A Novel D2D Data Offloading Scheme for LTE Networks

Transcription

1 A Novel DD Daa Offloading Scheme for LTE Neworks Zehua Wang and Vincen W.S. Wong Deparmen of Elecrical and Compuer Engineering The Universiy of Briish Columbia, Vancouver, Canada {zwang, Absrac Downloading remoe files (e.g., picures, videos) from online social neworks via smar user equipmens (UEs) (e.g., smarphones, ables) is becoming popular. Friends who are nearby may wan o download he same files shared by heir muual acquainance. People can obain hese files in a device-o-device (DD) manner via opporunisic connecions o reduce heir paymen for daa service. This is referred o as DD daa offloading. However, DD communicaions on unlicensed specrum using Blueooh or WiFi-Direc may no mainain high daa rae when many DD pairs nearby need o communicae simulaneously. Since DD connecions are ransien, i is imporan o improve spaial reuse of communicaion resources and increase he daa rae of opporunisic DD communicaions. In his paper, we propose a scheme o reuse he downlink licensed specrum of cellular neworks for DD daa offloading. Our proposed scheme includes deermining he availabiliy of digial files on neighbouring devices, esimaing he channel gains, and performing channel allocaion and power conrol for DD pairs. Simulaion resuls show ha our proposed scheme does no affec he exising cellular UEs and i can also offload more daa raffic when compared wih WiFi-Direc on an unlicensed specrum. I. INTRODUCTION Nowadays, accessing online social neworks (OSNs) via smar user equipmens (UEs) (e.g., smarphones, ables) is popular. Mobile users can use heir smar UEs o download picures or videos via Long Term Evoluion (LTE) neworks. The sudy in [] shows ha by, he average size of daa consumed by a smarphone in each monh will reach.7 GB and 69% of mobile daa raffic will be video sreaming. These recreaional files may no need o be downloaded a once. Meanwhile, he sudy in [] shows ha digial files requesed by users are closely relaed wih heir social relaions. Thus, friends who are nearby can obain delay olerable files in a device-o-device (DD) manner by connecing heir UEs opporunisically [3]. This is referred o as DD daa offloading. DD daa offloading using he indusrial, scienific and medical (ISM) specrum has been sudied in [] [5]. On one hand, sharing unlicensed frequency channels wihou coordinaion proocols may cause severe inerference beween muliple ransmier and receiver pairs. On he oher hand, when muliple DD pairs nearby need o communicae simulaneously, disribued coordinaion usually reserves communicaion resources for only one DD pair. Thus, he DD communicaion via unlicensed specrum using Blueooh or WiFi Direc [6] may no suppor parallel DD ransmissions nearby. Since DD connecions are ransien, i is imporan o improve spaial reuse of communicaion resources and increase he daa rae for DD communicaions. We refer o he UEs ha communicae wih he enb as he cellular UEs and refer o he UEs ha communicae in DD manner as DD UEs. The work in [7] allocaes licensed specrum o discover nearby DD UEs by sending he discovery signals. Yang e al. in [] propose o se up DD connecions based on he UE posiions known by LTE neworks. Since licensed specrum in he LTE nework is allocaed by he base saion or he evolved node B (enb) [9], i is convenien o reuse he licensed specrum for DD communicaions by applying resource allocaion and power conrol [] []. In paricular, he work in [] uses relay nodes o conduc DD communicaions. In conras, he works in [] and [] do no require relay nodes o ransfer daa beween DD UEs. In his paper, we use Blueooh or WiFi-Direc o discover neighbouring UEs and deermine he availabiliy of files wihin he neighbourhood. The size of hese signaling packes is acually much smaller han he size of files o be ransferred. Thus, he nearby UEs and he available files in a neighbourhood can be discovered