Information Caching Strategy for Cyber Social Computing Based Wireless Networks

Transcription

1 Received 4 December 26; revised 23 Februry 27; ccepted 2 April 27. Dte of publiction 28 April 27; dte of current version 6 September 27. Digitl Object Identifier.9/TETC Informtion Cching Strtegy for Cyber Socil Computing Bsed Wireless Networks XING ZHANG, (Senior Member, IEEE), YONGJING LI, YAN ZHANG, (Senior Member, IEEE), JIAXIN ZHANG, HAILING LI, SHUO WANG, AND DANYANG WANG X. Zhng is with the Beijing Advnced Innovtion Center for Future Internet Technology, Beijing University of Technology, nd the Key Lbortory of Universl Wireless Communictions (Ministry of Eduction), Beijing University of Posts nd Telecommunictions (BUPT), Chin Y. Li, J. Zhng, S. Wng, nd Dnyng Wng re with the School of Informtion nd Communiction Engineering, Beijing University of Posts nd Telecommunictions (BUPT), Beijing, 876, Chin Y. Zhng is with University of Oslo nd Simul Reserch Lbortory, Norwy H. Li is with CNCERT/CC & Institute of Informtion Engineering, CAS, Beijing, 29, Chin CORRESPONDING AUTHOR: X. ZHANG (zhngx@ieee.org) ABSTRACT Cyber socil computing hs brought gret chnges nd potentil intelligent technologies for wireless networks. Among these technologies, informtion cching strtegies re promising pproches to chieving lower dely, higher throughput nd energy efficiency (EE) of user equipment (UE) in 5G wireless networks, by deploying intelligent cching nd computing t the mobile edge. However, the sttic informtion cching strtegies ignore the relevnce of trffic fluctution mong different bse sttions (BSs) nd the vrince of users interests. Thus in this pper, n informtion cching strtegy for cyber socil computing bsed wireless network is proposed, tking dvntges of two lyer socil cyberspces in both trffic correltion between BSs nd the socil reltionship between UEs. In the first lyer, bse sttion socil network (BSSN) is constructed bsed on the socil reltionship between BSs, which is defined s socil-tie fctor (STF). In the second lyer, the Indin Buffet Model (IBM) is used to describe the socil influence of one UE to nother. To reduce bse sttion s trffic lod, users with similr socil interest cn shre the contents they cched with ech other. Therefore, device-to-device (D2D) communiction is tken s the underly to cellulr networks in our proposed informtion cching strtegy. By utilizing the socil chrcteristic of BSSN, the very importnt BSs (VIBSs) with higher verges STF re selected. Then the norml smll cells (NSCs) within the VIBS s coverge re linked to the VIBSs only nd the other unique smll cells (USCs) will be routed bck into the core network (CN) directly. Limited cche nd bckhul cpcity in the whole network re only shred by VIBSs nd USCs. UEs will communicte with ech other vi D2D links only if they hve i) similr interests, ii) enough encounter durtion between users nd iii) re djcent with ech other. Otherwise, the UE shll obtin the required contents vi cellulr networks. With the tool of stochstic geometry process, key performnce indictors, e.g., coverge probbility, network throughput power consumption, EE, dely nd offloded trffic re studied. Both the theoreticl nd numericl results show tht the proposed informtion cching strtegy for cyber socil computing cn chieve high coverge probbility, throughput, EE nd dely by optimizing STF threshold, VIBS coverge nd D2D communiction rdius. INDEX TERMS Informtion cching strtegy, socil cyberspces, D2D communiction I. INTRODUCTION Nowdys, more nd more people surf on the Internet vi vrious smrt mobile terminl devices such s phones, ipds, lptops, etc, nd people tke most of the entertinment on cyber such s wtching videos, reding news, plying gmes, chtting with friends, etc. Therefore, the mount of trffic dt in mobile cellulr networks increses exponentilly. To stisfy the incresing trffic requirement, improving the network throughput needs to be considered preferentilly [], [2]. Among ll the solutions, deploying intelligent cching nd computing t the mobile edge is efficient nd ble to cope with the demnd [3]. By deploying cche t BSs, users cn quickly obtin the informtion they required. However, the sttic informtion cching strtegies re not suitble VOLUME 5, NO. 3, JULY-SEPT ß 27 IEEE. Trnsltions nd content mining re permitted for cdemic reserch only. Personl use is lso permitted, but republiction/redistribution requires IEEE permission. See for more informtion. 39

