. User equipments (UEs) in close proximity, with higherSignal to Interference and Noise Ratio (SINR) between them,may communicate directly instead of through the eNodeB (theLTE base station). UEs communicating directly, by employingD2D connectivity, can achieve better performance than thatoffered via eNodeB (two-hops) by offloading eNodeB re-sources. Some appealing applications of D2D communicationsare video streaming, online gaming, media downloading, peer-to-peer (P2P) file sharing etc. It would be highly advantageous,if some UEs download contents through eNodeB while otherUEs retrieve it through D2D communication and thereby avoidcongestion at the eNodeB.

If the D2D users are assigned dedicated LTE resource blocks

(RBs), there is no interference between cellular users andD2D users. This is only possible when sufficient amount ofresources are available. However when D2D communicationshares the same resources with cellular communication; theinterference of D2D communications to the cellular networkneeds to be restricted to maintain a target performance level ofthe cellular network. The eNodeB controls the transmit powerof D2D connections in order to restrict the interference of D2Dtransmitters to the cellular network.

There are manifold benefits of enabling D2D communi-cation in a cellular network under the control of LTE sys-tem. With control over D2D connectivity, the operators canincorporate the D2D service exploiting existing cellular ar-chitecture, offer new services with new revenue opportunities.The eNodeB can employ power control mechanism for D2Dconnections and guarantee limited interference to the cellularnetwork. Moreover, with planned resource allocation, eNodeBcan decide whether to allocate dedicated resources for D2Dservices when ample amount of resources available. In LTEsystem, resource management is fast and operates in high time-frequency resolution; eNodeB controlled D2D services is thusrationalized.

A. Relevant Works and Motivations

There has been considerable research on spectrum sharingbetween cellular network and infrastructure-less wireless net-works [4], [5], [6]. Due to heavier download traffic, uplink(UL) spectrum is under-utilized in frequency division duplex(FDD) based cellular system with equal bandwidths allocatedfor UL and downlink (DL) transmission. [5] suggested thatnetwork capacity can be improved significantly when ad hocusers make use of unoccupied UL sub-channels. The authors in[6] also proposed a spectrum reuse protocol where D2D usersare only allowed to communicate with each other during theUL frame of the network. This is due to the fact that duringUL only the base station (BS) is exposed to interference bythe D2D users. In this scheme, the D2D user measures itspathloss from the known BS power and the received powerduring DL control frame. The D2D user then calculates itstransmit power such that SINR ratio of the BS does not fallbelow threshold level.

In [3], the authors investigated another D2D communicationthat is based on the statistics of the SINR of all users. In this

scheme, SINR behavior is formulated based on the positionof the D2D pair. Then power control method is applied tothe D2D communication such that SINR degradation of thecellular link does not fall behind a certain level. The resultsshowed that the SINR distribution of the D2D users (withdistance constraint) is comparable to that of the cellular userin most of the cell area. The authors also observed that theUL resource sharing is more beneficial for the system whenthe D2D pair is farther away from the BS. When the D2Dpair is closer to the BS, the DL resources sharing performsbetter. In [7], the authors studied D2D communication usingWi-Fi ad-hoc mode in the license exempt bands. The authorsconcluded that mobile devices (in close proximity) whencommunicate directly instead of infrastructure modes, canreduce contentions/collisions and also avoid bottleneck onan access point. However, the cellular operators are morelikely to nourish D2D communication on a licensed band as acontrolled underlay to a cellular network. With controlled D2Dcommunication, they can design the network with guaranteedlower level of interference as well as make additional profitout of it.

Several authors studied D2D communication over cellulararchitecture in the context of P2P file sharing [8], [9], [10]. In[9], the authors suggested that an extended peer (non cellularuser) from P2P network can communicate with cellular usersas a client/server based communication between them. Inthis way cellular users can participate indirectly in the P2Pnetwork, using the extended peers as proxies and also avoidthe costly competition for resources. [10] proposed a P2P filesharing application for cellular users using session initializa-tion protocol (SIP) as control protocol and then elaborated themodifications that should be made to SIP in order to meet therequirements of that application.

Cognitive radio (CR) technology, a very active researcharea for the past two decades, also reuses the under-utilizedspectrum in an opportunistic manner [11], [12], [13]. CRssense the radio frequency environment to detect temporaland spatial “holes” in the spectrum and thereby avoid theinterference with the primary users. However, the sensingability of secondary users demands high level of complexity todetect weak primary signals. Even CR system requires somedegree of coordination among the secondary users to ensurefairness of spectrum usages [12].