efficienly. Moreover, some delay olerable files may have deadlines. When he deadline is reached, he remaining par of he file has o be downloaded via he LTE nework. To offload more daa raffic from LTE neworks, DD UEs and he enb need o sor and schedule hese daa offloading asks. When he licensed uplink specrum in he LTE nework is reused beween DD UEs, he cellular UE may be required o increase is ransmission power o mainain a arge signal o inerference plus noise raio (SINR) a he enb. Hence, we propose o reuse he licensed downlink specrum of LTE neworks for DD daa offloading. The licensed downlink specrum is allocaed by enb in erms of resource blocks (s). Boh mobile users and wireless service providers can benefi from DD daa offloading. Consider Fig. as an example. There are five UEs (i.e., u, u, u 3, u, and u 5 ) and one enb. The paymen of daa service can be reduced for u and u 3 if hey can obain he digial files from UEs u and u, respecively. On he oher hand, when he channel gains of communicaion links beween DD UEs (i.e., h, and h,3 ) are much higher han he channel gains of inerference links (i.e., h enb,, h, o u, and h enb,3, h,3 o u 3 ), he s used o serve u 5 can be reused by wo pairs of DD UEs wih power conrol [3]. The wireless service provider can allocae hese downlink s o some

2 Fig.. UE h, u h, UE u UE h,3 u 3 UE inerference link communicaion link h enb, h,3 u u5 h enb,5 h UE u5 enb,3 enb Reusing he s in LTE neworks for DD daa offloading. oher cellular UEs and increase is revenue. In his paper, we propose an opimized scheme o reuse he downlink s in LTE neworks for DD daa offloading. Our main conribuions are summarized as follows: We propose o evaluae he uiliy of offloading demands wih heir deadlines. We hen approve he demands by solving a weighed maching problem in a direced graph. We design a scheme o deermine he channel gains of communicaion and inerference links by using Zadoff- Chu sequence []. We formulae an opimizaion problem o joinly allocae s and perform power conrol. Simulaion resuls show ha our proposed scheme can offload more daa raffic when compared wih daa offloading using WiFi-Direc on an unlicensed specrum. The res of he paper is organized as follows. In Secion II, we inroduce he sysem model and presen our proposed scheme wih modules. We formulae and solve problems for hese modules in Secion III. The simulaion resuls are presened in Secion IV. Conclusion is given in Secion V. II. SYSTEM MODEL AND PROPOSED SCHEME We consider each mobile user is equipped wih one UE. The erms mobile user and UE are used inerchangeably. Le U denoe he se of UEs. Le F denoe he se of all digial files. A popular digial file k F may be requesed by muliple users a differen ime wih various delay olerance. We consider a ime sloed sysem. Le Si denoe he se of digial files ha have been compleely downloaded and cached by user i Uin ime slo. Le Q i denoe he se of files ha have been requesed by user i and hose files have no been compleely downloaded in ime slo. Thus, he ses Si and Q i change over ime. Specifically, when a delay olerable digial file k F is firs being requesed by user i in ime slo, file k is included in se Q i and a deadline i,k >is specified by user i. If file k has no been compleely downloaded via DD daa offloading a ime i,k, user i will use he LTE nework o finish downloading he remaining par of file k. When file k has been downloaded compleely a ime slo, he se Q i is updaed by Q i \{k} and he se S i is updaed by S i [{k}. We assume a UE can cache a number of mos recenly downloaded digial files in a firs-in firs-ou (FIFO) manner. Each UE i U can be informed of he exisence of oher UEs in close proximiy by lisening o Blueooh or WiFi- h,5 Direc beacons. Le Ni denoe he se of neighbouring UEs discovered by UE i in ime slo. For each user j Ni \N i, users i and j exchange he ideniy of digial files (e.g., Fig Fig.. Modules on UEs: Local ask scheduling Channel gain esimaion Parallel DD daa ransmission Availabiliy informaion Offloading demand DD pairing DD channel profile Offloading conrol Modules on enb: Global ask scheduling s allocaion & power conrol Basic idea in proposed scheme for DD daa offloading. UE 3 S = 3 UE Q UE Q = S =3 Q UE =, 3,,d, =,3,,d, Uplink in LTE Blueooh/WiFi-Direc Connecion,5,,d, 5,,3,d 5,3 enb UE 5 S = 5 Q =3 5 Examples ha no all daa offloading demands can be approved. uniform resource locaor) in ses Si and Sj o deermine he available files on each oher, which is referred o as availabiliy informaion. Compared wih file ransfer, he size of availabiliy informaion is small and can be exchanged quickly via Blueooh or WiFi-Direc connecions. We now inroduce our DD daa offloading scheme, which is compaible wih he curren LTE neworks. The inpu and oupu for each module in our proposed scheme are presened in Fig. o show a blueprin. Problems in hese modules will be formulaed and solved in Secion III. Wih module, afer deermining he availabiliy of digial files in ime slo, each UE i U selecs one available file from se Q i and uploads a DD daa offloading demand o enb. A DD daa offloading demand is defined as a four-uple in Caresian space U F Z +. Consider a daa offloading demand (i, j, k, i,k) uploaded by UE i as an example. I means UE i requess o download file k Q i from UE j U wih a deadline i,k. The problem for selecing he digial file will be presened in Secion III-A. No all DD offloading demands received by he enb will be approved. Examples are shown in Fig. 3. Consider UEs and are conneced wih UE 3 via Blueooh or WiFi-Direc in ime slo. Assume Q = Q = and boh UEs and have been informed S3 =. Then, he offloading demands (, 3,,,) and (, 3,,,) are uploaded by UEs and o enb, respecively. These wo demands canno be approved simulaneously since UE 3 can only ransmi file o one UE. Similarly, if UEs and 5 upload heir demands o reques files from each oher, only one demand can be approved. Le P denoe all demands received by enb in ime slo. Module on he enb shown in Fig. deermines wha demands in P are approved. We refer o he oupu of module as DD pairing, which is a maching [5, Ch. ] in a weighed direced graph G = (V, A,w ) wih a uiliy funcion w : A 7! R +. Specifically, V = S (i,j,k, i,k )P{i, j}, A = S (i,j,k, i,k )P{(j, i)}, where (j, i) defined on U denoes a direced edge from j o i. The uiliy funcion w is defined as w ( ((j, i)) = e i,k ) for each offloading demand (i, j, k, i,k) P, where < < is a sysem parameer. The proposed uiliy funcion is monoonically decreasing since he uiliy of approving an offloading demand for a file

3 one LTE frame : ms subframe subframe slo slo slo slo LTE slo :.5ms RE for cell-specific reference signal RE for proposed reference signal Fig.. Proposed reference signals are ransmied on resource elemens (REs) in an LTE resource grid. wih small deadline is larger han approving a demand wih a longer deadline. This is because approving a demand wih a smaller deadline is likely o save more s since he file no compleely downloaded by is deadline has o be downloaded via he LTE nework. Using an exponenial funcion in he uiliy funcion is because exending a sufficien large deadline should marginally decrease is uiliy. Le direced graph ˆM =(ˆV, Â, ŵ ) ( ˆV V, Â A, ŵ : Â 7! R + ) denoe he DD pairing informaion in ime slo, which is deermined a enb by solving a maximum weigh maching problem formulaed in Secion III-B. The direced graph ˆM is broadcased from enb o UEs. Only hose UEs in se ˆV are allowed o paricipae in DD daa offloading on licensed specrum, eiher as a DD daa offloading ransmier (if in se T = S {j}) or as a DD daa offloading receiver (if in se R (j,i)â = S {i}). (j,i)â Le Y denoe he se of cellular UEs in ime slo. We now inroduce an approach for UEs in se R [Y o deermine he receiving channel gain from each UE in se T on each