2 becuse of people s interests chnging nd the populrity of certin content vries. Thus, studying the cyber users socil interest nd behvior is necessry nd meningful. Erlier, the socil network refers to the socil network sites (SNSs) like MySpce, Fcebook, Twitter nd Instgrm where people connect with ech other, shre contents nd disseminte informtion on cyber [4]. Therefore, understnding how users behve when they re connecting SNSs is more useful in studying socil influences, improving the design of content distribution systems [7]. Some studies utilized the dt of these SNSs to study the socil reltionship nd impcts mong users, the ptterns of users socil interest nd behvior re summrized [8] []. In [8], the uthors conceive wireless virtul socil network which describes the wy people seeking informtion vi socil network. A model bsed on Indin Buffet Process is proposed to reflect the influences of one user s choice to the others nd the distribution of contents in the users online socil networks (OSN) [9]. People s predictble socil ptterns re exploited to improve the content delivery performnce nd lower endto-end dely in time criticl ppliction in []. Thus, cyber socil computing is significnt to study the socil ties mong users, nd how strong these ties re. As the users behvior hve gret impcts on the vrition of trffic [5], they re nlyzed to chieve better performnce of networks. Besides, D2D communiction hs emerged to support the trnsmission of contents between UEs, nd become n underly to cellulr networks. D2D communiction underlys over cellulr systems hs become promising technique to overcome the imminent wireless cpcity crunch [6]. The mjority of the trffic in cellulr pertins to the downlod of populr content such s videos or mobile pplictions [], nd offloding these downlods to the D2D tier cn reduce the lod in the cellulr network s infrstructure [8]. When two users re involving D2D communictions, they should stisfy the following requirements: Strong socil tie, high similrity of interest nd re djcent to ech other [2]. Besides, the encounter durtion for the two users in D2D communiction must be long enough to trnsmit the contents. Mny studies ttempt to model the distribution of the cll holding time, which exhibits the sme property with encounter durtion between individuls, nd the gmm distribution hs been shown to be n ccurte model [3], [4]. Becuse the socil interest nd behvior of users influence the trffic fluctution nd cched contents of BSs, the socil network of BSs is built to exploit the socil chrcteristic nd reltionship between BSs [5]. In [6] the trffic fluctution hs lredy derived s n importnt socil chrcteristic to indicte the similrity of trffic vrition of BSs. By utilizing the socil reltionship between BSs, the cche cn be ppropritely deployed to improve the network performnce. In this pper, we propose n informtion cching strtegy for two-lyer socil cyberspce, which is illustrted in Figure. The top lyer represents the BSSN, nd the lines with different colors indictes the socil reltionship mong FIGURE. Two-lyer socil cyberspce. BSs, BSs connected with green lines hve stronger socil-tie thn the BSs connected with purple lines, nd the BSs connected with pink lines hve the strongest thn other BSs. The socil chrcteristic of BSs is exploited by utilizing the bse sttion socil network (BSSN) in our previous study [], nd the socil reltionship between BSs is modeled s the sociltie fctor (STF) bsed on the dt collected from prcticl network. The BSs re divided into three kinds: Very importnt bse sttions (VIBSs) re selected with the STF higher thn the threshold, nd the norml smll cells re selected within the VIBSs coverge, ll the rest BSs in the network re unique smll cells (USCs). We ssume the network is bckhul nd cche cpcity limited. As the VIBS hs higher STF, it could represent the trffic vrition of other BSs. The USC hs less similrity with other BSs so tht they cnnot ssocite with other kinds of BSs, but while the NSCs cn be served by VIBSs becuse of the physicl loction. Tking ll these into considertion, the VIBSs nd the NSCs shre the limited cche nd bckhul cpcity together. The other lyer of the socil cyberspce is the user lyer which is reflected s the bottom lyer of Figure. Different users hve different interests, some might be interest in pinting or reding while others re not. Users connected with lines in different colors symbolizes different chrcteristics 392 VOLUME 5, NO. 3, JULY-SEPT. 27