Several authors emphasized the advantages of cellu-lar/WLAN interworking systems as a two-tier overlayingstructure [14], [15]. In this structure, disjoint WLANs pro-vide local coverage in hotspot areas, while cellular networksprovide wide area coverage. Mutual interference and ver-tical handoff (switching between two access networks) areinevitable problems of this type of coexisting network.

B. Contributions

In this paper, we address the spectrum sharing problembetween 3GPP-LTE cellular network and underlaying D2Dcommunication network. We identify and analyze the inter-ference problem of the primary cellular network caused by

Internet

eNodeB eNodeB

S-GWMME

P-GW

UE UE

D2D

EPC

E-UTRAN

User Data Interface

Control Interface

Fig. 1. System level overview of LTE architecture.

the D2D transmitter during the UL and DL phases separately.To the best of our knowledge, we first formulate the problemof D2D shared radio resource allocation problem as a MINLPproblem where eNodeB controls D2D connections. We alsopropose an alternative greedy heuristic algorithm which candiminish interference to the primary cellular network utiliz-ing channel gain information. The proposed greedy heuristicD2D radio resource allocation algorithm does not requireany modification to the scheduling of the primary cellularnetwork; rather it utilizes information provided by the primaryscheduler. We also perform extensive simulation to prove theefficacy of the proposed algorithm over the random selectionof radio resources.

II. NETWORK AND CHANNEL MODELING

A. Network Model

The entire LTE architecture is termed as EPS (EvolvedPacket System) which comprises of Evolved Packet Core (E-PC) and the core network named as evolved UMTS TerrestrialRadio Access Network (E-UTRAN). The E-PC consists ofMobility Management Entity (MME), Serving Gateway (S-GW), and packet data network (PDN) Gateway (P-GW). TheE-UTRAN only comprises the evolved base stations (alsoknown as eNodeBs). User equipments (UEs) communicatethrough eNodeB, whereas the eNodeBs are interconnectedand also connected to MME, and S-GW. Fig. 1 depicts thehigh-level system overview of the LTE system architecture.In this architecture, UEs may communicate directly over theD2D links. However, eNodeB needs to establish the D2Dconnection and also remains in control of resource allocationto limit the interference experienced by the cellular receivers.In [2], the authors illustrated two options for D2D sessionsetup and management along with interference coordination toprotect the primary LTE cellular network. In 3GPP-LTE IP-based systems, P-GW performs the routing from/to the Internet

1

2

12

1 2 3 10

One Radio Frame, T = 10 msf

T = 0.5 ms = 7 OFDM symbolslot

Time

Frequency

subframe

number

M

15KHz

12x15=180kHz

RB

Fig. 2. The LTE downlink physical resource.

and also is aware of which eNodeB the UE is served. There-fore, P-GW is able to detect potential D2D from the sourceand destination of IP addresses of the UEs. Alternatively, theauthors also proposed a new type of dedicated radio bearerthat enables D2D communications and stays in control of thesession setup and the radio resources. We consider cellularnetwork as the primary network and D2D communication onlyuse shared channel to improve the overall performance.

B. Radio Resource and Access Technology

In LTE system bandwidth is divided into equal size physicalRBs. Each RB physically occupies (0.5ms) 1 slot in the timedomain and 180 kHz in the frequency domain with subcarrierspacing of 15 kHz. Fig. 2 illustrates the LTE DL physicalresource. LTE has adopted Orthogonal Frequency DivisionMultiplex (OFDM) based radio interface due to its higherspectral efficiency and resilience against multi-path delayspread. There is, however, one important difference betweenthe feasible assignments on the UL and DL shared channels.In LTE, the multiple access scheme for DL (from eNodeB toUEs) is OFDM access (OFDMA). The radio access technologyin UL is single carrier frequency division multiple access (SC-FDMA) due to its characteristics of low peak-to-average powerratio (PAPR) that enables higher transmit power efficiency forthe UEs. The physical properties of SC-FDMA require thatRBs allocated to a single user to be contiguous in frequency.The minimum scheduling period in the frequency domain isone physical RB; therefore the smallest unit of resource thatcan be assigned is two RBs.

C. Channel Model

In order to measure two important parameters SINR andchannel gain, we consider both distant dependent macroscopicpathloss and shadow fading pathloss. For urban and suburbanareas, the macroscopic pathloss between an eNodeB and anUE at a distance d meter is:

LdB(d) = 40(1− 4× 10−3hb) log10(d/1000)−18 log10(hb) + 21 log10(fc) + 80 (1)

here, fc is the carrier frequency in MHz, and hb is the basestation antenna height (in meter). For different scenarios anddetails, the readers are encouraged to read [16].