channel b B. Specifically, he se B denoes all channels in he LTE nework and each channel in B has he same bandwidh as an. We consider fla fading in each channel and consider he fading over differen channels o be frequency selecive. The idea is o ransmi he feaured reference signals from UEs in se T o UEs in se R [Y. Specifically, all reference signals ransmied by he UEs are generaed by he same roo Zadoff-Chu sequence [] bu wih differen cyclic shifs. Thus, he correlaion of wo reference signals ransmied by differen UEs in se T is zero. Le c j denoe he cyclic shifed sequence generaed on UE j T in ime slo. The sequence c j for UE j can be generaed by a predefined rouine based on he received DD pairing informaion ˆM. Thus, by receiving ˆM and running he same rouine, UEs in se R [Y can be cerain ha sequence c j is ransmied by he UE j. The grid of s in he curren LTE neworks is shown in Fig., where wo resource elemens (REs) are reserved for he ransmissions of our reference signals. We refer o he duraion of an (i.e.,.5 ms[9]) as an LTE slo. An LTE slo is much shoer han a ime slo. Le h, j,i,b denoe he channel gain o be deermined via RE (Fig. ) on in channel b from UE j T o UE i R [Y a LTE slo. We furher denoe g, i,b as he signal received by UE i a on RE. Then, g, i,b is he summaion of T reference signals, i.e., g, i,b = P`T h, `,i,bc`. Since he correlaion of wo Zadoff-Chu sequences wih differen cyclic shifs of he subcarriers same roo sequence is zero, we have c `c j =if ` 6= j and c `c j = C if ` = j, where C is a consan deermined by he roo sequence []. Thus, g, i,b c j = P`T h, `,i,b c`c j = h, j,i,b C and h, j,i,b = g, i,b c j /C hold for each UE j. Meanwhile, he channel gain h, j,i,b from UE j o UE i via RE (Fig. ) can be deermined wih similar approach. We firs apply inerpolaion o esimae he channel gains a oher REs besides REs and. Then, he channel gain from UE j o UE i on he of channel b a, denoed by h j,i,b, is he average value of he channel gains of all REs in channel b. Hence, h j,i,b can be deermined a UE i for each j T. Le h i,b R T + denoe he vecor of he curren receiving channel gains deermined on UE i R [Y for UEs in se T on he of channel b a. We refer o H i = [h i, h i, B ]T R B T + as he DD channel profile of UE i a LTE slo. The DD channel profile, which is he oupu of module 3 on each UE i R [Y, is uploaded via LTE uplink o he enb. The enb is aware of he downlink channel gain from enb o each UE i R [Y a due o he cell-specific reference signals used by he LTE nework. Thus, we can formulae an opimizaion problem in Secion III-C o joinly allocae downlink s and conrol he ransmission power for each UE in se T. The opimizaion problem is solved by module (Fig. ) a enb and he soluion is referred o as offloading conrol. The offloading conrol message conains boh he allocaion and he power conrol informaion on licensed specrum, which is broadcased o hose UEs in se ˆV. Module 5 (Fig. ) on each UE in se ˆV performs daa offloading in parallel by using he allocaed s wih conrolled power o improve he spaial reuse. III. PROBLEM FORMULATION AND SOLUTIONS We have specified he modules 3 and 5 (Fig. ) of our proposed scheme in Secion II. In his secion, we formulae and solve problems for modules,, and menioned above. A. Local Task Scheduling Problem We use he deadline o evaluae he urgency of a daa offloading ask. For each DD UE i U in ime slo, he following problem needs o be solved o deermine which digial file k Q i should firs be downloaded from a neighbouring UE: i,k argmin kq i subjec o k Q i \ [ jn i S j. (a) (b) Consrain (b) implies ha a file k seleced by UE i o download mus be available on one of he neighbouring UEs. We consider each mobile user i Uhas limied number of delay olerable files in se Q i. Problem () can be solved on UE i by comparing he deadline of all digial files in se Q i, wih compuaional complexiy O( Q i P jn S i j ).