3 in socil interests. To reflect the socil chrcteristics of users, we dopt the Indin Buffet Model proposed in [9], the uthor hs lredy proved the model perfectly fits the rel dt in the simultion. The users require the contents in order, nd ech user cn require the old content which hs been required by former users, or the new content which hs no history of requirement. Ech user s requirement is influenced by the former users choice, which is similr with people choosing dishes when eting buffets. In our proposed informtion cching strtegy, we ssume tht the D2D communiction users cn lwys get the old contents from their D2D prtners. Other requirements of D2D communiction re: The two users encounter durtion must be longer thn the minimum time of content trnsmission nd re djcent to ech other. When user requires content from the work, the BS will first identify whether the required content pertins to n old content or new content. If the user requires n old content, it will try to find D2D prtner first if there exists one stisfying ll the requirements, nd he will tke D2D communiction, otherwise, the user gets the content from the cellulr network. The min contributions of this pper re summrized s follows: A two-lyer socil cyberspce for both trffic correltion between BSs nd the socil reltionship between UEs is constructed. A bse sttion socil network bsed on the STF between BSs nd the Indin Buffet Model is used to model the socil UEs mutul influence. We lso utilizes D2D communiction for UEs who cn shre the loclly stored content to reduce the trffic lod t the bse sttion. Specificlly, the UEs will communicte with ech other vi D2D links only if certin fvorble conditions re met. Otherwise, the UE shll obtin the required contents vi cellulr networks. We proposed n informtion cching strtegy bsed on the two-lyer socil cyberspce which tkes dvntges of the fetures of two lyers, nd by cyber socil computing, the most populr contents of the next moment cn be predicted nd the contents cched in BSs cn be dynmiclly replced. Tking the D2D communiction s the underly of cellulr network, the trffic lod t BSs cn lso be reduced by offloding it to the user lyer. Key performnce indictors, e.g., coverge probbility, cche hitting probbility, network throughput, power consumption, EE, network verge dely nd trffic offloded by D2D communiction re derived. The theoreticl nd numericl results show tht our cching strtegy cn chieve high coverge probbility, network throughput nd energy efficiency by optimizing the STF threshold, VIBSs cover rdius nd D2D communiction rdius. The rest of this pper is orgnized s follows: Section II presents the BSSN model nd Indin Buffet Model, nd users encounter durtion is modeled s Gmm distribution. The network model of our cching strtegy is lso presented nd contents ccess sitution of users is discussed. Key performnce indictors such s coverge probbility, the cche hitting probbility, network throughput, trffic offloded by D2D communictions, power consumption, verge dely nd EE re derived in Section III. Theoreticl nd numericl results re obtined nd nlyzed in Section IV. Finlly, conclusions re drwn in Section V. II. SYSTEM MODEL In this section, we will briefly present our previous study on BSSN nd STF model. The Indin Buffet Model nd the distribution of users encounter durtion re lso introduced. Lter, the network model of our cching strtegy is presented nd the contents ccess situtions of users re discussed. A. BSSN AND STF MODEL In our previous work [7], the BSSN is constructed to visulize the hidden reltionship between BSs by utilizing the theory of socil network. The STF is modeled to indicte the strength of the reltionship. The STF is defined s: r ij ¼ E S ð i mðs i ÞÞ S j m S j ; () sðs i Þs S j where S i S j is the trffic volume series of BS iðþ:m j ð Si Þis the men of S i, sðs i Þ is the stndrd devition of S i nd E is the expecttion. r ij vries between nd þ, where is totl positive correltion, is no correltion, nd is totl negtive correltion [6]. The result of curve fitting ppliction in MATLAB shows tht the probbility density function (PDF) of STF follows the Gussin distribution with the goodness of fit: SSE ¼ 6:25e :5, men vlue m ¼ :2829, nd the vrince vlue s 2 ¼ :655. B. USERS ENCOUNTER DURATION Mny studies hve reserched the users behvior nd the moving trck to predict the mobility of users. In [9], the uthors identify tht humn mobility shows very high degree of temporl nd sptil regulrity, nd tht ech individul returns to few highly frequented loctions with significnt probbility. The encounter durtion mong individuls follows continuous distribution, nd exhibits the sme property with cll holding time. Thus, some studies find out tht the gmm distribution Gðk; uþ hs shown to be n ccurte model [9], [3], [4]. k nd u re two prmeters tht define the shpe of the distribution. Assuming X n is the contct durtion of UE i nd UE j, nd N i; j is the number of encounter times, the estimte of the expected contct durtion length M i; j is M i; j ¼ P n X n=n i; j, nd P the vrince I i; j which reflects the fluctution is: I i; j ¼ n ðx n M i; j Þ 2 =N i; j. The encounter durtion distribution is derived s X G ðk; uþ ¼ GðMi; 2 j =I i; j; I i; j =M i; j Þ, nd the PDF is given fðx; k; uþ ¼ u k GðÞ k xk e x u : (2) VOLUME 5, NO. 3, JULY-SEPT