The slow fading or shadowing on a wireless channel iscaused by obstacles in the propagation path of the radio wavesand obviously is location-dependent [17]. When shadowing ismodeled only by log-normally distributed random variables,the variables do not meet the spatial correlation properties.Finding an accurate fading model is a myth. However, wegenerate realistic looking patterns as described in [18]. Gaus-sian random fields are generated first and then convoluted two-dimensionally along with the following exponential correlationfunction.

R(∆d) = e−|∆d|

dcorrln 2 (2)

where dcorr is the decorrelation distance. We assume that theimpact of fast fading is averaged out over a certain time period.

Each eNodeB consists of three sectoral antenna with thefollowing radiation pattern as described in [19]:

Adb(θ) = −min[(θ/θ3dB)2 , Am

]; −1800 ≤ θ ≤ 1800

θ3dB is the 3 dB beam width, and Am maximum attenuation.The total pathloss, which includes the antenna gain, betweeneNodeB B and the user u, is:

PLdB,B,u(.) = LdB(d) + log10(Xu)−AdB(θ) (3)

where Xu lognormal shadow fading pathloss of user u. Thelinear gain between the eNodeB and a user u is

GBu = 10−PLdB,B,u/10 (4)

For D2D communication, the gain between two UEs u andv is Guv = Kuvd−α

uv [20]. Here, duv is the distance betweentransmitter u and receiver v, α is a constant path loss expo-nent and Kuv is a normalization constant. The normalizationconstant depends on the radio propagation properties of theenvironment, and also accounts for the effects of coding gain,spreading gain, beamforming, etc.

III. PROBLEM DEFINITION

D2D commutation takes place underlaying the primarycellular network. We acknowledge the radio resources to bevaluable and D2D communication shares the same resourceswith cellular communication; rather than using dedicatedresources. Hence, the interference of D2D communicationsto the cellular network needs to be restricted to maintaina target performance level of the cellular network. If thedistance between D2D is too long, the D2D will require highpower. In that case, the primary cellular network would sufferfrom interference. Fig. 3 illustrates the interference problemwhen primary cellular network shares channel with the D2Dnetwork. During the download period of cellular network (seeFig. 3(a)), cellular UE0 is exposed to interference when anyD2D transmitter (UE1, or UE3) transmits using the sameallocated subband. Also, D2D receiver (UE2, or UE4) willsuffer from interference from the UE0, whichever shares the

UE0 downlink

UE2UE1

Pair 1

Interferenceduetopair1

UE3

UE4

Pair2

Interfere

ncedue

to pair 2

InterferenceduetoeNodeB

(a) Download period

UE0 uplink

UE2 UE1

Pair 1

InterferenceduetoUE0

UE3

UE4

Pair 2Interferen

ce due toUE0

Interferencetoduetopair1

(b) Uplink period

Fig. 3. Interference problem when primary cellular network shares channelwith the D2D network.

same subband with the cellular UE0. One can easily speculatethat the amount of interference will depend not only onD2D transmit power but also spatial distance between D2Dtransmitter and the cellular users. On the other hand, duringthe UL period of the cellular network (see Fig. 3(b)), theD2D transmitter (UE1 or UE4) causes interference only to theimmobile eNodeB. D2D receiver (UE2 or UE3) also suffersfrom interference from the UE0. Anyone can whisper that theamount of interference depends on the transmitter power andthe spatial distance as well. Hence, intelligent selection ofthe shared RB would result in better performance in termsof network throughput. According to the suggestion [2], wedeny D2D connections if the the maximum distance betweenD2D is higher than dmax = 25 m. Since, eNodeB coordinatesthe D2D session setup, we assume that both UEs need to bein the same cell for possible D2D connection.

IV. PROBLEM FORMULATION

The eNodeB schedules the primary cellular network instandard way, and our solution does not require any modi-fication for D2D resource allocation. However, we want toallocate the same assigned RB(s) of any cellular UE c to oneof D2D connections for which total network throughput isincreased. We formulate the problem of assigning appropriateRBs for underlaying D2D communication as an optimizationproblem that achieves higher throughput without impairingthe existing cellular network. Consider a time division duplex(TDD) cellular network with identical splitting of UL and DLresources. The total number of available RBs during the DLand UL is equal to M and N . This is due to the fact thatbecause of TDD, there is only a single-carrier frequency andthe UL and DL transmissions are separated in time, and alsoon cell site basis; however this does not affect our proposedscheme. We assume infinitely backlogged model where datais always available from each UE. We also assume that moreadvanced intercell interference mitigation scheme works ontop of our scheme. The eNodeB serves a set C = {1, . . . , C}of cellular users and also coordinates a set D = {1, . . . , D}D2D pairs. According to our assumptions, C >> D.