4 Algorihm : Deermine he maximum weigh maching ˆM =(ˆV, Â, ŵ ) in direced graph G =(V, A,w ). ini E := ;,  := ;. for (j, i) A do 3 if hi, ji 6 E hen E := E [hi, ji, w (hi, ji) :=w ((j, i)). 5 if hi, ji E and w (hi, ji) <w ((j, i)) hen 6 w (hi, ji) :=w ((j, i)). 7 ( ˆV, Ẽ, w ):=WMA((V, E, w )). for hi, ji Ẽ do 9 if (i, j) A and w ((i, j)) = w (hi, ji) and(j, i) /  hen  :=  [ (i, j), ŵ ((i, j)) := w (hi, ji). if (j, i) A and w ((j, i)) = w (hi, ji) and(i, j) /  hen  :=  [ (j, i), ŵ ((j, i)) := w (hi, ji). 3 oupu ˆM := ( ˆV, Â, ŵ ). B. Global Task Scheduling Problem Le M denoe a V V binary marix, where an elemen m j,i =(or m j,i =) represens ha he daa offloading demand uploaded by UE i o obain daa from UE j is approved (or no). Thus, he global ask scheduling problem a ime can be formulaed as maximize M subjec o (j,i)a m j,iw ((j, i)) V (a) m j,i apple, i V, (b) j= V m j,i apple, j V. (c) i= Problem () can be solved by firs casing G =(V, A,w ) as an undireced graph, where he maximum weigh maching can be deermined in polynomial ime by he weighed maching algorihm (WMA) [6]. Then, we conver he oupu of WMA o a direced graph which can be proven as he maximum weigh maching of he direced graph G. We firs explain Algorihm and hen prove is correcness. In Algorihm, we firs cas he direced graph G as an undireced graph (V, E, w ) (Lines -6). Specifically, we include a weighed undireced edge hi, ji in se E if eiher direced edge (j, i) or (i, j) is in se A and is weigh is eiher w ((j, i)) or w ((i, j)) correspondingly. If boh edges (j, i) and (i, j) are in se A, hen w (hi, ji) is he larger one of w ((j, i)) and w ((i, j)). The weighed maching algorihm proposed in [6] is used by aking he undireced graph (V, E, w ) as he inpu. Is maximum weigh maching ( ˆV, Ẽ, w ) is deermined (Line 7) as he oupu. For each undireced edge hi, ji Ẽ (Line ), i is convered o a direced edge (i, j) or (j, i) by comparing w ((j, i)) and w ((i, j)) wih w (hi, ji) (Lines 9 and ). The chosen one is added in he direced edge se  wih corresponding weigh (Line or ). Theorem : ˆM = (ˆV, Â, ŵ ) is he maximum weigh maching of he direced weighed graph G. Proof: We firs prove he following lemma: Lemma : In a direced graph G =(V, A,w ), if here are wo verices in se V such ha boh direced edges (i, j) and (j, i) are in se A and if w ((i, j)) >w ((j, i)), hen he direced edge (j, i) is no in is maximum weigh maching. Proof: Assume a maching of G conains direced edge (j, i), hen we can compose a new maching by removing direced edge (j, i) and including direced edge (i, j). Due o w ((i, j)) >w ((j, i)), he new composed maching mus have a larger summaion of weighs, which complees he proof. Moreover, since each UE i U selecs one digial file o download from one neighbouring UE j Ni in ime slo, he indegree of each verex in he se V is. Thus, afer removing he direced edge wih he smaller weigh from all -cycles (a cycle [5, pp. 5] wih wo edges) from G, we obain an oriened graph [5, pp. 7] of he undireced graph (V, E, w ) in Algorihm. Since he undireced graph ( ˆV, Ẽ, w ) deermined by WMA is he maximum weigh maching of he undireced graph (V, E, w ), is oriened graph, ˆM =(ˆV, Â, ŵ ) deermined in Algorihm, is he maximum weigh maching of direced graph G. The soluion of problem () is furher validaed wih a random nework scenario in Secion IV-A. C. Join Allocaion and Power Conrol Problem Le D denoe he se of cellular UEs in ime slo. Cellular UEs have o be saisfied wih higher prioriy han he DD UEs. In he LTE neworks, a cellular UE in se D can be assigned one or more s in se B a each LTE slo, bu each can be allocaed o one cellular UE only. Thus, he enb is aware of a funcion D : B 7! D. We consider a arge SINR on b Ba has o be mainained above a hreshold b for he cellular UE d D and d = D(b). If