4 require the contents just like of which they choosing the dishes t buffet in the resturnt. Assuming there re N users in the networks, nd K files to be required, K ¼ K h þ K, where K h represents the number of files tht hve been required by previous users, nd K represents the number of files tht do not hve required history. Users re rnked from to N, the first user selects ech file with equl probbility of "=K nd ends up with the number of files following Poisson ðþ " distribution, where " is the prmeter which determines the probbility of whether the user chooses to select content or not. For the subsequent users n ¼ 2;...; N, the probbility of hving file k lredy belongs to previous UEs is m n k =n, where m n k is the number of users prior to n who select dish k [2]. UE n will lso require m n new files which re not required by the previous users following Poisson ð"=nþ, which is proved in the Appendix A of [9], so tht the number of old files required by customer n is given s m h n ¼ XK k¼ m n k n ¼ m n m n : (4) FIGURE 2. System model. Thus, the probbility of qulified contct durtion is given by Z Xmin w i; j ¼ fðu; k; u Þdu ¼ g k; X min u ; (3) GðÞ k where X min is the miniml contct durtion required for trnsmitting one content dt pcket successfully [], since our nlysis of the network is more meticulous thn trditionl heterogeneous network tht we cn dig ech user s interestin requiring content, so tht the successful D2D communiction requires D2D users trnsmitting t lest one content dt pcket nd the miniml contct durtion must be considered in successful D2D communiction. gðk; X Xmin min u Þ¼R u t k e t dt is the lower incomplete Gmm function. C. INDIAN BUFFET MODEL The Indin Buffet Model is bsed on the Indin Buffet Process (IBP) [8] which is importnt for lerning the content populrity distribution nd predicting the content tht users my require. The IBP is stochstic process which models resturnt problem where ech dinner smple is from some subset of n infinite selection of dishes on offer t buffet [9]. The first customer will choose its preferred dishes eqully s the dishes re ll new to him. However, once the first customer complete the choice, the following customers will be influenced by the feedbck of the first customer. Therefore, the probbility of ech dish being selected by the following customers re chnged becuse of the previous customers feedbcks. The uthor in [9] utilizes the IBP to describe the users socil influence nd the circumstnces tht users select nd D. NETWORK MODEL As shown in Figure 2, the downlink trnsmission in this network is considered. By dopting the stochstic geometry theory, BSs re modeled s independent homogeneous Poisson Point Processes (PPPs) denoted s F B, with corresponding density of. Assuming the VIBSs, the NSCs nd the USCs re ll distributed s PPPs nd will be explined in next section. The users re positioned with PPP F u with the density of UE. A threshold of STF r is set to select VIBSs. A VIBS s cover rdius is set to select NSCs which re locted within the VIBSs coverge. NSCs re linked to VIBSs with limited fronthul while USCs nd VIBSs occupy the limited bckhul to the core network. Users will get required contents from either D2D communiction or cellulr network which is depended on the judgement mde by BSs. The stndrd pth loss propgtion model is used with pth loss exponent > 2. All the links in this network re Ryleigh fding chnnels following exponentil distribution with men of : h b;i expðþ, h j;i expðþ. We ssume the network is n interference-limited scenrio, so tht the dditive white Gussin noise s 2 is neglected. E. CONTENT ACCESS The D2D communiction is tken s the underly of the cellulr network in our proposed informtion cching strtegy. We ssume the users require contents in the set order. Before user n get the required content, the serving BS judges whether the content hs history of requirement by previous users. If it belongs to the old contents, the user n will try D2D communiction first. He will serch the D2D prtners within the rnge of set rdius, fter tht, the user n will select the best prtner who hs enough encounter durtion within the rnge, nd they become D2D pir. At the sme time, user n gets the old content from his D2D 394 VOLUME 5, NO. 3, JULY-SEPT. 27