A. Downlink Phase Interference Coordination

During the download phase of cellular network, any UEis exposed to interference when any D2D user is allowed totransmit using the same allocated subband. And the amountof interference will depend not only on D2D transmit powerbut also the channel gain between D2D transmitter and thecellular users. Gcd denotes the channel gain between UEs cand d. GBc denotes the channel gain between the eNodeBand UE c. When D2D devices utilizes DL RBs, any D2Dtransmitter d causes interference to cellular user c while usingRB(s) assigned to it. The DL SINR of user c is

γDLc =

PBGBc

N0 + I +∑

d xdcPdGcd

(5)

The SINR of the D2D pair d is

γDLd =

∑c xd

cPdGdd

N0 + I +∑

c xdcPBGdB

(6)

Here, xdc represents a binary variable which satisfies xd

c = 1if D2D pair d uses RB(s) assigned to cellular user c. N0

accounts for receiver’s noise figure and thermal noise density;I accounts for intercell interference. Let us define the ratesrDLc and rDL

d corresponding to the SINR γDLc and γDL

d

as determined by the Shannon capacity model. We wantto maximize the sum rate of the primary cellular UEs andsecondary D2D UEs. The problem is formulated as a MINLPproblem as below:

Maximize

C∑c

mcrDLc +

D∑

d

C∑c

xdcmcr

DLd (7)

PBGBc ≥ γDLc,tgt

(N0 + I +

∑

d

xdcPdGcd

); ∀c ∈ C (8)

∑c

xdcPdGdd ≥ γDL

d,tgt

(N0 + I +

∑c

xdcPBGdB

); ∀d ∈ D (9)

∑c

xdc ≤ 1; ∀d ∈ D (10)

∑

d

xdc ≤ 1; ∀c ∈ C (11)

Here, mc denotes the number of RBs allocated to the cellularuser c at each time slot during the DL period. Constraints in(8), and (9) guaranty the target SINR of the cellular UE andD2D communication, respectively. Constraints in (10) ensurethat each device shares at most one user’s RB(s). Whereasconstraints in (11) ensure that at most one D2D pair sharesany user’s RB(s).

B. Uplink Phase Interference Coordination

During the UL phase of the cellular network, the D2Dtransmitter causes interference only to the immobile eNodeB.Any D2D receiver is also exposed to interference from the

UE whose subband is being shared. The SINR of the D2Dreceiver d during the UL phase is γUL

d and calculated as:

γULd =

∑c yd

c PdGdd

N0 + I +∑

c ydc PcGcd

(12)

γceNB =

PcGcB

N0 + I +∑

d ydc PdGdB

(13)

Here, ydc represents a binary variable which satisfies yd

c = 1if D2D pair d uses RB(s) assigned to cellular user c. Let usdefine the rates rUL

d and rceNB corresponding to the SINR γUL

d

and γceNB as determined by the Shannon capacity model. We

want to maximize the sum rate of the primary cellular UEs andsecondary D2D UEs. The problem is formulated as a MINLPproblem as below:

Maximize

C∑c

ncrceNB +

D∑

d

C∑c

ydc ncr

ULd (14)

PcGBc ≥ γc,tgteNB

(N0 + I +

∑

d

ydc PdGdB

); ∀c ∈ C (15)

∑c

ydc PdGdd ≥ γUL

d,tgt

(N0 + I +

∑c

ydc PcGcd

); ∀d ∈ D

(16)

∑c

ydc ≤ 1; ∀d ∈ D (17)

∑

d

ydc ≤ 1; ∀c ∈ C (18)

Here, nc denotes the number of RBs allocated to the cellularuser c at each time slot during the UL period. Constraints in(15), and (16) guaranty the target SINR of the cellular UE andD2D communication, respectively. Constraints in (17) ensurethat each device shares at most one user’s RB(s). Whereasconstraints in (18) ensure that at most one D2D pair shares ofany user’s RB(s).

