Performance Analysis of Device-to-Device Communications with Dynamic Interference Using Stochastic Petri Nets

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 12, NO. 12, DECEMBER 2013 6121

Performance Analysis ofDevice-to-Device Communications with

Dynamic Interference Using Stochastic Petri NetsLei Lei, Member, IEEE, Yingkai Zhang, Xuemin (Sherman) Shen, Fellow, IEEE,

Chuang Lin, Senior Member, IEEE, and Zhangdui Zhong

AbstractIn this paper, we study the performance of Device-to-Device (D2D) communications with dynamic interference. Inspecific, we analyze the performance of frequency reuse amongD2D links with dynamic data arrival setting. We first considerthe arrival and departure processes of packets in a non-saturatedbuffer, which result in varying interference on a link based on thechange of its backlogged state. The packet-level system behavioris then represented by a coupled processor queuing model, wherethe service rate varies with time due to both the fast fadingand the dynamic interference effects. In order to analyze thequeuing model, we formulate it as a Discrete Time MarkovChain (DTMC) and compute its steady-state distribution. Sincethe state space of the DTMC grows exponentially with thenumber of D2D links, we use the model decomposition and someiteration techniques in Stochastic Petri Nets (SPNs) to derive itsapproximate steady state solution, which is used to obtain theapproximate performance metrics of the D2D communications interms of average queue length, mean throughput, average packetdelay and packet dropping probability of each link. Simulationsare performed to verify the analytical results under differenttraffic loads and interference conditions.

Index TermsDevice-to-device communication; performanceanalysis; coupled processor model; stochastic petri nets.

I. INTRODUCTION

DEVICE-TO-DEVICE (D2D) communications are com-monly referred to as the type of the technologies that en-able devices to communicate directly with each other withoutthe infrastructure, e.g, access points or base stations. Bluetoothand WiFi-Direct are two most popular D2D techniques inthe market, bothing working in the unlicensed 2.4 GHz ISM

Manuscript received December 31, 2012; revised March 17 and June 24,2013; accepted September 18, 2013. The associate editor coordinating thereview of this paper and approving it for publication was C.-F. Chiasserini.

This work was supported by the National Basic Research Program of China(973 program: 2010CB328105), the National Natural Science Foundation ofChina (No. 61272168, No.60932003), the Fundamental Research Funds forthe Central Universities (No. 2012JBM003, No. 2010JBZ008), the State KeyLaboratory of Rail Traffic Control and Safety (Contract No. RCS2011ZT005),Beijing Jiaotong University.

L. Lei and Z. Zhong are with the State Key Laboratory of Rail TrafficControl and Safety, Beijing Jiaotong University, China.

Y. Zhang is with the Wireless Signal Processing and Network (WSPN)Lab, Beijing University of Posts & Telecommunications, China.

X. Shen is with the Department of Electrical and Computer En-gineering, University of Waterloo, Waterloo, Ontario, Canada (e-mail:[email protected]).

C. Lin is with the Department of Computer Science and Technology,Tsinghua University, Beijing, China.

Digital Object Identifier 10.1109/TWC.2013.101613.122076

bands. Cellular networks, on the other hand, do not sup-port direct over-the-air communications between user devices.However, with the emergence of context-aware applicationsand the accelerating growth of Machine-to-Machine (M2M)applications, D2D function plays a more and more importantrole since it facilitates the discovery of geographically closedevices and reduce the communication cost between thesedevices. To seize the emerging market that requires D2Dfunction, the mobile operators and vendors are exploring thepossibilities of introducing D2D function in the cellular net-works to develop the network assisted D2D communicationstechnologies [1][4]. The most significant difference betweenthe cellular network assisted D2D communications and thetraditional D2D technologies such as WiFi-direct is that theformer works in the licensed band of cellular networks withmore controllable interference and the base station or thenetwork can assist the D2D user equipments (UEs) in variousfunctions, such as new peer discovery methods, physical layerprocedures, and radio resource management algorithms.

In the emerging new cellular networks such as 3rd Genera-tion Partnership Project (3GPP) Long Term Evolution (LTE),orthogonal time-frequency resources are allocated to differentusers within a cell to eliminate intracell interference. Theintroduction of D2D function may bring two categories ofpotential intracell interference into cellular networks: interfer-ence between different D2D users and interference betweena cellular user and one or multiple D2D users. The formercategory of interference arises when the radio resources arereused by multiple D2D users, while the latter arises when theradio resources allocated to a cellular user are reused by oneor more D2D users. The latter category of interference canbe avoided by statically or semi-statically allocating a groupof dedicated resources to all the D2D users at the cost of re-duced spectrum efficiency. Most of the existing work on D2Dcommunications focus on the design of optimized resourcemanagement algorithms using a static interference model,where each D2D link or cellular link is assumed to be saturatedwith infinite backlogs and constantly cause interference to theother D2D links or cellular links [5][11]. The base stationcan either centrally determine the radio resource allocationof D2D connections along with the cellular connections [4],[5] or let the users perform distributed resource allocationby themselves [3], [6]. In this paper, we focus on the firstcategory of interference (i.e., interference between D2D links)

1536-1276/13$31.00 c 2013 IEEE

6122 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 12, NO. 12, DECEMBER 2013

with centralized resource allocation and study the performanceof D2D communications using a dynamic interference model,where the finite amount of data arrives to the links at eachtime slot. Thus, the links do not always have data to transmitand cause interference to the other links. In order to focuson the dynamic interference, we consider the Full Reuse (FR)resource sharing strategy in this paper, where all the availableresources are reused by all the D2D links with non-emptyqueues in each time slot.

In the queuing model for such system, the service rate ateach queue varies with time due to two factors: the fast fadingeffect of the wireless channel and the dynamic arrival anddeparture of packets which results in the dynamic variationof interference from a link when its status changes frombusy to idle or vice versa. When only the second factor isconsidered, the system can be modeled as a coupled-processorserver, where the service rates at each queue vary over timeas governed by the backlogged state of the other queues. Thecomplex interaction between the various queues renders anexact analysis intractable in general and steady-state queuelength distributions are known only for exponentially dis-tributed service in two-queue systems [12][14]. In [15], [16],the flow-level performance in wireless networks with multiplebase stations is investigated, which can be formulated as acoupled-processor model in the single user class scenario. Dueto the complex nature of the coupled-processor model, [15]derives bounds and approximations for the key performancemetrics by assuming maximum and minimum interference inneighbours of the reference cell and [16] studies the perfor-mance gains of intercell scheduling in a two-cell network and asimplified symmetric network, respectively. In [17], the upperand lower bounds on the moments of the queue length ofthe coupled-processor model are obtained by formulating amoments problem and solving a semidefinite relaxation ofthe original problem. This analysis method is applied in [18]to study the impact of user association policies on flow-level performance in interference-limited wireless networks.Although the semidefinite relaxation method can be appliedto study the coupled-processor model with more than twoqueues, the size of the formulated optimization problem scalesexponentially as the number of queues increases.

In this paper, we consider both the fast fading and dynamicinterference effects to better depict the variations of theservice rate in the practical D2D system. The Finite-StateMarkov Channel (FSMC) model with respect to the Signalto Interference and Noise Ratio (SINR) is constructed foreach link, which not only captures the fast fading effect ofthe wireless channel as in the traditional FSMC model basedon Signal to Noise Ratio (SNR) partition [19], [21], but alsoconsiders the dynamic variation of interference due to thechanging backlogged states of the other active links. Basedon the above FSMC model, we formulate a coupled processorqueuing model for such a system with time-varying servicerates and propose an analytical method to derive the statetransition probability matrix and steady-state distribution ofthe underlying Markov process. However, the scalability ofproposed analytical method with large number of links islimited by the exponentially increasing state space of theMarkov process. Therefore, we use the model decomposition

and iteration approach in Stochastic Petri Nets (SPN) [22],[23] to deal with the coupling between the service rates ofthe different links. Specifically, we formulate the SPN modelfor the above queuing system with multiple D2D links anddecompose it into multiple near-independent subnets, whereeach subnet corresponds to a queuing system with a single linkand is solved separately. Because the subnets are correlated,after solving the problem in each subnet to get its steady-statestatistics, the distributions are exported to other subnets toderive their approximate state transition probability matricesand steady-state distributions, and this is conducted iteratively.Finally, performance metrics such as throughput and packetdropping probability can be obtained from the steady-statedistribution of the Markov process.

In recent years, SPN has been used for the performanceanalysis of wireless networks by some work [24][27]. Perfor-mance evaluation of resource management has been conductedin cellular networks [25], multihop networks [26], and IEEE802.16 networks [27] using SPN. In [24], we have usedthe model decomposition and iteration approach [22], [23]to analyze the performance of opportunistic scheduling incellular networks. However, the interaction between differentsubsystems is due to scheduling instead of interference asin this paper. This model decomposition approach has beenimproved in [28] to deal with the following two potentiallimits: (1) proofs of convergence are usually difficult to obtain;(2) spurious states might be introduced, which do not existin the exact model. Moreover, [29] eliminates the need for aKronecker consistent model decomposition. Since our modelin this paper is a Kronecker-consistent model, and we canprove the convergence of the fixed-point iteration while nospurious states are introduced due to the decomposition, westick to the decomposition method in [22], [23] due to itssimplicity. To the best of our knowledge, this work is the firstone which addresses the problem of subnets interaction dueto dynamic interference in D2D communications.

In [30], the performance of D2D communications in cellularnetworks with dynamic packet arrivals is analyzed using SPN.However, the transition probabilities of the channel statesunder dynamic interference are obtained by Monte-Carlosimulation due to the computation complexity of the numericalmethod. We extend the work by developing an analyticalapproach based on level crossing rate of the wireless channelto derive the transition probabilities of the channel states,which greatly reduce the computation complexity. Moreover,we perform extensive numerical analysis and simulation withvarying number of D2D links to evaluate the performancemetrics such as throughput, delay and dropping probability. Inaddition, we also consider the Batch Bernoulli arrival process.