5 prtner. If user n required new content or he could not find D2D prtner stisfied ll the D2D communiction requirement, he will obtin the content from cellulr network. Thus, the content ccess cses re summrized s following: The user n ðn 6¼ Þ obtins the old content from his D2D prtner by tking D2D communiction. The user n ccesses to the VIBS, nd obtin the old or new content from locl cching spce of the VIBS. The user ssocited with the VIBS, but the required content is not cched in the BS, so tht the content will be fetched by the VIBS from the core network vi bckhul links. The user ccesses to n NSC, the request is sent to its corresponding VIBS, nd the user gets the content from the cche in corresponding VIBS. The user ccesses n NSC, but the required content is not cched in the corresponding VIBS, so tht the user obtined the content from the core networks fetched by the VIBS vi bckhul links. The user ccesses to USC nd obtined the content from the USC s locl cche. The user ccesses to USC, but the required content is not cched in the BS, so tht the content will be fetched by the USC from the core network vi bckhul links. III. PERFORMANCE ANALYSIS In this section, we will first nlyze the probbility of new nd old contents the nth user my require nd the proportion of different kinds of BSs. Some key performnce indictors including coverge probbility, the cche hitting probbility, network throughput, trffic offloded by the D2D lyer, power consumption, verge dely nd EE will be derived. A. PROBABILITY OF REQUIRING NEW AND OLD CONTENTS As is mentioned in Section II, the nth user requires m n new files, nd the m n Poisson " n,ndthemenvlue is n ".Thetotlnumberoffiles required by user n denotes s m n, nd m n Poisson ðþ, " the expecttion vlue is ", so tht the expected number of old contents tht the nth user requires is m h n ¼ Efmh n g¼efm ng Efm ng¼ð ð n Þ=n Þ": (5) Bsed on the nlysis of [9], we know tht the number of old contents is the difference of two Poisson distribution nd follows the Skellm distribution. Although the Probbility mss function (PMF) nd the PDF of the number of old contents is given in [9], our priority of considertion in this pper is the cche deploying nd replcing strtegy, nd how to improve the performnce of the whole network. The probbilities of new nd old contents required by users only hve impcts on the proportion of the number of users tking D2D communiction or ssociting with the FIGURE 3. Requirements for successful D2D communiction. cellulr network. Thus, we simplify the probbilities of the new contents nd the old contents by utilizing the men vlues of m n, m n nd m h n, which represent the verge probbility for the requirements of user n. The probbility tht the content required by user n is lso required by previous users is P oc ¼ mh n ¼ Efmh n g m n Efm n g ¼ ððn Þ=nÞ" " ¼ n ; (6) n nd the probbility tht the required content does not hve required history is P nc ¼ Efm n g Efm n g ¼ ð=nþ" ¼ " n : (7) B. PROBABILITY OF D2D COMMUNICATION In our network, the successful D2D communiction requires similrity in socil interest, djcent physicl loction nd enough encounter durtion s is shown in Figure 3. The Indin Buffet Model gives the probbility tht two users require the sme content which represents the similrity of users socil interest. As the loctions of the users follow the PPP distribution with the density UE, the probbility of the typicl user cn find potentil D2D prtner within the D2D communiction distnce is P UE ¼ e UEpr 2 UE; (8) where r UE is the distnce between the typicl user nd ny other users in the network, the PDF of r UE is f rue ðþ¼ r 2p UE e p UEr 2. The probbility of qulified contct durtion w i; j is given in Section II, so tht the probbility of successful D2D communiction is P D ¼ e UEpr UE 2 w i; j P oc (9) Similrly, the probbility of the typicl user obtins the required content from cellulr network is: VOLUME 5, NO. 3, JULY-SEPT

6 P C ¼ P D () ¼ e UEpr UE 2 P oc þ P nc : () w i; j C. PROPORTION OF DIFFERENT BSS To deploy the cche cpcity properly, we divide the BSs into three different kinds: The VIBS, the NSC nd the USC, nd the three kinds of BSs re distributed PPP respectively ccording to the feture of Poisson distribution. The threshold of STF r is the key prmeter to select VIBSs. As the STF follows the Gussin distribution with the men vlue m nd the vrince vlue s 2, the proportion of VIBSs is P v ¼ r m erfc p 2 s ffiffi : (2) 2 The selected VIBSs hve higher STF, which mens they hve stronger socil ties with other BSs, nd lrger similrity of trffic vrition with other BSs. Since the VIBS is more typicl nd representtive, it cn serve some BSs which is locted in the rnge of prticulr distnce, these BSs re served by the corresponding VIBS nd defined s NSCs without equipping cche, the NSCs cn only link to the corresponding VIBS