V. GREEDY HEURISTIC RB SELECTION ALGORITHM

The aforementioned MINLP problem is notoriously hardto solve otherwise impossible within such a short schedulingperiod (1 millisecond). Therefore, we propose an alternativegreedy heuristic RBs selection algorithm. During the DLphase, higher values of γDL

c and γDLd would aid in increased

cell and D2D throughput, respectively. Through careful obser-vation of Eq. (5), it is obvious that lower channel gain (Gcd)between UE c, and D2D transmitter d will result in higherγDL

c ; i.e. will cause less interference to UE c. Therefore, anyUE with higher channel quality indicator (CQI) can share RBsassigned to them with the D2D transmitter with lower channelgain between them. Here, we assume that the eNodeB receiveschannel gain (between UE and D2D user) information throughcontrol; as it remains in control of D2D connections.

For the UL phase, by observing Eq. (12), and Eq. (13); weintuitively decide that lower Gcd or increase spatial distance

Algorithm 1: Downlink D2D RB allocation schemeC : Sorted list of CQIs for all downlink UEs in decreasing order1Gcd : Channel gain matrix2D : List of D2D connections yet to be assigned RBs3mc : Number of RBs assigned to cellular UE c4begin5

c ← 1;6while D 6= φ or c == C do7

Pick RBs with cth largest value;8Find the D2D transmitter d for which channel gain is minimum;9

γDLc ← PBGBc

N0+I+PdGcd;10

γDLd ← PdGdd

N0+I+PBGdB;11

if γDLc ≥ γDL

c,tgt and γDLd ≥ γDL

d,tgt then12Share all RBs of the UE c with D2D connection d;13D = D − {d};14c ← c + 115

else16Do not assign RBs to D2D connection d17

end18end19

end20

Algorithm 2: Uplink D2D RB allocation schemeC : Sorted list of CQIs for all uplink UEs in decreasing order1Gcd : Channel gain matrix2D : List of D2D connections yet to be assigned RBs3mc : Number of RBs assigned to cellular UE c4begin5

c ← 1;6while D 6= φ or c == C do7

Pick RBs with cth largest value;8Find the D2D transmitter d for which channel gain is minimum;9

γULd ← PdGdd

N0+I+PcGcd;10

γceNB ← PcGBc

N0+I+PdGdB;11

if γULd ≥ γDL

d,tgt and γceNB ≥ γc,tgt

eNB then12Share all RBs of the UE c with D2D connection d;13D = D − {d};14c ← c + 115

else16Do not assign RBs to D2D connection d17

end18end19

end20

between cellular UE and D2D receiver d would result in higherD2D throughput. Howsoever, the paramount importance isto limit the interference at the eNodeB caused by the D2Dconnections. Both algorithms are summarized in Fig. 1, and2.

VI. SIMULATION METHODOLOGY

We perform extensive simulations to evaluate the efficacy ofthe proposed RBs sharing methods. We developed a simulatorusing C ++ programming language that realistically modeledthe LTE system. Table I summarizes a list of simulation param-eters and their default values. We generated distant dependentmacroscopic pathloss including antenna gain with 7-cell, 3sectors hexagonal layout. Fig. 4 shows the aforementionedpathloss map. Shadow fading, when approximated only by alog-normal distribution, cannot accurately model the location-dependent characteristics. Hence, we generated realistic space-correlated shadow fading pathloss applying 2D convoluted

TABLE ISIMULATION PARAMETERS AND VALUES

Parameter ValuesSpectrum allocation(UL/DL) 20 MHzCarrier frequency 2 GHzNumber of subcarriers per RB 12Neighboring subcarrier spacing 15KHzRB bandwidth 180 kHzNumber of available RBs 100Max eNodeB Tx power 20 WMax antenna gain l5 dBiMax UE Tx Power 250mW [24dBm]Modulation and coding scheme (MCS) QPSK: 1/6, 1/3, 1/2, 2/3

16QAM: 1/2 , 2/3, 3/464QAM: 1/2 , 2/3, 3/4, 4/5

Slot duration 0.5msNumber of symbols per slot 7(1 Pilot+6 Data)3dB beam width, θ3dB 650

Maximum attenuation, Am 20 dBCell-level user distribution UniformNumber of active users per cell 10, 20, 30, 40, 50User speed 5 km/hLog-normal shadowing Standard deviation 8 dBShadowing decorrelation distance dcorr =20 mDistance attenuation L = 35.3 + 37.6× log(d)Log-normal shadow fading 8 dB standard deviationCell layout Hexagonal grid, 3-sector sites, 21

sectors in totalCell radius 167m (500m inter-site distance)UE noise figure 9 dBUE thermal noise density -174 dBm/Hz

Macroscopic pathloss distribution [dB]

x location (meter)

y lo

catio

n (m

eter

)

−600 −400 −200 0 200 400 600 750−750

−600

−400

−200

0

200

400

600

750

−750 50

60

70

80

90

100

Fig. 4. Distance dependent macroscale pathloss map (in dB, includingantenna gain) with 7 eNodeBs (in white circle).