The remainder of the paper is organized as follows. InSection II, the system model of D2D communications is firstintroduced, followed by the formulation of a coupled processorqueuing model based on the system model. In Section III, thetransition probabilities of the Markov process underlying thequeuing model are analyzed in order to derive the steady-stateprobabilities. In Section IV, we formulate the correspondingSPN model, and use the model decomposition and iterationmethod in SPN to approximate the system performance. Theanalytical results are verified by simulation under different

LAI et al.: PERFORMANCE ANALYSIS OF DEVICE-TO-DEVICE COMMUNICATIONS WITH DYNAMIC INTERFERENCE USING STOCHASTIC PETRI NETS 6123

D2D destination

UE 1

D2D source UE 1

D2D source UE 2

D2D destination

UE 2

D2D link

Potential interfering linkBase

Station

(a) (b)

Queue 2

Server 1

Service Rate

Time slot

Queue 1

Queue 1

Queue 2

Queue 1

Queue 2

D2D link 1

D2D link 2Server 2

Fig. 1. Cellular wireless networks with D2D communications capability (a) resource sharing between two D2D links; (b) queuing model of (a).

traffic loads and interference conditions in Section V. Finally,conclusion is given in Section VI. As many symbols are usedin this paper, Table I summarizes the important ones.

II. THE COUPLED PROCESSOR QUEUEING SYSTEMConsider a cellular wireless network with D2D commu-

nication capability. Fig.1(a) illustrates the case where twoD2D links share radio resources with each other. A D2D linkconsists of a source D2D UE transmitting to its destinationD2D UE. Potential interfering link exists for D2D link 1 (resp.D2D link 2) from the transmitter of source D2D UE2 (resp.source D2D UE1) to the receiver of destination D2D UE1(resp. destination D2D UE2). Since there are two categoriesof links, i.e., D2D links and potential interfering links, all linksmentioned are referred to the D2D links by default in the restof the paper. We consider that each source D2D UE maintainsa queue with finite capacity to buffer the dynamically arrivingdata. An interfering link is potential since it only exists whenthe queue of the source D2D UE associated with its transmitteris non-empty. The transmission rate of each link is determinedby its own SINR, which varies with time due to the fast fadingeffects of its own wireless channel, and also the fast fadingeffects and changing on-off status of its potential interferinglink. Moreover, the transmission rates of different links varyasynchronously over time. Therefore, Fig.1(a) can be modeledas a queuing system as illustrated in Fig.1(b), where there aretwo queues and each one is served by a private single-server,whose service rate is determined by the transmission rate ofthe corresponding link.

Let D = 1, . . . , D denote the set of non-overlapping links.Each link maintains a queue at the source D2D UE, and eachqueue has a finite capacity of K < packets, where packetsare assumed to be of the same size B bits. For each queue i,packets arrives according to Poisson distribution with averagerate i packets/sec. The transmission in the time is slot-by-slotbased and each slot has an equal length T . In each time slot,the resource can be allocated to one or more links, dependingon the resource sharing and scheduling strategies. Althoughthere are various resource sharing strategies between the links,there are two extreme cases which incur the maximum andminimum interference, respectively:

Full Reuse (FR): The links reuse all the available re-sources, causing interference to each other. However, the

links get the largest amount of resources to use. Orthogonal Sharing (OS): The links use orthogonal re-

sources with each other where no interference exists.However, each link gets the least amount of resourcesto use.

In this paper, we focus on the performance of FR strategywith dynamic interference. This resource sharing strategy isnot a practical one, since it may cause excessive amount ofinterference between D2D links. However, it is an extremecase which incurs the maximum interference but achieves thebest frequency reuse. On the other hand, the queuing perfor-mance of the other extreme case, i.e., the OS strategy, can beanalyzed using the method in our earlier work [24]. Since thepractical resource allocation strategies try to achieve the besttradeoff between frequency reuse and interference control, theperformance of the two extreme cases provide lower boundsfor the performance of practical resource allocation strategies.Moreover, the packet-level performance evaluation method ofother practical resource allocation methods can be developedbased on those of FR and OS strategies.

Assume that the instantaneous channel gains of the Dtransmitters Txi on the source D2D UEs of links i D withthe D receivers Rxj on the destination D2D UEs of linksj D remain constant within a time slot, the value of whichat time slot t can be represented by a D-by-D channel gainmatrix Gt, where item Gij,t denotes the channel gain betweenthe transmitter Txi of link i and the receiver Rxj of link j.The channel gain matrices Gt and Gt in different time slotst = t can be different due to the fast fading effects of thewireless channel. Let Ii := {Iji}jD\{i} denote the set ofpotential interfering links of link i, where Iji is the potentialinterfering link from the transmitter of link j D\{i} to thereceiver of link i. Note that the channel gain of link i is Gii,t,while the channel gains of set of potential interfering linksIi are {Gji,t}jD\{i}. Let P = {Pi}i=1,...,D be the D-by-1power matrix that determines the transmission power of everylink i.

In this paper, we assume that a link does not always havedata to transmit, and its transmitter first examines whether thequeue is empty or not at the beginning of every time slot t.Only when the queue is non-empty shall it move the packetsout of the queue for transmission and thus cause interference to


TABLE ISUMMARY OF IMPORTANT NOTATIONS

Category Symbol DefinitionConstant D The number of D2D links

K The buffer (queue) capacity in terms of packetsB Packet lengthT Time slot durationL The number of channel states of a single D2D link in FSMCl The SINR threshold between the l-th state and (l + 1)-th state of a D2D link in FSMCRl The transmission rate of any D2D link in the l-th state of the FSMC model

Set D The set of D2D links i, i = 1, . . . ,DIi The set of potential interfering links of D2D link iJi,t The subset of D2D links excluding link i with non-zero queue length at the beginning of time slot t

Variable i The mean packet arrival rate of the Poisson arrival process of D2D link iIji The potential interfering link from the transmitter of D2D link j to the receiver of D2D link i where j is other than iri,t The instantaneous data rate of D2D link i at time slot tQi,t The queue length of D2D link i at the beginning of time slot ti,t The queue status (empty or non-empty) of the D2D link i at the beginning of time slot t

SINRi,t The SINR value for each D2D link i at time slot tii,t The SNR value of D2D link i at time slot tji,t The virtual SNR value of potential interfering Iji at time slot tHi,t The channel state of the FSMC model of D2D link i at time slot tAi,t The number of packets arrived to link i during time slot t

Hi,t+1 The virtual channel state of the FSMC model of D2D link i at time slot t+ 1, which assumes thatthe queue status of all the other links j D\{i} is the same at time slots t and t+ 1

Vector Qt D-dimensional, where the ith element is Qi,t, i = 1, . . . , Dt D-dimensional, where the ith element is i,t, i = 1, . . . ,DHt D-dimensional, where the ith element is Hi,t, i = 1, . . . ,Di,t D-dimensional, where the jth element is ji,t , j = 1, . . . ,D

Probability p(n,h)

(l,k)Pr.{ Qt+1 = h, Ht+1 = n| Ht = l, Qt = k}

ph(l,k)

Pr.{ Qt+1 = h| Ht = l, Qt = k}pn(l,k,h)

Pr.{ Ht+1 = n| Ht = l, Qt = k, Qt+1 = h}phili,ki

Pr.{Qi,t+1 = hi|Hi,t = li, Qi,t = ki}pni(li,k,h)

Pr.{Hi,t+1 = ni|Hi,t = li, Qt = k, Qt+1 = h}pni(li,v ,w)

Pr.{Hi,t+1 = ni|Hi,t = li, {j,t}jD\{i} = v, {j,t+1}jD\{i} = w}p(li,ni)(v ,w)

Pr.{Hi,t = li, Hi,t+1 = ni|{j,t}jD\{i} = v, {j,t+1}jD\{i} = w}pni(li,v ,w)

Pr.{Hi(t + 1) = ni|Hi(t + 1) = li, {j,t+1}jD\{i} = v, {j,t+1}jD\{i} = w}pli,ni(v ,w)

Pr.{Hi(t + 1) = li,Hi(t + 1) = ni|{j,t+1}jD\{i} = v, {j,t+1}jD\{i} = w}pni(li,ki,hi)

Approximate probability of Pr.{Hi,t+1 = ni|Hi,t = li, Qi,t = ki, Qi,t+1 = hi}pli,ni(ki,hi)

Approximate probability of Pr.{Hi,t = li, Hi,t+1 = ni|Qi,t = ki, Qi,t+1 = hi}pliv Pr.{Hi,t = li|{j,t}jD\{i} = v}l,k Steady state probability, limt Pr.{ Ht = l, Qt = k}ili,ki

Steady state probability, limt Pr.{Hi,t = li, Qi,t = ki}kj,hj Steady state probability, limt Pr.{Qj,t = kj , Qj,t+1 = hj}

Performance Qi The average queue length of link iMetrics T i The mean throughput in terms of packets/s of link i

Di The average packet delay of link ipid The packet dropping probability of link i


the other links. We consider the transmission capability of linki during time slot t as ri,tTB , where ri,t is its instantaneousdata rate in terms of bits/s and is the integer no bigger than. Here, we assume that if a packet could not be transmittedcompletely due to time-slot expiration at the end of time-slott, which can be foreseen at the beginning of this time slot,the entire packet is not transmitted in this time slot but at thenext time-slot (t+1). Although a packet can be truncated fortransmission in a realistic scenario, for example, by the RLCprotocol in the LTE systems, assuming the realistic scenariowill make the queue length a real number instead of an integer,which will result in the infinite state space of the Markovchain. Therefore, we make this approximation in our analysis,while keeping the approximation under control by adjust thepacket size B. This approximation has also been used inexisting work in literature [21], [31], where the service rateis given in terms of packets/s instead of bits/s. If the numberof packets in the queue of link i at the beginning of time slott is less than its transmission capability during time slot t,padding bits shall be transmitted along with the data accordingto the LTE standard. Arriving packets are placed in the queuethroughout the time slot t and can only be transmitted duringthe next time slot t+1. If the queue length reaches the buffercapacity K , the subsequent arriving packets will be dropped.Let Qt = {Qi,t}iD denote the queue length of every link iin terms of packets at the beginning of time slot t, and Ai,tdenote the number of packets arrived to link i during the timeslot t, which is a Poisson distributed stationary process withmean iT . According to the above assumption, the queuingprocess evolves following

Qi,t+1 = min[K,max[0, Qi,t ri,tTB

] +Ai,t]. (1)From the above discussion, the SINR of every link i de-

pends on the subset of other links in the system with non-zeroqueue length at the beginning of time slot t, which is denotedas Ji,t. Let t = {i,t}iD, where i,t = 1(Qi,t > 0)denotes the queue status (empty or not) of the link i at thebeginning of time slot t. Note that t can take 2D possiblevalues v, v = 1 . . . , 2D. Therefore, we have Ji,t := {j D\{i} : j,t = 1}. The SINR value for each link i at timeslot t is given by the following formula:

SINRi,t =PiGii,t

Ni +

jD\{i} PjGji,tj,t(2)

=ii,t

1 +

jD\{i} ji,tj,t,

where ii,t := PiGii,tNi is the SNR value of link i, and Ni is thenoise power on link i. Similarly, ji,t := PjGji,tNi , j D\{i}can be referred to as the virtual SNR value of the link Iji,where it is virtual since Iji is an interference link insteadof a link and it is in fact the interference to noise ratioconsidering the physical meaning.