for meeting the limittion of bckhul. The proportion of NSCs is P n ¼ e P vpr v 2 ð Pv Þ; (3) where r V is the distence between the typicl NSC nd the corresponding VIBS, nd the PDF of r V is f rv ðþ¼ r 2pe pr2. All the rest BSs re sorted s USCs, the USCs re with lower STF thn the threshold r nd cnnot be served by the VIBSs which should be equipped with cche nd re ble to link to the core networks vi bckhul links. The proportion of USCs is P u ¼ e P 2 vpr v ð Pv Þ: (4) D. COVERAGE PROBABILITY For the convenience of nlysis, typicl user is chosen to be locted t the origin, nd the distnces to the serving BS nd the D2D prtner re defined s r nd r 2. Besides, the prmeters of three kinds of BSs re exctly the sme. The downlink signl-to interference rtio (SIR) re SIR B ¼ P B h b ;ir P b2fb=b P B h b;i rb;i þ P j2fj b jip UE h j;i rj;i (5) P UE h j ;ir2 SIR D ¼ P b2fb P Bh b;i rb;i þ P j2fj=j b ji P UE h j;i rj;i ; (6) where P B nd P UE re the trnsmission power of BSs nd terminl devices respectively. b ji in (5) indictes the presence of interference from D2D communiction to cellulr communiction, b ji only tkes the vlue of or which presents whether there exists n interference. For b ji in (6) represents the interference from the other D2D pirs tht shre spectrum resources with user j nd user i [9]. The coverge probbility of cellulr communiction is P C cover ¼ E r ½P½SIR B > Tjr ŠŠ Z ¼ P½SIR B > Tjr Š f r ðþdr r r > Z ¼ e pr2 P½h b;i > r TI jr Š2pr dr r > P B! ¼ exp X2 2 p ;g g P g T G 2 ; T 2 (7) where p ; ¼ ; p ;2 ¼ b j;i nd ^ ¼ = ¼ ; ^ 2 ¼ UE =. T is the SIR threshold for cellulr communiction. The gmm function is defined s G ðyþ¼ R y þx dx; y. 2 Similrly, the coverge probbility of D2D communiction is P D cover ¼ E r 2 ½P½SIR D > Tjr 2 ŠŠ! ¼ exp X2 2 p ;g g P g T G 2 ; where ^ ¼ = UE ; ^ 2 ¼ UE = UE ¼. The coverge probbility of the network is T 2 (8) P cover ¼ P C P C cover þ P D P D cover : (9) E. CACHE HITTING PROBABILITY Our proposed network is cche limited, the totl number of cched files in the network is N f nd ech file is of L length. The simultion in [9] shows tht the IBP simulted trce ccurtely fits the red dt nd pproximtely follows the distribution. Thus we dopt the Zipf distribution to model the file populrity in the network, nd the probbility of the f th rnked content requested by terminl users is p f ¼ ðd Þf d, where f > nd d > reflects the skew of the populrity distribution. The cche cpcity is shred by the VIBSs nd the NSCs eqully N f N v ¼ ðp v þ P u Þ (2) N u ¼ N f N v P v ðp v þ P u Þ : (2) When the typicl user tkes cellulr communiction nd ssocited with the VIBS, the cche hitting probbility nd missing probbility re P v h ¼ P v Z Nv P v m ¼ P v p f df ¼ Nv d Pv (22) Z Nv p f df ¼ N d v P v : (23) 396 VOLUME 5, NO. 3, JULY-SEPT. 27

7 When the typicl user tkes cellulr communiction nd ssocited with the NSC the cche hitting probbility nd missing probbility re P n h ¼ P n Z Nv P n m ¼ P n p f df ¼ Nv d Pn (24) Z Nv p f df ¼ N d v P n : (25) When the typicl user tkes cellulr communiction nd ssocited with the USC, the cche hitting probbility nd missing probbility re P u h ¼ P u Z Nu P u m ¼ P u p f df ¼ Nu d Pu (26) Z Nu p f df ¼ N d u P u : (27) F. NETWORK THROUGHPUT AND OFFLOADED TRAFFIC When the typicl user n tkes D2D communiction, the dt rte is given by R D ¼ W log 2 ð þ SIR D Þ; (28) nd the throughput of the D2D communiction comes to T D ¼ P D R D : (29) When the user tkes cellulr communiction, the dt rte depends on which kind of BSs the user ssocited with. If the user n ssocites with the VIBS nd hits the cche, the dt rte will be R B h ¼ W log 2ð þ SIR B Þ; (3) where W is the bndwidth of the network, nd if the user did not get the required content from the locl cche in VIBS, the dt rte will be limited by the bckhul, which is R B m ¼ W log 2ð þ SIR B Þ if W log 2 ð þ SIR B Þ C bh C bh if W log 2 ð þ SIR B Þ C bh ; (3) where C bh ¼ C bh2 ¼ P v þ P n m þ ; (32) Pu m C is the totl bckhul limittion of core network. The fronthul is shred with both the NSCs nd the USCs C 2 C fh ¼ min ; C 2 ; (33) ð P v Þ=P v where C 2 is the fronthul limittion of VIBS, nd C fh > C bh. C The throughput will be given s T V ¼ P V h R V h þ PV m R V m ; (34) where the verge dt rte is given below fter derived nd simplified Z!! R V h ¼ þ h þ X2 g p ;g P 2 g hg 2 dh Z C e bh R V m ¼ þ h þ X2 h 2 g p ;g P 2 g hg 2 h 2 (35)!! dh: (36) Similrly, the throughput when the nth user ssocite with the NSC nd USC re given Z C e fh T n ¼ P n h þ h þ X2 Z C e bh þ P n m þ h þ X2 Z T u ¼ P u h þ h þ X2 g p ;g P 2 g hg 2 Z C e bh þ P u m þ h þ X2 h 2 g p ;g P 2 g hg 2!! dh h 2!! g p ;g P 2 g hg 2 dh h 2 g p ;g P 2 g hg 2 h 2!! dh (37)!! dh: (38) When the typicl user n tkes cellulr communiction, the throughput is denoted s T c ¼ P V T V þ P n T n þp u T u : (39) In our proposed network, different rnked users hs different probbilities of selecting the contents, so the totl network throughput is described s T ¼ XN n¼ P oc ðp D T D þ P C T c ÞþP nc T c : (4) The trffic offloded by D2D communiction lyer is denoted s T ol ¼ T P nc T c ; (4) where T represents the network throughput when there is no D2D lyer, nd is derived s T ¼ P v h R v olh þ Pn h R n olh þ Pu h R u olh þ P v m R v olm þ Pn m R n olm þ Pu m R u olm: (42) VOLUME 5, NO. 3, JULY-SEPT