Uncorrelated lognormal pathloss distribution [dB]

x location (meter)

y lo

catio

n (m

eter

)

−100 0 100−200

−150

−100

−50

0

50

100

150

200

−15

−10

−5

0

5

10

15

Correlated lognormal pathloss distribution [dB]

x location (meter)

y lo

catio

n (m

eter

)

−100 0 100−200

−150

−100

−50

0

50

100

150

200

−30

−20

−10

0

10

20

30

Fig. 5. Uncorrelated and space-correlated lognormal shadow fading pathlossmap (in dB).

Pathloss distribution [dB]

x location (meter)

y lo

catio

n (m

eter

)

−600 −400 −200 0 200 400 600 750−750

−600

−400

−200

0

200

400

600

50

60

70

80

90

100

110

Fig. 6. Simulation region (of size 1500 × 1500 m2) showing pathloss(dB) that includes antenna gain, distant dependent macroscale pathloss, andlognormal shadow fading pathloss.

function with decorrelated distance dcorr = 20 m [18]. Fig.5 shows the shadow fading map with random (uncorrelated)shadow fading variables (left) and with convolutional corre-lated shadow fading variables (right). The mean and standarddeviation of the shadow fading variables are 0 dB and 10dB. Fig. 6 shows the pathloss map (dB) that includes antennagain, distant dependent macroscale pathloss, and lognormalshadow fading loss. Users are distributed uniformly insidethe scenario. The SINR is calculated from the received signalpower and interference power level. The CQI are calculatedfrom the SINRs at the UEs and then fed back to eNodeB.Link adaptation is performed based on the CQI at the eN-odeB. Various modulation and coding schemes (MCS) (QPSK,16QAM, 64QAM) are considered with varying coding ratesranging from 1/8 to 4/5 (see [19], [21] for details). The ratecorresponds to a certain block error rate (BLER, %10) arealso estimated from the Ack/Nack of the past transmissions.The number of RBs is 100 for 20MHz bandwidth. UL andDL bandwidth is divided equally into m RBs. The UL/DLbandwidth ratio does not effect the proposed scheme. Threedifferent scheduling algorithms are employed for the primarycellular network: Round Robin (RR), Maximum Carrier toInterference ratio (Max C/I) and Proportional Fair (PF) withCQI feedback [22]. The throughput is calculated by mappingthe SINR to the ideal link-adaptation based LTE link-levelcapacity as follows:

η =

0, if γ < γmin

Weff × log2 (1 + γ) , if γmin ≤ γ < γmax

4.7, if γ ≥ γmax;

where η is the estimated spectral efficiency in bps/Hz, γ is theSINR and Weff (0.57) is the attenuation factor applied to theShannon bound. Weff accounts for link level implementationoverhead and also for overhead due to pilot, and controlchannel signal. We apply hard spectral efficiency, which givesηmax = 4.7 bps/Hz at γmax = 25 dB or higher. Also, η = 0

10 20 30 40 50 60 70 80 90 1000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Time (s)

Nor

mal

ized

thro

ughp

ut

Cell, RRCell, max CQICell, PF with CQICell+D2D, RRCell+D2D, max CQICell+D2D, PF with CQI

Fig. 7. Normalized network throughput for different scheduling algorithmswith the proposed greedy heuristic RB assignment algorithm [scenario 1: thenumber of D2D connection is 10% (of total UEs)].

10 20 30 40 50 60 70 80 90 1000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Time (s)

Nor

mal

ized

thro

ughp

ut

Cell, RRCell, max CQICell, PF with CQICell+D2D, RRCell+D2D, max CQICell+D2D, PF with CQI

Fig. 8. Normalized network throughput for different scheduling algorithmswith the proposed greedy heuristic RB assignment algorithm [scenario 2: thenumber of D2D connection is 20% (of total UEs)].

at γmin = −10 dB or lower..

VII. RESULTS

In this section, we analyze and discuss the simulationresults.