an b Bis no allocaed o any cellular UE, b =. We assume he maximum ransmission power on he enb is P enb and denoe is power componen on b a ime is p enb,b. For DD UE pair (j, i) Â, he maximum ransmission power from he ransmiing UE j T is denoed by P j. Le p j,b denoe he power componen of he DD ransmier j on b a. Then, he join allocaion and power conrol problem o maximize he hroughpu of DD daa offloading for each LTE slo is formulaed as maximize P D, p enb subjec o (j,i)â bb Wlog + p j,b h j,i,b WN +p enb,b h enb,i,b +P`T \{j} p `,bh `,i,b (3a) p enb,b h enb,d(b),b P`T p `,bh`,d(b),b WN + p j,b apple P j, bb b, b B, (3b) j T, (3c)

5 p enb,b apple P enb, bb (3d) p enb,b, b B, (3e) p j,b, b B, j T, (3f) where p enb =(p enb,,...,p enb, B ), P D is a T B marix wih p j,b for each j T, b B, W is he bandwidh of each b Ba, and N is he hermal noise specral densiy. Problem (3) is a nonlinear opimizaion problem, which can be solved by sequenial quadraic programming (SQP) mehod o deermine a local opimal soluion. An inuiive explanaion of he soluion is presened in Secion IV-A. IV. PERFORMANCE EVALUATION We firs validae our proposed scheme and hen presen is performance. The nework seing is as follows. The UEs are randomly deployed in a square coverage region of an enb, where he enb is locaed a he cenre. The DD communicaion range is 3 m. Each UE has a maximum ransmission power 3 dbm [7]. The enb has a maximum ransmission power Was. A user selecs a arge posiion o move wih random velociy from o m/s. A user says a a arge posiion from o min randomly and hen moves o he nex posiion. Since he velociy is relaively small, we consider sec for each ime slo. Each UE has 5 o 35 delay olerable files under reques. These files have random size from 5 o 5 MB and random deadlines from o 9 min. Each UE caches a number of mos recenly downloaded files. We consider boh pah loss and channel fading in our simulaions. For communicaion channels from enb o cellular UEs and all inerference channels, Rayleigh fading is considered. The Rician fading is applied for DD communicaion channels. A. Sysem Validaion We consider 3 DD UEs (he circles) and 5 cellular UEs (he squares) as shown in Fig. 5. Each cellular UE is assigned an a an LTE slo. The index of each cellular UE is he same as he ID of assigned o i. Thus, up o 5 can be reused by each pair of DD UEs wih power conrol. Fig. 5 is auomaically generaed. The head end of each solid arrow represens an offloading reques submied by he UE. We also show he uiliy (given by funcion w in Secion II) of he UE ha submis he reques. The ail end of an arrow represens he UE ha is responsible for sending daa. We firs show he correcness of Algorihm, which is used o approve offloading requess. Maching ˆM is represened by colouring he approved requess and heir uiliies in blue. For example, DD UEs 3 and submi daa offloading requess for files from each oher wih uiliies.337 and., respecively. Meanwhile, DD UE 9 submis a reques wih uiliy.3 for he daa from DD UE, which is approved. Algorihm is correc since i approves he reques ha maximizes he uiliy. We hen show he effeciveness of he channel esimaion approach (Secion II) in Fig. 5. The doed red arrows, which are from each approved DD ransmier o hree cellular UEs, represen he sronges hree inerference e e e 9 6.5e e e 3.e 3.e.e e.7e (m) Fig. 5. An auomaically generaed figure for a random nework scenario in an LTE slo, which is used o validae he proposed scheme. DD UE index Y (m) Resource block () index (dbm) Fig. 6. The power allocaion resul by solving problem (3). links esimaed in he corresponding s. The esimaion is valid since he sronges hree inerference links o cellular UEs are hose hree wih relaively close disance from he DD ransmier. The soluion of problem (3) is shown in Fig. 6. The enb