The corresponding instantaneous data rate ri,t is a functionof SINRi,t. In this paper, we assume that adaptive modulationand coding (AMC) is used, where the SINR values are dividedinto L non-overlapping consecutive regions and if the SINR

TABLE IISINR THRESHOLD AND RATES

channel state index l SINR Threshold (l1) (dB) Rates Rl (Kbs)2 -4.46 213.33 -3.75 328.24 -2.55 527.85 -1.15 842.26 1.75 1227.87 3.65 1646.18 5.2 2067.29 6.1 2679.7

10 7.55 3368.811 10.85 3822.712 11.55 4651.213 12.75 5463.214 14.55 6332.815 18.15 7161.316 19.25 7776.6

value SINRi,t of link i falls within the l-th region [l1, l),the corresponding data rate ri,t of link i is a fixed value Rl,i.e., ri,t = Rl, if SINRi,t [l1, l). Table II gives theAMC scheme in 3GPP LTE systems where L = 16 [20].As an example, l = 2 if SINRi,t [4.46dB,3.75dB),and R2 = 213.3Kbs. Since the AMC function can selectthe appropriate coding and modulation schemes according tothe instantaneous SINR of the wireless channel guaranteeingthat the packet error rate is above an acceptable value, wedo not consider the transmission errors. Although it is aninteresting and challenging research problem to study thetransmission errors due to factors such as imperfect ChannelState Information (CSI), it is outside the scope of this paper.

In order to achieve AMC, we assume that the BS hasknowledge of the channel gain matrix Gt and the queuestatus t of all the D2D links at each time slot t, so thatit can determine the modulation and coding schemes foreach D2D link with non-empty queues and inform the sourceand destination D2D UEs with downlink control signaling.Since network assisted D2D communications have not beenstandardized in 3GPP, there is currently no specific signalingprotocols for resource allocations of D2D connections. In [1],we have proposed a candidate signaling procedure.

Since the SINR value SINRi,t of any link i D can bederived according to (2), the wireless channel for each link ican be modeled as a FSMC with total L states, where Hi,trepresents the channel state of the FSMC model of link iat time slot t. Each state of FSMC corresponds to one non-overlapping consecutive SINR region and a fixed transmissionrate determined by the AMC algorithm. From (2), it canbe seen that the SINR value SINRi,t and thus the channelstate Hi,t of link i depends on the SNR value ii,t of linki and the virtual SNR values ji,t of its interfering linksIji, j D\{i}, and also the queue status j,t of the linksj D\{i}. For any link i D, since both the (virtual) SNRvalues ji,t, j D and the queue status j,t, j D\{i}remain constant within a time slot t, the SINR value SINRi,talso remains constant within a time slot.

There has been a lot of research on the finite state Markov


modeling of wireless fading channels, where interference isnot considered. Compared with these work, our FSMC has twoadditional complicating factors: (1) the fading of the potentialinterfering links of link i; (2) the variation of the set ofinterfering links of link i due to changing backlog status of theother links. For factor (1), the channel gain of an interferinglink in time slot t can be considered as only dependent on itschannel gain in time slot t 1 due to the time correlation asassumed in the existing first-order FSMC models. For factor(2), the queue length of a link at time slot t only depends onits queue length at time slot t1 and the channel state Hi,t1at time slot t 1, as will be discussed later and calculated in(4). Therefore, the channel state Hi,t of link i at time slot tonly depends on the channel gains of link i and its potentialinterfering links at time slot t 1 and the queue length ofthe other links at time slot t 1, which obeys the Markovianproperty.

The above D2D communications system can be formulatedby a coupled processor queuing model as follows. The definedqueuing system consists of a finite number, D, of queuesindexed by i = 1, 2, . . . , D, each of which has a serveri corresponding to a D2D link i. For any i, there is aPoisson distributed packet arrival process with mean iTfixed length packets of B bits, and a finite and discrete-time Markov chain Hi,t with total L states representing theevolution of the channel states of D2D link i. Associated withthe l-th (l {1, . . . , L}) state of the FSMC model of any linki is a fixed service rate Rl bits/sec of server i, which is anon-negative integer and the same for all the links. Since thewireless channels vary with time asynchronously for differentlinks, the transitions of the channel states are link dependentand the channel states Hi,t and Hj,t of any two different linksi = j at time slot t are not necessarily the same. If at time slott the queue i is non-empty with Hi,t in the l-th state, the queuei is served at a deterministic rate Rl, i.e., the queue is servedaccording to an L-state MMDP. For any link i D, sinceSINRi,t depends on {j,t}jD\{i} and thus {Qj,t}jD\{i},the state of Hi,t depends on the set of queues in the systemwith non-zero queue length, which corresponds to a coupledprocessor server.

III. STEADY-STATE SOLUTION OF THE QUEUEING SYSTEMLet Ht := {Hi,t}iD denote the channel states for every

link at time slot t. Let Qt as defined in Section II rep-resent the queue states for every link at time slot t. The(2 D)-dimensional discrete-time Markov chain (DTMC){( Ht, Qt), t = 0, 1, ...} can be used to represent the systembehavior of the above queuing system. The state number ofthe DTMC is ((K + 1) L)D, which grows exponentiallywith the increasing number D of D2D links. In this section,we focus on the derivation method of the exact transitionprobabilities and steady-state solution of the DTMC and leavethe state space explosion problem to the next section, wherethe model decomposition and iteration method in SPN is usedto decompose the DTMC with ((K + 1) L)D number ofstates to D DTMCs each with (K +1)L number of states.Based on the exact method discussed in this section, theapproximate transition probabilities and steady-state solutionof the decomposed DTMCs can be derived.

n,hl,kp

hl,kp

nl,k,hp

1ii i

hDi= l ,kp 1 ii nD

i= l ,k,hp

ii v w nl , ,p

i iv w

l ,n,p iv

lp

v w = v w

/

Fig. 2. The computation of p(n,h)

(l,k), which mainly consists of two com-

ponents. The first component ph(l,k)

represents the transition probability ofthe queue state, and the second component pn

(l,k,h)represents the transition

probability of the server state.

Let p(n,h)

(l,k)be the transition probability from state (l,k) to

state (n,h) of the Markov chain, where l := {li}iD , k :={ki}iD, n := {ni}iD and h := {hi}iD. Note that li, ni {1, . . . , L}, and ki, hi {0, . . . ,K}. We can first decomposep(n,h)

(l,k)into two components as

p(n,h)

(l,k)=Pr.{ Qt+1 = h| Ht = l, Qt = k} (3)Pr.{ Ht+1 = n| Ht = l, Qt = k, Qt+1 = h}

=ph(l,k)

pn(l,k,h)

,

where the first component ph(l,k)

is the transition probabilityof the queue state from Qt = k to Qt+1 = h, given thechannel state Ht = l; and the second component pn(l,k,h) isthe transition probability of the channel state from Ht = l toHt+1 = n, given the queue states Qt = k and Qt+1 = h.

In the rest of Section III, we will first discuss the com-putation method of ph

(l,k)and pn

(l,k,h), respectively, to get

the state transition probability p(n,h)

(l,k)of the Markov chain

of {( Ht, Qt), t = 0, 1, ...} according to (3). The frameworkof computing p(n,

h)

(l,k)in the following part is illustrated in

Fig.2. Then, we will derive the steady-state distribution ofthe Markov chain from its state transition probability matrixand prove that the steady-state distribution exists and is uniqueunder certain conditions in Theorem 3.

A. Transition Probability of the Queue State ph(l,k)

According to (1) and given ri,t = Rli , we have

phili,ki = Pr.{Qi,t+1 = hi|Hi,t = li, Qi,t = ki} (4)

=

Pr.(Ai,t = hi ki + i) if ki > i, hi = KPr.(Ai,t = hi) if ki i, hi = KPr.(Ai,t K ki + i) if ki > i, hi = KPr.(Ai,t K) if ki i, hi = K


where i = RliTB , and Pr.(Ai,t = a) = (iT )a

a! eiT

due to Poisson assumptions.Since the transition probability phili,ki of each link i depends

only on its own server and queue states, we have

ph(l,k)

=

Di=1

phili,ki . (5)

B. Transition Probability of the Server State pn(l,k,h)

According to eq.(2), the value of SINRi,t is determinedby the (virtual) SNR vector i,t := {ji,t}jD and the queuestatus vector {j,t}jD\{i}. Therefore, we have the followingtheorem.

Definition 1: Denote the set of all possible values of Qias SQi , the set of all possible values of Q as SQ :=D

i=1 SQi , which is the Cartesian product of SQi , i D,and the set of all possible values of {Qj}jD\{i} as S iQ :=

jD\{i} SQj . Partition S iQ into 2D1 non-overlapping re-gions S iv , v = 1, . . . , 2D1, such that the subset of queueswith non-zero queue length is identical within each region,i.e., if {Qj}jD\{i} S iv , then {j}jD\{i} = v .

Theorem 1: Qt = k SQi S iv and Qt+1 = h SQi S iw , the values of pni(li,k,h) are the same and can bedenoted as pni(li,v,w).

The proof of Theorem 1 is straightforward from (2). There-fore, we try to derive the value of pni(li,v,w), which equals

pni(li,v,w) =p(li,ni)(v,w)

pliv(6)

where

p(li,ni)(v,w)

= Pr.{Hi,t = li, Hi,t+1 = ni|{j,t}jD\{i} = v,{j,t+1}jD\{i} = w} (7)

pliv = Pr.{Hi,t = li|{j,t}jD\{i} = v} (8)1) Derivation of pliv : Since Hi,t = li, we have

SINRi,t [(li1), li). Therefore, according to (2),i,t at time slot t belongs to the convex polyhedronli := {i|ii (li1)

jD\{i} jij,t (li1), ii

li

jD\{i} jij,t < li , i 0}. The (virtual) SNRregions corresponding to the channel state li and li + 1 areseparated by the hyperplane iili

Dj=1,j =i jij,t = li .