A. Performance of heuristic D2D RB assignment algorithmfor different scheduling algorithms

D2D commutation takes place underlaying the primary cel-lular network. We limit the number of possible D2D connec-tions as a percentage of total UEs in the network. We performthe simulations for a period of 500 seconds on two scenarioswhere the number of D2D connections are restricted to 10%and 20% of the total UEs in the work. As the patterns remainthe same, we present the results for the first 100 seconds toshow the initial variations. To show how the proposed D2D RBassignment algorithm performs, we plot sum rate of cell andD2D throughput. We also plot cell throughput to investigatewhether primary cellular network does not degrade due to

10 15 20 25 30 35 40 45 500

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Number of active UEs per eNodeB

Nor

mal

ized

net

wor

k th

roug

hput

Cell, RandomCell, HeuristicCell+D2D, RandomCell+D2D, Heuristic

Fig. 9. Comparisons between random and heuristic D2D RB assignmentalgorithms in terms of normalized aggregate throughput.

interference from the D2D connections. Fig. 7 and Fig. 8shows normalized throughput variation with time for RR, MaxC/I, PF with CQI scheduling algorithms in the case of scenario1 and 2, respectively. For both scenarios, RR performs poorlyin terms of cell and sum of cell and D2D throughput than thatof other two scheduling algorithms. This is due to that RRdoes not consider current channel condition and allocates equalresources to all users. Max C/I performs the best in terms ofcell and sum of cell and D2D throughput by exploiting channelcondition. Moreover, for all algorithms network throughput ishigher when D2D communication integrated underlaying thecellular network.

B. Comparisons between random and heuristic D2D RB as-signment algorithms

To compare the proposed D2D RB assignment algorithmwith the random assignment algorithm, we plot the normalizedthroughput with varying number of active UEs in the cell(see Fig. 9). We kept the number D2D connections constant(20% of the cell UEs). The cell throughput is higher for theproposed heuristic algorithm than the random algorithm forall active number of UEs in the cell. This is due to the factthat random D2D RB assignment causes significant amountof interference on the primary cellular network. For example,with the number of 30 active UEs per site, normalized cellthroughput is 7% higher in the case of heuristic algorithm thanthe random algorithm. As the proposed heuristic algorithm ismore efficient to avoid interference to the cell UEs than therandom algorithm, the sum of cell and D2D throughput isalso higher for the heuristic algorithm than that of the randomalgorithm. For example, with the number of 30 active UEs persite, normalized cell throughput is 9% higher in the case ofheuristic algorithm than the random algorithm.

C. Impact of number of D2D connections on network through-put

Finally we present the impact of number of D2D connec-tions on the network throughput with varying number of active

10 15 20 25 30 35 40 45 500

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Number of active UEs per eNodeB

Nor

mal

ized

net

wor

k th

roug

hput

Cell, 10%, HeuristicCell+D2D, 10%, HeuristicCell, 20%, HeuristicCell+D2D, 20%, Heuristic

Fig. 10. Impact of the number of D2D connections on normalized aggregatethroughput.

UEs for the proposed greedy heuristic algorithm. The numberof D2D connections is either 10% or 20% of the numberactive UEs. For any specific number of active UEs, the cellthroughput is lower with 20% D2D connections than that ofwith 10% D2D connections. For example with 40 active UEsper eNodeB, the cell throughput is 5% lower with 20% D2Dconnections than that of with 10% D2D connections. Withmore D2D connections, the interference is higher and the thecell throughput is lower. However, for any specific numberof active UEs, the sum throughput of cell and D2D users ishigher with 20% D2D connections than that of 10% D2Dconnections. With higher D2D connections the sum throughputincreases from the contribution of D2D communications.

VIII. CONCLUSION

D2D communication offers an advantageous complement toinfrastructure mode. The network performance improves whenD2D communications share radio resources with the primarynetwork. However, D2D transmitter causes interference to theUE receiver during the DL period and to the immobile eNodeBduring the UL phase. In this paper, we proposed spectrumsharing strategies between 3GPP-LTE cellular network and un-derlaying D2D communication network that is capable of di-minish interference. We performed extensive simulations withrealistic scenario and LTE-standard parameters. Simulationresults showed that the proposed greedy heuristic algorithmimproves network performance in terms sum of cell and D2Dthroughput without causing significance harm to the primarycellular network for all standard scheduling algorithms of theprimary network. The proposed heuristic algorithm selectsthe appropriate shared radio resource for which channel gainbetween the UE receiver and D2D transmitter during the DLphase and channel gain between the D2D transmitter andthe eNodeB during the UL phase are lower. The results alsoshowed cell and sum of cell and D2D throughput is higher forthe proposed heuristic algorithm than the random algorithm.With the increase of D2D connections, although the cellthroughput has decreased, the sum of cell and D2D throughput

has increased significantly with the proposed algorithm. Asthe eNodeB remains in control of D2D connections, D2Dcommunication is a promising integration for LTE Advancednetwork.

REFERENCES

[1] Cisco, “Visual networking index: Forecast and methodology, 2008-2013,” White Paper, 2009.