allocaes small ransmission power on s 9,, and for DD UE 5 due o is high inerference channel gains o cellular UEs 9,, and (Fig. 5). However, much higher ransmission power is allocaed on s 7,, 3, and 7 for DD UE 5 since is inerference channel gains o corresponding cellular UEs are small. Similar resuls can be found from Figs. 5 and 6 for oher DD ransmiers. B. Performance Comparison We consider s for our simulaions wih W = khz. We consider.g as he base line o show he daa offloading abiliy of WiFi-Direc wih MHz unlicensed frequency band, where he carrier sense muliple access wih collision avoidance echnique is applied. All he simulaion resuls are obained by hrsimulaion ime. We deploy cellular UEs and increase he number of DD UEs from o in a square coverage region of an enb wih lengh of m. We compare he performance of he proposed scheme wih 5 or cached files on each DD UE. When he required SINR for each cellular UE in each is db, he simulaion resuls are given in Fig. 7. For our proposed scheme and.g, he amoun of offloaded daa increases wih he number of DD UEs since here are more available files nearby and more DD communicaion pairs in he nework. Our proposed scheme ouperforms he daa offloading in.g because muliple DD pairs nearby canno communicae simulaneously in.g since he ransmission beween one pair of DD UEs suspends ohers by he collision avoidance echnique. Thus, some ransien

6 Amoun of offloaded daa (GB) Maximum cached files wih proposed scheme Maximum cached files wih.g Maximum 5 cached files wih proposed scheme Maximum 5 cached files wih.g % = 66.7% more daa is offloaded by increasing he number of cached files from 5 o % = 33% more daa is offloaded % = 59% by our proposed scheme more daa is offloaded by compared wih.g. increasing he number of cached files from 5 o. 6 Number of DD UEs Fig. 7. Resuls wih varying number of DD UEs and wih differen maximum number of cached files on each DD UE. Amoun of offloaded daa (GB) Cellular UEs SINR 7dBwihaskscheduling Cellular UEs SINR db wih ask scheduling Cellular UEs SINR 7dBwihouaskscheduling Cellular UEs SINR db wihou ask scheduling % = 9.3% 6 more daa can be offloaded by using ask scheduling compared wih he case 3.9 % =.% wihou ask scheduling. more daa can be offloaded by using ask scheduling compared wih he case wihou ask scheduling. 6 Lengh of coverage region (m) Fig.. Resuls wih varying size of he simulaion region and wih differen SINR requiremens of cellular UEs. daa offloading opporuniies are wased. However, by joinly allocaing he s and performing power conrol, muliple DD pairs nearby can communicae simulaneously in our proposed scheme. When we have DD UEs, 33% more daa can be offloaded wih he proposed scheme compared wih WiFi-Direc operaed on unlicensed specrum. We also find ha he amoun of offloaded daa is proporional wih he maximum number of cached files on each DD UE. Almos 6% more daa is offloaded when he maximum number of cached files increases from 5 o. We sudy he effeciveness of ask scheduling, i.e., he effeciveness of solving problems () and () in he proposed scheme. We deploy DD UEs and cellular UEs in a square coverage region of an enb wih side lengh varying from m o km. For he case wihou solving problems () and (), we le each DD UE randomly selec a file o upload is daa offloading demand and he enb randomly approving received offloading demands wihou concerning heir deadlines. We also compare he performance of proposed scheme when differen SINR are required by cellular UEs. We assume he required SINR of each cellular UE in each is 7 or db in each se of simulaions. When each DD UE caches files, simulaion resuls are given in Fig.. The amoun of offloaded daa decreases when he lengh of he coverage region increases from m o km. This follows he explanaions for Fig. 7 