Assume there are two links in the system, an illustrationof the SINR regions and its equivalent (virtual) SNR vectorregions of link 1 when the channel state H1,t = 1, 2, 3 and{2,t} = {1} is shown in Fig.3. Therefore, the steady-stateprobability that Hi,t = li given that {j,t}jD\{i} = v canbe derived as

pliv =

li

f(i)di =

li

jD

(f(ji)dji). (9)

where f(i) is the joint probability distribution function (pdf)of the stationary random process {ji,t}jD , and f(ji) is thepdf of ji,t. The second equality is due to the independence

crosses downward across threshold a given

SINR1,t

1

0

2

1, 1tH =

1, 2tH =

1, 3=tH

11, t

21, t

1

2

3

0

1, 1tH =

1, 2tH =

1, 3=tH

2,{ } {1}t =

a

1/ 1a

SINR1,t crosses downward across threshold 1

11,t 21, 1/ 1t a =

21,t crosses upward across threshold given 1/ 1a 11,t a =

crosses across the hyperplane 11, 21,{ , }t t 11,

121,1t

t

=

+

a takes value from to 1

1

23

from to2 1

Fig. 3. (Virtual) SNR regions corresponding to the channel states and possibletransitions of Hi from li into ni when v = w.

between the r.v. elements in the set {ji}jD . For a Rayleighfading channel with additive white Gaussian noise, the re-ceived instantaneous SNR, ji, is exponentially distributedwith mean ji.

Therefore, pliv can be derived by the integration of amultivariate exponential function over a convex polyhedron,which can be equivalently written as

pliv =

0

( ub(li,Jt)lb(li,Jt)

f(ii)dii) jJt

f(ji)dji,

(10)where ub(li,Jt) = li + li

jJt ji, lb(li,Jt) = li1 +

li1

jJt ji. Therefore, the integration limits of ii can bewritten as affine functions of ji, j Jt, while the integrationlimits of ji, j Jt are all from 0 to . Therefore, the closed-form expression for pliv can be written as

pliv =jJt

1

ji(

exp((li1)/ii)kJt

(1/ki + (li1)/ii)

exp(li/ii)kJt

(1/ki + li/ii)). (11)

2) Derivation of p(li,ni)(v,w): The transition of channel stateHi from li to ni can be due to two categories of factors:(1) the fading of link i and its potential interfering links; (2)the variation of the set of interfering links of link i from vto w due to the changing backlog status of the other links.Therefore, we consider that v = w and v = w, respectively,and derive the value of p(li,ni)(v,w) under each scenario. In theformer scenario, factor (2) doesnt exist and we can focus onthe SINR variation due to the fading effects.

a) When v = w: the values of {j,t}jD\{i} and{j,t+1}jD\{i} remain the same during two consecutive timeslots, and thus the set of interference links {Iji}jJt and


{Iji}jJt+1 remain the same in time slot t and t+1. Therefore,the transition of Hi,t in state li to Hi,t+1 in state ni canonly be due to the variations of the (virtual) SNR vector{ji}j{i}Jt in time slot t and t + 1. In [19], in order toderive the transition probability of the FSMC based on SNRpartition, the product of the SNR level crossing rate and thetime slot interval is used to approximate the joint probabilitythat the channel states are in adjacent states in time slot tand t+ 1, respectively. Furthermore, it assumes that the jointprobabilities that the channel states are in different and non-adjacent states in two consecutive time slots are zero. In thispaper, we use a similar method to approximate the transitionprobabilities of the FSMC based on SINR partition. However,since the state transition of the SINR-based FSMC can be dueto the variations of any of the (virtual) SNR elements in the(virtual) SNR vector {ji}j{i}Jt , the derivation method ismuch more complicated. We first use a similar assumptionas in the SNR-based FSMC that the state transition of Hi,tcan only occur between adjacent states, i.e., p(li,ni)(v,w) = 0, if|ni li| > 1. Then, we try to derive the transition probabilitiesbetween adjacent states as

p(li,li+1)(v,w)

NI(li)T, (12)p(li,li1)(v,w)

NI((li1))T,where NI(li) is level crossing rate in terms of the SINR,i.e., the expected number of times per second the SINR passesdownward across the threshold li . This approximation issimilar to the method in [19]. NI(li)T , whose value issmaller than one, can be explained as the probability that theSINR passes downward across the threshold li in a time slotinterval T .

In order to derive the value of NI(li)T , we considera small time interval t 0. Therefore, NI(li)t can beexplained as the probability that the SINR passes downwardacross the threshold li in a small interval t.

Theorem 2: We find the SINR value of link i crossesdownward the threshold li in a small time interval t ifone of the mutually exclusive and exhaustive eventualities inthe set {Ej}j{i}Jt occurred, where Ej is defined as theevent that the (virtual) SNR ji passes downward (or upward)across a threshold j,li , which equals

j,li =

{1

li(ii li(1 +

kD\{i,j} kij,t)), if j = i,

li(1 +

kD\{i} kij,t), if j = i,(13)

while the other (virtual) SNR values in the set {ki}k{i}Jtremain unchanged. The probability of event Ej can be calcu-lated as

Pr.(Ej) =

Nj(j,li)t

k{i}Jt\{j}f(ki)dki,

(14)where Nj(), j = 1, . . . , D is the level crossing rate of the(virtual) SNR ji at , which is the expected number of timesper second the (virtual) SNR ji passes downward (or upward)across the threshold , and can be calculated according to theDoppler shift fm and the normalized threshold /ji [19].

Proof: Due to the equivalence between the SINR region[(li1), li) and the (virtual) SNR region li when thechannel state Hi,t = li, (li = 1, . . . , L), the SINR valueof link i crosses downward the threshold li if and onlyif the (virtual) SNR vector i passes across the hyperplaneiili

jD\{i} jij,t = li from the convex polyhedron

li+1 to li in a small interval t. Consider a (virtual)SNR ji for j {i}

Jt, given the values of the restof the (virtual) SNR elements in the set {ki}k{i}Jt ,we can derive the (virtual) SNR threshold j,li accordingto (13) so that if ji crosses downward (or upward) j,lithen the (virtual) SNR vector i crosses the hyperplaneii li

Dj=1,j =i jij,t = li from the convex polyhedron

li+1 to li . Since Nj()t is the probability that the(virtual) SNR ji passes downward (or upward) across thethreshold in a small interval t, and the values of the restof the (virtual) SNR elements in the set {ki}k{i}Jt canbe taken over the field of non-negative real numbers as longas j,li > 0 with the pdf function f(ki), the probability ofevent Ej can be derived as (14).

Note that the probability of multiple simultaneous varia-tions of the (virtual) SNR elements in the vector i in asmall interval t are prohibited in the sense that each suchmultiple event is of order o(t). Therefore, we will find theSINR value of link i crosses downward the threshold li , orequivalently the (virtual) SNR vector i crosses the hyperplaneiili

jD\{i} jij,t = li from the convex polyhedron

li+1 to li in a time interval t, if one of the mutuallyexclusive and exhaustive eventualities Ej , j {i}

Jtoccurred.

Example 1: As illustrated in Fig.3 where there are two linksand three channel states of Hi,t, i = 1, 2, we consider the link1 and assume that {2,t} = {2,t+1} = v = {1}, i.e., link1 suffers interference from link 2 in both time slots t andt+1. SINRi,t crosses downward across threshold 1 so thatH1,t = 2 and H1,t+1 = 1. This is equivalent to {11,t, 21,t}crosses across the hyperplane 11,t1+21,t = 1 from 2 to 1,which can only happen when 11,t crosses downward acrossthreshold a while 21,t = a/1 1 remains unchanged or21,t crosses upward across threshold 21,t = a/1 1 while11,t = a remains unchanged, where a takes values over theregion [1,).

According to Theorem 2, we have

NI(li)t =jD

Pr.(Ej). (15)

Let both sides of (15) be divided by t, we can derivethe level crossing rate NI(li) of the SINR value of link i.Combining (15) with (12), we have


11, 1 +t

21, 1 +t

1

2

3

0

1, 1 1+ =tH

1, 1 2+ =tH

1, 1 3+ =tH

1

0

2

3

1, 1=tH

1, 2=tH

1, 3=tH

1

0

2

3

1, 1 1+ =tH

1, 1 2+ =tH

1, 1 3+ =tH

SINR1,t SINR1,t+1

t t+1

1

0

2

3

1, 1=tH

1, 2=tH

1, 3=tH

0

11, t

21, t

1

2

3

1, 1 1+ =tH

1, 1 2+ =tH

1, 1 3+ =tH

11, 1 +t

21, 1 +t

SINR region

2,{ } {0} = =t v 2, 1{ } {1}+ = =t w

2,{ } {0} = =t v 2, 1{ } {0}+ = =t v

2, 1{ } {1}+ = =t w

(virtual) SNR region

Fig. 4. Possible transitions of Hi from li into ni when v = w.

p(li,li+1)(v,v)

Dj=1

0

Nj(j,li)Tk{i}Jt\{j}

f(ki)dki,

p(li,li1)(v,v)

Dj=1

0

Nj(j,(li1))Tk{i}Jt\{j}

f(ki)dki. (16)

b) When v = w: the values of {j,t}jD\{i} and{j,t+1}jD\{i} are different during two consecutive timeslots. Therefore, the transition of Hi,t from state li to Hi,t+1in state ni can be due to not only the variations of the (virtual)SNR vector {ji}j{i}Jt in time slot t and t+1, but also thechange of interference link set from {Iji}jJt to {Iji}jJt+1 .Therefore, we can no longer assume that the state transitionof Hi,t can only occur between adjacent states. Given v andw, we find the channel state Hi,t in li and Hi,t+1 in ni,respectively, if one of the three following (mutually exclusiveand exhaustive) eventualities occurred:

1) that due to (virtual) SNR variations from i,t to i,t+1,we would have found the channel state Hi,t in state liand Hi,t+1 in state li+1 at the beginning of time slot tand t+ 1, respectively, if the value of {j,t+1}jD\{i}remains unchanged as {j,t}jD\{i}, which equals v;however, since {j,t+1}jD\{i} changes to w at thebeginning of time slot t + 1, Hi,t+1 is in state niinstead of state li + 1 with the same set of underlying

(virtual) SNR vector values i,t+1; the multi-transitionpath described above is denoted as Hi,t = li Hi,t+1 = li + 1 Hi,t+1 = ni;

2) that similar to event 1), except that Hi,t+1 = li1, i.e.,Hi(t) = li Hi,t+1 = li 1 Hi,t+1 = ni;

3) that similar to event 1), except that Hi,t+1 = li, i.e.,Hi(t) = li Hi,t+1 = li Hi,t+1 = ni.