8 TABLE. Importnt vribles. Vrible v i;j C C 2 N f r r V r UE Mening Probbility of qulified contct durtion Totl bckhul limittion of core network Fronthul limittion of VIBS Totl number of cched files in the network Threshold of socil tie fctor VIBS cover rdius Mximum vlue of D2D communiction rdius The verge dt rte bove re simplified s Z!! R v olh ¼ R n olh ¼ þ h þ h2 G dh h 2 Z e C fh!! R u olh ¼ þ h þ h2 G dh h 2 R V olm ¼ R u olm ¼ R n olm Z C e bh!! ¼ þ h þ p ;hg 2 dh: h 2 (43) G. POWER CONSUMPTION AND ENERGY EFFICIENCY The network power consumption is constructed of two prts, the D2D communiction prt nd the cellulr communiction prt. As the throughput is derived on the view of users, We will derive the network power consumption from users ngle, nd is derived s Pow ¼ XN ðp D Pow D2D þ ð P D ÞPow c Þ; (44) n¼ where Pow D2D is the trnsmission power of terminl devices, nd the Pow c is the power consumption which contins three kinds of BSs power consumption, the bsic BS power consumption, cching power consumption nd bckhul power consumption. The expression of the EE of the network is EE ¼ T Pow ; (45) which is the rtio of totl network throughput to the network power consumption. TABLE 2. Simultion prmeters. Prmeter Vlue Prmeter Vlue Bsic Prmeters 4 :=m 2 P B 38 dbm Cche Relted W MHz d.2 C 3 Mbps C 2 5 Mbps L 24 Mbits N f D2D Relted P D 23 dbm X min 2 s " 2 k u UE :5; ; 2=m 2 Dely Relted t hd : ms t bh :5 ms t 2hd :5 ms t 2bh 3:5 ms t fh :5 ms which is the product of the delys of different conditions nd the corresponding probbilities. t hd nd t bh re the cche retrieve nd bckhul dely for VIBSs, t 2hd, t 2bh nd t fh re cche retrieve, bckhul nd fronthul dely for SCs, nd t 2bh > t bh > t fh > t 2hd > t hd. As the successful D2D prtners re ll locl users within the edge of the network nd re closed with ech other, the dely of D2D communiction t D is short enough to neglect. Thus, the verge network dely minly depends on the dely of cellulr network communiction. IV. NUMERICAL RESULTS In this section, we evlute the performnce of the cching strtegy for two-lyer socil cyberspces. Relted simultion prmeters re given in Tble 2. The users encounter durtion relted prmeters re tken from [9], the power relted prmeters re from [2], [2]. The coverge probbility of the network is shown in Figure 4. The different colors of lines denote the vrition in the densities of terminl users nd BSs. For the convenience of comprison, only the density of users is chnged, s the coverge probbility is relted to the rtio of the density of users to the density of BSs. As is shown in Figure 4, the H. NETWORK AVERAGE DELAY As n importnt KPI, the network dely hs huge influence on user s Qulity of Experience (QoE). In our pper, the verge network dely is denoted s t ¼ P c ðp V h t hd þ P V m 2t bh fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} get ccess to VIBSs þ P n h t hd þ 2t fh þ P n m 2t fh þ 2t bh fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} get ccess to NSCs þ P u h t 2hd þ P u m 2t 2bhÞþ fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} P D t fflfflffl{zfflfflffl} D ; get ccess to USCs D2D communiction (46) FIGURE 4. Coverge probbility. 398 VOLUME 5, NO. 3, JULY-SEPT. 27

9 FIGURE 5. Network throughput: () The impct of VIBS cover rdius on throughput; (b) The impct of D2D communiction rdius on throughput; (c) The impct of socil-tie fctor (STF) threshold on throughput. coverge probbility goes down when the SIR threshold becomes higher, becuse more requirements could not meet the SIR threshold nd re filed to sent to the BSs or to the D2D prtners, nd the coverge probbility goes down quickly t first, when the threshold becomes lrger nd the slope turns smll. The lines with color red, green nd blue represent the rtios of users density to the BSs density re 5, nd 2 respectively. When the rtio is smller, the coverge probbility is much higher becuse of less interference mong users, nd ech BS serves fewer users with fewer requirements t the sme time, so tht the chnnel