since he densiy of DD UEs is decreased. By solving problems () and (), almos % more daa raffic can be offloaded compared wih he case wihou ask scheduling. We also find differen SINR required by cellular UEs marginally changes he amoun of offloaded daa. This is because join allocaion and power conrol provide high flexibiliy in daa offloading conrol, which can adap differen SINR requiremens of cellular UEs. V. CONCLUSION In his paper, we proposed a scheme o reuse he downlink s of LTE neworks for DD daa offloading. We presened our sysem modules, where he local ask scheduling problem, global ask scheduling problem, channel esimaion, and join allocaion and power conrol problem are included. We validaed he correcness of global ask scheduling and presened he soluion of he join allocaion and power conrol problem wih a random nework scenario. We also conduced simulaions o evaluae he performance of our proposed scheme. Resuls show ha, wihou violaing he downlink SINR requiremens of cellular UEs, our proposed scheme can offload more daa han using WiFi-Direc in an unlicensed specrum. REFERENCES [] Cisco, Cisco visual neworking index: Global mobile daa raffic forecas updae, 3-. [Online]. Available: hp:// paper c-56.hml [] A. Beach, M. Garrell, S. Akkala, J. Elson, J. Kelley, K. Nishimoo, B. Ray, S. Razgulin, K. Sundaresan, B. Surendar, M. Terada, and R. Han, Whozha? Evolving an ecosysem for conex-aware mobile social neworks, IEEE Nework, vol., no., pp. 5 55, Jul.. [3] E. Başuǧ, M. Bennis, and M. Debbah, Living on he edge: The role of proacive caching in 5G wireless neworks, IEEE Commun. Mag., vol. 5, no., pp. 9, Aug.. [] G. Zyba, G. Voelker, S. Ioannidis, and C. Dio, Disseminaion in opporunisic mobile ad-hoc neworks: The power of he crowd, in Proc. of IEEE INFOCOM, Shanghai, China, Apr.. [5] A.-K. Pieilänen and C. Dio, Disseminaion in opporunisic social neworks: The role of emporal communiies, in Proc. of ACM MobiHoc, Hilon Head, SC, Jun.. [6] WiFi Alliance, Wi-Fi peer-o-peer (PP) echnical specificaion,. [Online]. Available: Direc.php [7] Y. Zhao, B. Pelleier, P. Marinier, and D. Pani, DD neighbor discovery inerference managemen for LTE sysems, in Proc. of In l Workshop on DD Communicaion Wih and Wihou Infrasrucure, Alana, GA, Dec. 3. [] M. J. Yang, S. Y. Lim, H. J. Park, and N. H. Park, Solving he daa overload: Device-o-device bearer conrol archiecure for cellular daa offloading, IEEE Veh. Technol. Mag., vol., no., pp. 3 39, Mar. 3. [9] 3GPP TS 36. V.., Evolved universal erresrial radio access (E-UTRA) physical channels and modulaion, Jun.. [] M. Hasan, E. Hossain, and D. I. Kim, Resource allocaion under channel uncerainies for relay-aided device-o-device communicaion underlaying LTE-A cellular neworks, IEEE Trans. Wireless Commun., vol. 3, no., pp. 3 33, Apr.. [] G. Fodor, E. Dahlman, G. Mildh, S. Parkvall, N. Reider, G. Miklós, and Z. Turányi, Design aspecs of nework assised device-o-device communicaions, IEEE Commun. Mag., vol. 5, no. 3, pp. 7 77, Mar.. [] P. Phunchongharn, E. Hossain, and D. I. Kim, Resource allocaion for device-o-device communicaions underlaying LTE- Advanced neworks, IEEE Wireless Commun. Mag., vol., no., pp. 9, Aug. 3. [3] 3GPP TS 36.3 V.., Evolved universal erresrial radio access (E-UTRA) physical layer procedures, Jun.. [] D. C. Chu, Polyphase codes wih good periodic correlaion properies, IEEE Trans. Inf. Theory, vol., no., pp , Jul. 97. [5] D. Jungnickel, Graphs, Neworks and Algorihms, Fourh Ediion. Springer,. [6] H. N. Gabow, Daa srucures for weighed maching and neares common ancesors wih linking, in Proc. of ACM-SIAM Symposium on Discree Algorihms, San Francisco, CA, Jan. 99. [7] 3GPP TS 36. V.., Evolved universal erresrial radio access (E-UTRA) user equipmen (UE) radio ransmission and recepion, Jul..