Example 2: As illustrated in Fig.4 where there are two linksand three channel states of Hi,t, i = 1, 2, we consider the link1 and assume that {2,t} = v = {0} and {2,t+1} = w ={1}, i.e., link 1 suffers no interference from link 2 in time slott, but it receives interference in time slot t+ 1. We considerthat H1,t is in state 3 and H1,t+1 is in state 1 as shown in theupper part of Fig.(4) and examine the events that can causethis to happen as shown in the lower part of Fig.(4), whichmaps the SINR regions corresponding to the three states ofH1,t to the (virtual) SNR regions. The solid lines mean thatthere is a change in the (virtual) SNR or the queue status,while the dotted lines mean no change has happened. First,assume that {2,t+1} is unchanged as {2,t} = {0}, andH1(t + 1) can transit to the adjacent state 2 or remain inthe same state 3 as H1(t). The (virtual) SNR region 1,t+1corresponding to both states are shown in Fig.(4) as light grayand medium gray areas. Second, since {2,t+1} is {1} insteadof {0}, H1,t+1 is in state 1 instead of state 2 or state 3, and itscorresponding (virtual) SNR region 1,t+1 falls in the regionsurrounded by the bold lines, i.e., the overlapping regions ofthe horizontal stripped area with the light gray and mediumgray areas. The multi-transition paths of the above two eventsare H1,t = 3 H1,t+1 = 3 H1,t+1 = 1 and H1,t = 3 H1,t+1 = 2 H1,t+1 = 1.

Therefore, we have

p(li,ni)(v,w)

=p(li,li+1)(v,v)

pni(li+1,v,w) + p(li,li1)(v,v)

pni(li1,v,w)+ p

(li,li)(v,v)

pni(li,v,w), (17)

where p(li,li)(v,v) (resp. p(li,li+1)(v,v)

, p(li,li1)(v,v)

) is the conditionaljoint probability that Hi,t is in state li and Hi,t+1 is instate li (resp. li + 1, li 1), given that {j,t+1}jD\{i} ={j,t}jD\{i} = v , whose value can be derived from (16).On the other hand, pni(li,v,w) (resp. p

ni(li+1,v,w)

, pni(li1,v,w))represents the conditional probability that Hi,t+1 is in state ni,given that {j,t+1}jD\{i} = v , Hi,t+1 is in state li (resp.li + 1, li 1), and {j,t+1}jD\{i} = w, i.e.,pni(li,v ,w) =Pr.{Hi(t+ 1) = ni|Hi(t+ 1) = li, {j,t+1}jD\{i}

= v, {j,t+1}jD\{i} = w}. (18)

Given that Hi(t + 1) = li and {j,t+1}jD\{i} =v , the (virtual) SNR vector i,t+1 belongs to the convexpolyhedron li := [i|ii (li1)

jD\{i} jij,t+1

(li1), ii li

jD\{i} jij,t+1 < li , i 0}. Sim-ilarly, given Hi(t + 1) = ni and {j,t+1}jD\{i} =w, i,t+1 also belongs to the convex polyhedron ni :={i|ii (ni1)

jD\{i} jij,t+1 (ni1), ii

ni

jD\{i} jij,t+1 < ni , i 0}. Therefore, the


(virtual) SNR vector i,t+1 should belong to the convexpolyhedron li

ni , and

pni(li,v,w) =pli,ni(v,w)

pli+1v, (19)

where

p(li,ni)(v,w)

=Pr.{Hi(t+ 1) = li, Hi(t+ 1) = ni|{j,t+1}jD\{i} = v, {j,t+1}jD\{i} = w}

=

li

ni

f(i,t+1)di,t+1, (20)

pli+1v = Pr.{Hi(t+ 1) = li|{j,t+1}jD\{i} = v}. (21)The denominator of (19) can be derived according to (11).

Similar to (10), the numerator of (19) is also the integration ofa multivariate exponential over a convex polyhedron accordingto (20). However, since

ii [max{lb(li,Ji,t), lb(ni,Ji,t+1)},min{ub(li,Ji,t), ub(ni,Ji,t+1)}

], (22)

the integration limits of ii cannot be written as affine func-tions of ji, j D\{i}. Since integration over an arbitraryconvex polyhedron is a non-trivial problem, and it has beenshown that computing the volume of polytopes of varyingdimension is NP-hard, we present a relatively simple methodto calculate the integration of (20) in the appendix A.

Similar to (19), we can derive the values of pni(li+1,v,w)and pni(li1,v,w). Taking these values into (17), we can derivethe value of p(li,ni)(v,w) when v = w.

Now we have derived both the values of the denominator(by (11)) and numerator (by (16) when v = w and by (17)when v = w) in (6), we can finalize the calculation ofpni(li,v,w) and thus p

ni(li,k,h)

.

Finally, we have

pn(l,k,h)

=Di=1

pni(li,k,h)

. (23)

C. Steady-State Probability of Markov chain { Ht, Qt}Define the transition probability matrix P = [p(n,

h)

(l,k)] and

the steady-state probability matrix = [l,k], where l,k limt Pr.{ Ht = l, Qt = k}. Each element of the transitionprobability matrix P can be derived from (3) combining with(5) and (23).

Theorem 3: The stationary distribution of the Markov chain( Ht, Qt) exists; is unique, and > 0.

Theorem 3 is proved in Appendix B. Then, the stationarydistribution of the ergodic process { Ht, Qt} can be uniquelydetermined from the balance equations

= P, e = 1. (24)where e is the unity vector of dimension (L (K + 1))Dand can be derived as the normalized left eigenvector of Pcorresponding to eigenvalue 1.

(a)

(b)ic iq i s

1ih 2ih iLh

( 1, )L Litr

( ,2)Litr

1, ... ,i D=

( , 1)L Litr

( ,1)Litr

(2, )Litr

(1, )Litr

(1,2)itr

(2,1)itr

3ih

(1,3)itr

... (2,3)itr

(3,2)itr

...

......

...

(3,1)itr

...

(3, )Litr

( ,3)Litr

1, ... ,i D=

Fig. 5. The DSPN Model of FR strategy. (a) The service process in SPN,(b) The queue process in DSPN.

IV. MODEL DECOMPOSITION AND PERFORMANCEAPPROXIMATION USING DSPN

A. The DSPN ModelThe analytical method in the previous section for the multi-

user system faces the challenge of the exponentially enlargedstate space, which makes it unacceptable for a large numberof links. Since directly solving the queuing model suffers thehigh computational complexity, in this section, we formulatethe SPN model of the above queuing system and use the modeldecomposition and an iteration procedure of SPN to simplifythe analysis.

The (2 D)-dimensional discrete-time Markov chain{( Ht, Qt), t = 0, 1, ...} can be seen as a sampled-time Markovchain of a continuous-time semi-Markov process sampled atevery T interval, while the continuous-time semi-Markovprocess can be modeled as a Deterministic Stochastic PetriNets (DSPN). The DSPN consists of a SPN for representingservice processes and a DSPN for representing queuing pro-cesses. The SPN, as shown in Fig.5(a), is composed of Dsubnets and each subnet i corresponds to the L-state Markovmodulated service process of link i. Each subnet is describedby places ({hil}Ll=1) and transitions {tr(l,n)i }Ll,n=1

l =n. The DSPN,

as shown in Fig. 5(b), models the queuing behavior of thelinks and can be characterized by places {qi}iD, and transi-tions {ci}iD , {si}iD. The meanings of all the places andtransitions are described as follows.hil: a place for the l-th channel state of link i.tr

(l,n)i : exponentially-distributed timed transitions for the

channel state transitions of link i. When tr(l,n)ifires, the channel state transits from l to n. Thefiring rate of tr(l,n)i can be derived as

(l,n)i =

pn(l,k,h)

/T , where pn(l,k,h)

can be obtained by (6).Therefore, (l,n)i depends on the queue states of theother links before and after hil transits to hin, i.e.,whether M(qj), (j D\{i}), is equal to or larger


than zero, where M() is a mapping function froma place to the number of tokens assigned to it. Notethat transitions from any channel state hil, l {1, . . . , L} to any other channel states hin, n {1, . . . , L}\{i} are possible as proved in Lemma1.

qi: a place for the queue state of link i.ci: an exponentially-distributed timed transition de-

noting new packet arrivals from link i, with firingrate i. When it fires, one packet arrives at thequeue place qi.

si: a deterministic timed transitions for service pro-cess. When it fires, one packet is transmitted fromthe queue place qi. Its firing rate i depends onthe marking of the places {hil}Ll=1, i.e.,

i = RlTB

/T, if M(hil) = 1, l = 1, . . . , L,(25)

where M(hil) is either 1 or 0, which representswhether link i is in its l-th channel state or not.

B. Model Decomposition and IterationAccording to [22], the original DSPN can be decomposed

into a set of near-independent subnets. By decomposition,the original multiuser system is represented by D subsystems,each of which consists of one subnet in Fig. 5(a) and onesubnet in Fig. 5(b). Obviously, if each subsystem can beanalyzed separately, the model decomposition can significantlyreduce the size of the state space in the analysis and achievesbetter performance in computational complexity. However,unfortunately, such model decomposition is not clean, i.e.,there exist interactions among subsystems. Specifically, forany subsystem i D, the firing rate of transition si depends onthe marking of the places {hil}Ll=1, which in turn depends onthe markings of the queue places qj of those links j D\{i}.As a consequence, the markings of all other subsystems haveto be available at the same time in order to solve the i-thsubsystem, which is not possible in the decomposed model. Inorder to solve this dilemma, we use the fixed point iterationmethod in Stochastic Petri Nets [23]. First, the steady-stateprobabilities of the markings instead of the instant markingsof the other subsystems are used as the input of subsystem i inorder to derive its steady-state probabilities of the markings.Next, fixed point iteration is used to deal with the cycles inthe model solution process.

Let {Hi,t, Qi,t} denote the the sampled-time Markov chainfor the i-th DSPN subsystem. Let i := [ili,ki ] denote thesteady-state probabilities of {Hi,t, Qi,t} , where ili,ki limt Pr.{Hi,t = li, Qi,t = ki}. In order to derive thesteady-state probabilities i of subsystem i, we have to firstderive the transition probability matrix Pi = [p(ni,hi)(li,ki) ]. First,phi(li,ki) can be derived according to (4). Next, we try to analyzethe transition probability of the channel state from li to nigiven the queue states (ki, hi). According to (6), we can derivethe value of pni

(li,k,h). However, since only the queue state

(ki, hi) of subsystem i is given instead of (k,h), we assumethat the steady state probabilities of all the other subsystems

{j}DjD\{i} are known and derive the approximate transitionprobability pni(li,ki,hi) as follows:

pni(li,ki,hi) =

{kj}jD\{i}S iQ{hj}jD\{i}S iQ

pni(li,k,h)

jD\{i}

kj ,hj . (26)

where kj ,hj limt Pr.{Qj,t = kj , Qj,t+1 = hj} is thejoint steady-state probability that the queue length of link jis kj in time slot t and hj in time slot t + 1. Therefore, wehave,

kj ,hj =

Llj=1

phjlj ,kj

jlj ,kj , (27)

where phjlj ,kj can be obtained by (4).Finally, we have

p(ni,hi)(li,ki)

= phi(li,ki)pni(li,ki,hi)

, (28)which gives the transition probability matrix Pi, and thesteady state probabilities i can be derived similar to (24).