[2] K. Doppler, M. Rinne, C. Wijting, C. Ribeiro, and K. Hugl, “Device-to-device communication as an underlay to LTE-advanced networks,”IEEE Commun. Mag., vol. 47, no. 12, pp. 42–49, Dec. 2009.

[3] C.-H. Yu, O. Tirkkonen, K. Doppler, and C. Ribeiro, “On the perfor-mance of device-to-device underlay communication with simple powercontrol,” in IEEE Vehicular Technology Conference, Spring, 2009, pp.1–5.

[4] K. Huang, V. K. N. Lau, and Y. Chen, “Spectrum sharing betweencellular and mobile ad hoc networks: transmission-capacity trade-off,”IEEE J. Sel. Areas Commun., vol. 27, no. 7, pp. 1256–1267, sep. 2009.

[5] K. Huang, Y. Chen, B. Chen, X. Yang, and V. Lau, “Overlaid cellularand mobile ad hoc networks,” in IEEE International Conference onCommunication Systems (ICCS), 2008, pp. 1560–1564.

[6] B. Kaufman and B. Aazhang, “Cellular networks with an overlaid deviceto device network,” 2008, pp. 1537–1541.

[7] B. Xing, K. Seada, and N. Venkatasubramanian, “An experimental studyon Wi-Fi Ad-Hoc mode for mobile device-to-device video delivery,” inIEEE INFOCOM Workshops, 2009, pp. 1–6.

[8] A. Berl, H. de Meer, and T. Hossfeld, “Peer-to-peer solutions forcellular networks,” in First International Conference on Advances inP2P Systems, 2009, pp. 54–59.

[9] A. Berl and H. de Meer, “Integration of mobile devices into popularpeer-to-peer networks,” in Next Generation Internet Networks, 2009, pp.1–9.

[10] L. Li and X. Wang, “P2P file-sharing application on mobile phones basedon SIP,” in 4th International Conference on Innovations in InformationTechnology, 2007, pp. 601–605.

[11] S. Haykin, “Cognitive radio: brain-empowered wireless communica-tions,” IEEE J. Sel. Areas Commun., vol. 23, no. 2, pp. 201–220, feb.2005.

[12] A. Sahai, R. Tandra, S. M. Mishra, and N. Hoven, “Fundamental designtradeoffs in cognitive radio systems,” in Proc. of the first internationalworkshop on technology and policy for accessing spectrum, 2006, pp.1–9.

[13] T. Yucek and H. Arslan, “A survey of spectrum sensing algorithms forcognitive radio applications,” IEEE Commun. Surveys Tuts., vol. 11,no. 1, pp. 116–130, 2009.

[14] W. Song and W. Zhuang, “Multi-service load sharing for resourcemanagement in the cellular/WLAN integrated network,” IEEE Trans.Wireles Comm., vol. 8, no. 2, pp. 725–735, feb. 2009.

[15] E. Stevens-Navarro and V. W. S. Wong, “Virtual partitioning for con-nection admission control in cellular/WLAN interworking,” in IEEEWireless Communications and Networking Conference, 2008, pp. 2039–2044.

[16] 3GPP TR 25.814, “Physical layer aspects for E-UTRA,” 2006.[17] F. Graziosi and F. Santucci, “A general correlation model for shadow

fading in mobile radio systems,” IEEE Commun. Lett., vol. 6, no. 3, pp.102–104, mar 2002.

[18] I. Forkel, M. Schinnenburg, and M. Ang, “Generation of two dimen-sional correlated shadowing for mobile radio network simulation,” inInternational Symposium on Wireless Personal Multimedia Communi-cation, 2004.

[19] 3GPP TR 36.942 V1.2.0, “E-UTRA radio frequency (RF) systemscenarios (release 8),” June. 2007.

[20] M. Zulhasnine, C. Huang, and A. Srinivasan, “Penalty function methodfor peer selection over wireless mesh network,” in (accepted for publi-cation) Vehicular Technology Conference Fall (VTC), Fall, 2010.

[21] P. Mogensen, W. Na, I. Z. Kovacs, F. Frederiksen, A. Pokhariyal, K. I.Pedersen, T. Kolding, K. Hugl, and M. Kuusela, “LTE capacity com-pared to the shannon bound,” in IEEE Vehicular Technology Conference(VTC), Spring, 2007, pp. 1234–1238.

[22] A. Haider and R. Harris, “A novel proportional fair scheduling algorithmfor HSDPA in UMTS networks,” in The 2nd International Conferenceon Wireless Broadband and Ultra Wideband Communications, 2007, pp.43–49.