Define xj := {kj ,hj}Kkj,hj=0. According to (26), the solu-tion i for the i-th subsystem can be obtained only when themeasure {xj}jD\{i} are known so that the transition matrixPi can be derived, and the value of {xj}jD\{i} dependson the solutions of all the other subsystems {j}jD\{i}according to (27). Obviously, if the D subsystems are solvedsequentially by index, the above requirement cannot be sat-isfied since only {xj}i1j=1 are known when solving the i-thsubsystem. In the following, the fixed point iteration methodis used to solve this problem.

Let {x1, . . . ,xD} be the vector of iteration variables of thefixed point equation

{x1, . . . ,xD} = z({x1, . . . ,xD}), (29)where the function z is realized by solving the D subsystemssuccessively with the subsystem solution method as describedin the previous subsection. That is, the function z can bedecomposed into D independent functions zi, i D, withzi representing the derivation of the solution i from themeasure {x1, . . . ,xD}, and the calculation of the measure xias a function zi of i according to (27), i.e.,

xi = zi(i({x1, . . . ,xD})) (30)

= zi({x1, . . . ,xD}).Obviously, the vector of measures of the D subsystems{x1, . . . ,xD} satisfies (29), which is referred to as the fixedpoint of this equation.

The fixed point can be derived by successive substitu-tion [23]. Let the initial vector of iteration variables be{x01, . . . ,x0D}. Each element of x0i (i D) can be set toan arbitrary value between 0 and 1. In the u-th iteration, wehave

{xu1 , . . . ,xuD} = z({u11 , . . . ,u1D }), (31)where the iteration variables are determined by the func-tion z based on the values of the last iteration, and the


function z is realized by solving the D subsystems succes-sively using the solution method as described above. Specif-ically, in solving the i-th subsystem in the u-th iteration,{xu1 , . . . ,xui1,xu1i , . . . ,xu1D } is used as the input to derivethe value of xui . After that, x

u1i is replaced by xui as input

in solving the rest of the subsystems (from the (i + 1)-th tothe D-th subsystems) during the u-th iteration.

The iteration is terminated when the differences between theiteration variables of two successive iterations are less than acertain threshold value. The convergence of the fixed pointiteration is proved in the Appendix C.

Given , the performance metrics such as the average queuelength, the mean throughput, the average packet delay and thepacket dropping probability can be derived as in [24].

The average queue length of link i equals

Qi =

Kki=0

Lli=1

ili,kiki, (32)

whereL

li=1ili,ki is the probability that Qi,t = ki.

The mean throughput of link i in terms of packets/s canbe expressed as

T i =

Lli=1

Kki=1

Tli,kiili,ki , (33)

where

Tli,ki =

{ ri,tTB /T if ki ri,tTB kiT if ki < ri,tTB

(34)

is the service rate of link i in terms of packets/s whenHi,t = li and Qi,t = ki. It depends on the minimumvalue of the channel transmission capability and theamount of packets in the queue of link i. Note thatthe service rate is 0 when queue i is empty (ki = 0).Therefore, T i is the sum over the whole system statespace of the product between the service rate of link iin state (li, ki) and the probability that the system is instate (li, ki).

The average packet delay of link i can then be calculatedaccording to Littles Law as

Di = Qi/T i, (35)

which is the average amount of time between the arrivaland departure of a packet for link i. Note that the meanthroughput T i equals the effective arrival rate of link i,which is the average rate at which the packets enter queuei.

Let Bili,ki be the random variable which represents thenumber of dropped packets of link i when Hi,t = li andQi,t = ki. Since K + b = Ai,t + max[0, k ri,tTB ],where b is the number of packets dropped during the t-thslot,

Pr.(Bili,ki = b) = Pr.(Ai,t = K+bmax[0, kiri,tT

B]).(36)

Then, the packet dropping probability pid of link i can beestimated as

pid =Average # of packets dropped in a time slot

Average # of packets arrived in a time slot

=

Lli=1

Kki=0

b=0 bPr.(Bli,ki = b)

ili,ki

iT. (37)

V. PERFORMANCE EVALUATIONIn this section, we verify our analytical model under dif-

ferent interference conditions by tuning the length of thepotential interfering links as shown in Fig.6. We considerthe path loss channel model 28 + 40 log10 d [32] for all theD2D links and potential interfering links, where d is thedistance between the transmitter and receiver in meter. Wenormalize the distance between a transmitter and a receiverwith mean SNR equals 0dB to be 1. The distance betweena transmitter and a receiver of a link is denoted as . Notethat we do not require the length of the links to be the samein our analytical model, and this assumption in our networktopology is only to facilitate us to focus on the variation ofthe potential interfering link length. Assume that the distancesbetween the pair of transmitters (resp. receivers) of link i andi + 1 are + (i 1) (i = 1, . . . , D 1). Therefore, thelength of the potential interfering link Iji of link i, wherej > i (resp. j < i), is

2 +

(j1k=i( +(k 1))

)2(resp.

2 +

(i1k=j( +(k 1))

)2). In this way, wecan increase (resp. decrease) the length of all the potentialinterfering links of link i by increase (resp. decrease) the valueof . We add (i 1) to in the distance between thetransmitters of link i and i+1 to ensure that the mean virtualSNR values ji and ki of any two potential interfering linksIji and Iki (k = j D\{i}) of link i are different, as requiredin (39).

The FSMC model has 16 states in total, and the SINRthresholds and the corresponding transmission rates in1.4MHz bandwidth for each service process are given in TableII as defined in the LTE system. The carrier frequency fand the time slot duration T are set to 2GHz and 1ms,respectively. The velocity of the terminals is set to be 3km/hso that the Doppler frequency becomes 5.56Hz. We let thebuffer size K = 50 packets, where the packet length B = 50bits.

We numerically solve the decomposed Markov model usingfixed point iteration and compare the performance measureswith those obtained by discrete-event simulations of a D2Dcommunications system with dynamic packet arrivals and fullfrequency reuse between D2D links. Both numeric method andsimulation are implemented in Matlab and all experiments arerun on 1.93GHz PC with 1.87GHz RAM. We increase thelink number D from 2 to 4 and the state space of Markovmodels before and after decomposition are shown in Table III.In the simulation, we generate Rayleigh fading channels by theJakes Model using a U-shape Doppler power spectrum [33]for every pair of transmitter and receiver. In each simulatedtime slot, packets arrive to every queue according to Poissondistribution with mean T . For those D2D links with non-empty queues, we derive their respective SINR values in thistime slot according to (2), where the channel gains of the D2D


TABLE IIISTATE SPACE, ITERATIONS, AND RUNTIME WITH DIFFERENT NUMBER OF

D2D LINKS

D State space Iters Runtime (s)original decomposed Numerics Simulation

2 1.632e3 816 4 18 334 ( = 1k)970 ( = 100k)

3 1.332e6 816 4 25 657 ( = 1k)1738 ( = 100k)

4 1.086e9 816 5 34 1136 ( = 1k)2400 ( = 100k)

1

2

D

1

2

D

... ...

D2D linkinterference link

+

( )2D + D-1 D-1

...

3 3

Fig. 6. The Network Topology. The distance between the transmitter andreceiver of a link is denoted as . The distance between the pair of transmitters(resp. receivers) of link i and i+ 1 is +(i 1).

links and interfering links are generated by the Jakes Model.The corresponding transmission rates for the derived SINRvalues in this time slot can thus be derived according to TableII. The simulations are run over 105 time slots and the time-average performance measures of every D2D link are obtained.The simulation runtime is given in Table III which varies withboth packet arrival rate and D2D link number D. In thenumerical method, we set the initial steady-state probabilityvector i of every link i to be { 1(K+1)L , . . . , 1(K+1)L},and the initial vector of iteration variables {x01, . . . ,x0D} canbe derived from (27). The number of iterations and runtimefor convergence are given in Table III. It can be observed thatthe runtime for the iterative numerics is much shorter than thesimulation runtime.

Fig.7(a)-7(d) show the mean queue length, mean through-put, mean delay and dropping probability averaged over allD2D links with varying arrival rates for D2D communicationssystems consisting of different numbers of links D, respec-tively. We choose = 0.5, = 0.3 and = 0.01. It can beseen that the numerical results match well with the simulationresults under every configuration. As expected, the systemperformance in terms of all the above four measures degradewith the increasing number of links due to the growing amountof interference to every link. Fig.7(a) reveals that the meanqueue length increases with the packet arrival rate and reachesthe maximum buffer size 50 when the arrival rate reaches100k packets/s. The variations of the mean throughput anddropping probability have a similar trend. Fig.7(b) shows thatthe difference in mean throughput when the numbers of linksD are different grows larger with the increasing arrival rate.

This is because the chances that the potential interfering linkshave data to transmit grow larger and thus the interference op-portunities are increased for every link. The increase in meanthroughput becomes insignificant when the arrival rate reaches100k packets/s for every D. This is not surprising since ithas almost reached the maximum transmission capacity of thesystem and the increasing arrival rate only results in increasingdropping probability. In Fig.7(d), it can be observed that thedropping probability is related to the maximum transmissioncapability of the system. Take D = 4 for example, thedropping probability reaches approximately 84% when thearrival rate is 100k packets/s, which means the system isoverload by a factor of 5.2 (i.e., dropping probability/(1-dropping probability)). On the other hand, Fig.7(b) shows thatthe maximum transmission capability or the maximum meanthroughput is 16k packets/s when D = 4, and 100k packets/sis indeed 5.2 times more than the maximum transmissioncapability of the system. Note that even in this saturated case,the amount of instantaneous interference received by a linkcan still be smaller than that under the infinite backlog trafficmodel, since the probability that the queue of any other linkbeing empty cannot be zero as the Markov chain underlyingthe queuing system is irreducible. Fig.7(c) shows that the delayincreases sharply when arrival rate increases from 1k packets/sto 20k packets/s, and then remains roughly the same whenthe arrival rate further increases. This is because when thearrival rate becomes larger than the maximum transmissioncapability of the system as discussed above, the system wouldhave become overload if not for the packet dropping mech-anism. Therefore, both the mean queue length and the meanthroughput quickly reach their respective maximum values asrevealed by Fig.7(a) and Fig.7(b). By Littles Law, the meandelay also remains the same after that.

Fig.8(a)-8(d) show the performance metrics with varyinginterference link length, where ranges from 0.1 to 0.7 andthe values of and remain the same as above. Thearrival rate is assumed to be 20K packets/s. The analyticalresults and the simulation results are very close, both ofwhich improve with the increasing interference link lengthand decreasing interference from the other links. Note thatwhen the interference link length is large, the performancegap between the scenarios with different numbers of linksbecome small, since the difference in the amount of receivedinterference is small in these topologies.

In the above numerical and simulation experiments, weassume that the packets arrive one at a time (as opposed toarriving in batches) following the Poisson process. However,we can extend the presented queuing model with the BatchBernoulli arrival process by setting Pr.(Ai,t = a) in (4) tobe the probability mass function of a Binomial distribution.We illustrate the numerical and simulation results with respectto dropping probability for D2D communications systemsconsisting of different numbers D of D2D links under BatchBernoulli arrival process in Fig.9 with varying mean packetarrival rate and in Fig.10 with varying interference link length.We observe that the numerical results match well with thesimulation results. Moreover, comparisons between Fig.9 andFig.7(d) and between Fig.10 and Fig.8(d) show that thedropping probability under Poisson arrival process and Batch


(a) mean queue length (b) mean throughput (packets/s)

(c) mean delay (ms) (d) dropping probability

Fig. 7. Performance metrics versus packet arrival rate for D2D communications systems consisting of different numbers D of D2D links ( = 0.5, = 0.3,and = 0.01).

Bernoulli arrival process are quite similar.

VI. CONCLUSION AND FUTURE WORKIn this paper, we have developed a numerical method to

investigate the performance of D2D communications withfrequency reuse between D2D links and dynamic data arrivalwith finite-length queuing. The system behavior is formulatedby a coupled processor queuing model, where the serviceprocess is characterized by a FSMC with each state cor-responding to a certain SINR interval. We first constructthe underlying DTMC of the queuing model and computethe state transition probabilities of the DTMC to derive itssteady-state distribution. Since the state space of the DTMCgrows exponentially with link number, we next formulatea DSPN model of the queueing system and use the modeldecomposition and iteration techniques in SPN to derive theapproximate steady-sate distribution of the DTMC with lowcomplexity. Finally, we obtain the performance metrics ofthe D2D communications from the steady-state distribution ofthe DTMC, whose accuracy has been verified by simulationresults.

In this paper, we focus on the narrowband communicationsscenario with flat fading wireless channels. When considering

broadband communications scenario with frequency-selectivefading channels, since the frequency-selective fading channelscan be turned into multiple parallel flat fading channels by theOrthogonal Frequency Division Multiplex (OFDM) technique,the channel state space in our FSMC model may grow expo-nentially with the increasing number of flat fading channelscorresponding to increasing system bandwidth. Therefore, wewill expand our numerical method for frequency-selectivefading channels using colored Stochastic Petri Nets for stateaggregation. Furthermore, since we have provided both numer-ical methods for the two extreme cases of resource sharingbetween D2D links, i.e., full reuse and orthogonal sharing,we will develop efficient interference management schemes.In addition, we will study two cases in D2D communicationssystem without making the assumption that a pair of sourceand destination D2D UEs always communicate via the samedirect over-the-air link. The first case is that the D2D UEssupport multi-hop transmission capability. Therefore, sincethere can be multiple routes for a flow with a fixed pair ofsource and destination UEs, the transmitting node has theoption to select the channel on which the flow should betransmitted. The second case is when the dynamic D2D mode


(a) mean queue length (b) mean throughput (packets/s)

(c) mean delay (ms) (d) dropping probability

Fig. 8. Performance metrics versus interference link length for D2D communications systems consisting of different numbers D of D2D links ( = 0.5, = 0.01, mean arrival rate is 20K packets/s. Since the length of the potential interfering link Iji of link i, where j > i (resp. j < i) and i, j D, is2 +

(j1k=i( +(k 1))

)2 (resp. 2 + (i1k=j( +(k 1)))2), the interference link length varies with on the x-axis).

selection is considered, where the traffic between a pair ofsource and destination D2D UEs can be transmitted either viathe D2D link or via the cellular uplink to the base station andthen relayed via the cellular downlink to the destination UE.The research on both cases are interesting and challenging.Although we assume that the packets arrive one at a timefollowing the Poisson process, only minor revision is neededto our numerical method if the packets arrive according to theBatch Bernoulli process as validated by simulation in SectionV. Extension of our numerical method to more sophisticatedarrival processes such as the Batch Markovian arrival process(BMAP) will be studied. Finally, it is interesting in applyingour numerical method to the performance evaluation of otherdynamic interference scenarios, e.g., wireless networks withmultiple base stations.

APPENDIXA. Calculation of the integration in (20)

Given {j}jD\{i} = v , {j}jD\{i} = w and v = w,we define Ivi := {Iji|j D\{i} : j = 1} as the subset of

interfering links Iji with j = 1, and Iwi := {Iji|j D\{i} :j = 1} as the subset of interfering links Iji with j = 1.Furthermore, we define Ivwi := Ivi

Iwi , which could be anempty set . Finally, we define Ivwi := Ivi \Ivwi and I vwi :=Iwi \Ivwi , which are the difference of subsets Ivi and Iwi andthe difference of subsets Iwi and Ivi , respectively. Since Ivi =Iwi , the relationship between Ivi and Iwi can be divided intothe following four mutually exclusive and exhaustive cases, asshown in Fig.(11):

1) Ivi = or Iwi = , and in both subcases, Ivwi = ;in the former subcase, we have Ivwi = , while in thelatter subcase, we have I vwi = ; in the following threecases, we implicitly assume that Ivi = and Iwi = ;

2) Ivi Iwi or Iwi Ivi , and in the former subcase, wehave Ivwi = Ivi , Ivwi = , while in the latter subcase,we have Ivwi = Iwi , I vwi = ;

3) Ivi Iwi = Ivwi = , and in this case, we have Ivwi =

Ivi and I vwi = Iwi ;4) Ivi Iwi , Iwi Ivi , and Ivwi = , and in this case, we

have Ivwi = and I vwi = .


Fig. 9. Dropping probability versus packet arrival rate for D2D communi-cations systems consisting of different numbers D of D2D links under BatchBernoulli arrival process ( = 0.5, = 0.3, and = 0.01).

Fig. 10. Dropping probability versus interference link length forD2D communications systems consisting of different numbers D ofD2D links under Batch Bernoulli arrival process ( = 0.5, =0.01, mean arrival rate is 20K packets/s. Since the length of the po-tential interfering link Iji of link i, where j > i (resp. j ni,

and if (li1) ii < li , we have A1 equals li(li1)

f(ii)dii

ii(li1)

1

0

fs1(s1)ds1 ii(li1)

s11

0

fs2(s2)ds2

ii(ni1)

s11

iini

s11

fs3(s3)ds3, (51) and if ii li , s1 iili 1, we have A2 equals

li

f(ii)dii

ii(li1)

1

iili

1fs1(s1)ds1

ii(li1)

s11

0

fs2(s2)ds2

ii(ni1)

s11

iini

s11fs3(s3)ds3,

(52)


else if ii li , s1 < iili 1, we have A3 equals li

f(ii)dii

iili

1

0

fs1(s1)ds1 ii(li1)

s11

iili

s11fs2(s2)ds2

ii(ni1)

s11

iini

s11

fs3(s3)ds3. (53)

3) If li < ni, and if (ni1) ii < ni , we have A1 equals ni

(ni1)f(ii)dii

ii(ni1)

1

0

fs1(s1)ds1 ii(li1)

s11

iili

s11fs2(s2)ds2

ii(ni1)

s11

0

fs3(s3)ds3, (54)

and if ii ni , s1 iini 1, we have A2 equals ni

f(ii)dii

ii(ni1)

1

iini

1fs1(s1)ds1 ii

(li1)s11

iili

s11fs2(s2)ds2

ii(ni1)

s11

0

fs3(s3)ds3, (55)

else if ii ni , s1 < iini 1, we have A3 equals ni

f(ii)dii

iini

1

0

fs1(s1)ds1 ii(li1)

s11

iili

s11

fs2(s2)ds2

ii(ni1)

s11

iini

s11fs3(s3)ds3.

(56)

Therefore, pli+1,ni(v,w) = A1+A2+A3 when li = ni, li > ni,and li < ni, respectively. Combining (39) with the aboveintegrations, and after mathematical manipulation, we have:

1) if li = ni,

pli,ni(v,w) =1

ii[jJ 1

1

ji][

jJ 2

1

ji][

jJ 3

1

ji]jJ 1

jJ 2

jJ 3jiji(F ((li1)) F (li))

kJ 1\{j}[1

ki 1

ji]

kJ 2\{j}

[1

ki 1

ji]

1kJ 3\{j}

[1

ki 1

ji], (57)

where

F (a) =exp(aii )

4iiji(2ajiji + ii(ji + ji))

(ii + aji)(ii + aji)(ii + aji)

1

(ajiji + ii(ji + ji)); (58)

2) if li > ni,

pli,ni(v,w) =1

ii[jJ 1

1

ji][

jJ 2

1

ji][

jJ 3

1

ji]

jJ 1

jJ 2

jJ 3

jiji exp(1

ji)

kJ 1\{j}[1

ki 1

ji]

kJ 2\{j}

[1

ki 1

ji]

1kJ 3\{j}

[1

ki 1

ji](F (ni , (li1))

F ((ni1), (li1)) F (ni , li)+ F ((ni1), li)), (59)

where

F (a, b) =exp(b( 1aji +

1ii

))a33iiji3ji

(ii + aji)(aiiji + bji(ii + aji))

1(biiji + a(bjiji + ii(ji ji)))

(60)3) if li < ni, pli,ni(v,w) can be derived according to (59)

except that (ni1) (resp. ni ) and (li1) (resp. li)switch places with each other.

Although we will not discuss the integration of (20) in case1, 2 and 3 in detail, we will prove the following Lemma, whichwill be used in Appendix B for the prove of Theorem 3.

Lemma 1: In Case 1 and Case 2, pli,ni(v,w) > 0 for anyli, ni {1, . . . , L} satisfying li ni (resp. li ni) whenIvi Iwi (resp. Iwi Ivi ); In Case 3 and Case 4, pli,ni(v,w) > 0for any li, ni {1, . . . , L}.

Proof: In order to prove pli,ni(v,w) > 0, we need to showthat the integration region li

ni is non-empty according

to (20). In case 1, since (41) defines the integration region whenIvi Iwi , we need to verify that for any li, ni {1, . . . , L} satisfying li ni, the upper limit of integra-tion for s3 is not always smaller than its lower limitswhen ii [(li1), li). In (41), if li ni, we haveii (li1) (ni1). Therefore, the upper limits ofintegration for s3, i.e., ii(ni1) 1 is larger than zerowhen ii > (ni1), and thus larger than the lower limitsof integration for s3, i.e., max{0, iini 1}. The scenariowhen Iwi Ivi can be proved in a similar way.

In case 2, since (43) defines the integration region whenIvi Iwi , we need to verify that for any li, ni {1, . . . , L} satisfying li ni, the upper limits of inte-gration for s1 and s3 are not always smaller than theircorresponding lower limits when ii [(li1),). In


(43), the upper limit of s1, i.e, ii(li1) 1 is largerthan zero when ii > (li1), and thus larger than thelower limits of integration for s1, i.e., max{0, iili 1}.Furthermore, since li ni and s1 < ii(li1) 1, theupper limits of integration for s3, i.e., ii(ni1) s11is larger than zero, and thus larger than the lower limitsof integration for s3, i.e., max{0, iini s1 1}. Thescenario when Iwi Ivi can be proved in a similar way.

In case 3, since (45) defines the integration region, weneed to verify that for any li, ni {1, . . . , L}, the upperlimits of integration for s2 and s3 are not alwayssmaller than their corresponding lower limits when ii [max{(li1), (ni1)},). In (45), the upper limits ofintegration for s2 and s3, i.e, ii(li1)1 and

ii(ni1)

1are larger than zero when ii > max{(li1), (ni1)},and thus larger than their corresponding lower limits ofintegration, i.e., max{0, iili 1} and max{0,

iini

1}. In case 4, since A1, A2 and A3 are all larger than zero

when li > ni, li < ni and li = ni, pli,ni(v,w) = A1+A2+A3 > 0 for any li, ni {1, . . . , L}.

B. Proof of Theorem 3We first prove the following lemmas.Lemma 2: The Markov chain ( Ht, Qt) is irreducible, if

K RLT .Proof: We can prove Lemma 2 by showing that for each

transition from state (l,k) to (n,h), there exists a multi-transition path with non-zero probability, which is denotedas (l,k) (n,h). Now we shall verify the following cases:

1) (l,k) (l, k), for any l = {li {1, . . . , L} :Rli T ki}iD.

First, we will prove that p{li+1}iD(l,k,k)

> 0,

p{li1}iD(l,k,k)

> 0 and pl(l,k,k)

> 0. From (23), we onlyneed to prove that p(li+1)

(li,k,k)> 0, p

(li1)(li,k,k)

> 0, andpli(li,k,k)

> 0. According to Theorem 1, this is equiv-alent to proving that p(li+1)(li,v,v) > 0, p

(li1)(li,v,v)

> 0,

and pli(li,v,v) > 0, wherek SQi S iv . This

is true from (6), since pliv > 0, and we havep(li,li+1)(v,v)

> 0, p(li,li1)(v,v)

> 0, and p(li,li)(v,v) > 0 basedon (16).

Then, we will prove that pkl,k > 0. Since Ai,t =ki max[0, ki RliT ] 0, we have pkili,ki > 0from (4) and pkl,k > 0 from (5).

Therefore, we have p({li+1}iD ,k)

(l,k)> 0,

p({li1}iD ,k)(l,k)

> 0 and p(l,k)

(l,k)> 0 from (3),

and there exists a multi-transition path from (l,k)to (l, k) as (l,k) ({li + 1}iD, k) . . . ({li 1}iD, k) (l, k) or(l,k) ({li 1}iD, k) . . . ({li + 1}iD, k) (l, k), where the probabilityof each transition is non-zero.

2) (l, k) (l,h). First, we will prove that phl,k > 0. Since Ai,t =hi max[0, ki Rli T ] 0 when Rli T ki,we have phili ,ki > 0 from (4) and p

hl,k

> 0 from(5).

Then, we will prove that pll,k,h > 0. From (23),we only need to prove that pl

i

(li ,k,h)> 0. According

to Theorem 1, this is equivalent to proving thatpli(li ,v,w)

> 0, where k SQi S iv and h SQi S iw . Since li |v > 0 and from (6), we needto prove that pl

i ,l

i

(v,w)> 0. According to (17), since

we have proved that p(li ,l

i +1)

(v,v)> 0, p

(li ,li1)

(v,v)> 0,

and p(li ,l

i )

(v,v)> 0 in the above discussion, we only

need to show that at least one of three probabilitiespli(li +1,v,w)

, pli(li1,v,w), and p

li(li ,v,w)

is non-zero. From Lemma 1, we have pl

i

(li ,v,w)is always

non-zero irrespective of the relationship between vand w. Therefore, we can prove that pl

l,k,h

> 0.

Therefore, the transition from (l, k) to (l,h) hasnon-zero probability from (3).

3) (l,h) (n,h). The proof is the same with (1).Combining (1), (2) and (3), we can prove that there exits

a multi-transition path with non-zero probability from state(l,k) to (n,h), i.e, (l,k) (l, k) (l,h) (n,h),where Rli T ki. Since K RLT , there always existssuch li that satisfies this condition.

Lemma 3: The Markov chain ( Ht, Qt) is homogeneous andpositive recurrent, if K RLT .

Proof: Since the transition probability matrix P is inde-pendent of the time slot t, the Markov chain is homogeneous[34]. From Theorem 3.3 in [34], ( Ht, Qt) is positive recurrent,since it has finite state space (L(K+1))D and is irreduciblefrom Lemma 2.

From Theorem 3.1 in [34], Theorem 3 is valid if and onlyif the Markov chain ( Ht, Qt) is irreducible, homogeneous andpositive recurrent, which have been proved in Lemma 2 andLemma 3.

C. Convergence of the fixed point iterationAccording to Theorem 2 in [23], in order to prove the

convergence of the fixed point iteration for the decomposedDSPN model as described in (17), it is sufficient to show thatthe following lemma is true.

Lemma 4: The Markov chain (Hi,t, Qi,t) corresponding tothe ith subsystem in the decomposed model is irreducible, ifK RLT .

Proof: Similar to the proof of Lemma 2, we can proveLemma 4 by showing that for each transition from state(li, ki) to (ni, hi), there exists a multi-transition path withnon-zero probability, which is denoted as (li, ki) (ni, hi).Comparing Lemma 4 with Lemma 2, we can prove that thereis a multi-transition path (li, ki) (li , ki) (li , hi) (ni, hi), where Rli T ki, if we prove that pli+1(li,ki,ki) > 0,pli1(li,ki,ki) > 0, p

li(li,ki,ki)

> 0 and pli

(li ,ki,hi)> 0. This is true


by (26), since we have proved that p(li+1)(li,k,k)

> 0, p(li1)(li,k,k)

> 0,

pli(li,k,k)

> 0, and pli

(li ,k,h)> 0 in Lemma 2, and we also

have kj ,hj 0 for any kj , hj SQj , j D\{i} and{kj}jD\{i}S iQ{hj}jD\{i}S iQ

jD\{i} kj ,hj = 1.

REFERENCES

[1] L. Lei, Z. Zhong, C. Lin, and X. Shen, Operator controlled device-to-device communications in LTE-advanced networks, IEEE WirelessMag., vol. 19, no. 3, pp. 96104, June 2012.

[2] G. Fodor, E. Dahlman, G. Mildh, S. Parkvall, N. Reider, G. Miklos,and Z. Turanyi, Design aspects of network assisted device-to-devicecommunications, IEEE Commun. Mag., vol. 50, no. 3, pp. 170177,Mar. 2012.

[3] M. S. Corson, R. Laroia, J. Li, V. Park, T. Richardson, and G. Tsirtsis,Toward proximity-aware internetworking, IEEE Wireless Mag., vol.17, no. 6, pp. 2633, Dec. 2010.

[4] K. Doppler, M. Rinne, C. Wijting, C. B. Ribeiro, and K. Hugl, Device-to-device communication as an underlay to LTE-Advanced Networks,IEEE Commun. Mag., vol. 47, no. 12, pp. 4249, Dec. 2009.

[5] C.-H. Yu, K. Doppler, C. B. Ribeiro, and O. Tirkkonen, Resourcesharing optimization for device-to-device communication underlayingcellular networks, IEEE Trans. Wireless Commun., vol. 10, no. 8, pp.27522763, Aug. 2011.

[6] X. Wu et al., FlashLinQ: a synchronous distributed scheduler for peer-to-peer ad hoc networks, in 2010 Allerton Conf..

[7] H. Wang and X. Chu, Distance-constrained resource-sharing criteriafor device-to-device communications underlaying cellular networks,Electron. Lett., vol. 48, no. 9, pp. 528530, Apr. 2012.

[8] M. Jung, K. Hwang, and S. Choi, Joint mode selection and powerallocation scheme for power-efficient device-to-device (D2D) commu-nication, in 2012 IEEE Veh. Technol. Conf. Spring.

[9] M. Zulhasnine, C. Huang, and A. Srinivasan, Efficient resource allo-cation for device-to-device communication underlaying LTE network,in Proc. 2010 IEEE Wireless Commun. Netw. Conf.

[10] H. Xing and S. Hakola, The investigation of power control schemesfor a device-to-device communication integrated into OFDMA cellularsystem, in 2010 IEEE International Symp. Personal Indoor MobileRadio Commun.

[11] C.-H. Yu, O. Tirkkonen, K. Doppler, and C. Ribeiro, Power optimiza-tion of device-to-device communication underlaying cellular communi-cation, in 2009 IEEE International Conf. Commun.

[12] G. Fayolle and R. Iasnogorodski, Two coupled processors: the reduc-tion to a Riemann-Hilbert problem, Z. Wahr. verw. Geb., vol. 47, pp.324351, 1979.

[13] A. G. Konheim, I. Meilijson, and A. Melkman, Processor-sharing oftwo parallel lines, J. Appl. Probab., vol. 18, no. 4, pp. 952956, 1981.

[14] F. Guillemin and D. Pinchon, Analysis of generalized processor-sharingsystems with two classes of customers and exponential services, J.Appl. Probab., vol. 41, no. 3, pp. 832858, 2004.

[15] T. Bonald, S. Borst, N. Hegde, and A. Proutiere, Wireless dataperformance in multi-cell scenarios, in Proc. 2004 ACM Sigmetrics

Documents

Performance Analysis of Device-to-Device Communications with Dynamic Interference Using Stochastic Petri Nets