qulity is ssured. When the threshold is higher, the coverge probbility becomes lower when the rtio is smller, becuse under the high STF threshold, D2D communiction is esier tken thn the cellulr communiction s the physicl distnce between two users is short nd hs less interference compred with cellulr communiction. The numericl results of network throughput re shown in Figure 5. In Figure 5(), The impct of VIBS cover rdius on network throughput is reflected. The threshold of STF tkes the vlues of, m,.4 nd.6. The D2D communiction rdius tkes the vlue of 2. As is shown in Figure 5(), the network throughput decreses when the rdius of VIBSs cover rnge increses. The lrger the rdius is, VIBSs cn serve more BSs, nd more smll cells turn into NSCs without cching nd bckhul cpcity. Thus, the throughput goes down. When the rdius rrives t some specific vlue, network throughput will not chnge even the rdius chnges, becuse the totl number of BSs is limited, when the rdius is lrger thn specific vlue, ll the other smll cells cn be served by the VIBSs, nd there is no USC exists. The proportion of VIBSs nd the NSCs is fixed vlue, nd the throughput is no longer chnging. When the STF threshold r increses, the network throughput decreses first nd increses lter. To further study the impct of r on the network throughput, we simulted the throughput with different vlues of STF threshold nd the result is shown in Figure 5(c), nd we will explin the vrition lter. Another obvious feture of Figure 5() is tht the four different lines hve crossings with others. The reson for this phenomenon is tht the proportion of three different kinds of BSs is decided by both the STF threshold r nd the VIBSs cover rdius, so tht the throughput of the network with lower vlue of r nd lrger VIBSs cover rdius my be equl with the network with higher vlue of r nd smller VIBSs cover rdius. By optimizing the STF threshold nd VIBS cover rdius jointly, the network throughput cn chieve 44 percent gin thn the lowest throughput. Figure 5(b) shows the impct of D2D communiction rdius on the network throughput. The threshold of STF tkes the vlues of, m,.4 nd.6. The VIBSs cover rdius tkes the vlue of. As is shown in the figure, the network throughput cn chieve 6 percent gin by selecting the D2D communiction rdius properly. When the rdius of mximum successful D2D communiction increses, the network throughput increses t first nd cn rech the mximum of network throughput, when the rdius is lrger thn 35 meter, the throughput is no longer chnging, s when the rdius increses, the probbility of users tking D2D communiction increses, but the interference from other D2D pirs will increse nd will result in the filure of trnsmitting contents vi D2D links. Besides, the requesting user will find best D2D prtner in the vilble physicl rnge, so the throughput will be stble becuse the probbility of successful D2D communiction for the typicl user is stble. When the STF threshold r increses, the network throughput decreses first nd increses lter, which reflects the exctly tend with the result in Figure 5(c). To exploit the impct of STF threshold r on the network throughput, We tke some vlues of rdius of VIBSs coverge nd D2D communiction distnce in pirs. As is shown in Figure 5(c), the influence of r is much lrger thn tht of both rdius, becuse the proportion of VIBSs is minly up to the vlue of r. Since the VIBS is equipped with cche nd bckhul, nd serves other BSs s well, it contributes to the network throughput most, so tht the network throughput decreses t first when the number of VIBSs becomes less. As the USCs shre limited cching nd bckhul cpcity, its number lso hs n obvious influence on the network throughput. When the threshold of STF becomes higher, the proportion of USCs increses nd results in the increse of the network throughput. When the difference of the number of VIBSs nd USCs chieves the mximum, the network throughput comes to the minimum s is shown in Figure 5(c). VOLUME 5, NO. 3, JULY-SEPT

10 FIGURE 6. Power consumption: () The impct of VIBS cover rdius on power consumption; (b) The impct of D2D communiction rdius on power consumption. The simultion results show tht the tendency of network EE hs similr properties with the network throughput for the sme reson. The proportion of three BSs nd the probbility tht whether the users tke D2D communiction or the cellulr communiction to obtin the required contents re the key elements of influencing the network EE. Figure 6 shows the power consumption results with the chnges of the VIBS cover rdius nd D2D communiction rdius. In Figure 6(), the mcro tend of the power consumption is incresing s the VIBS cover rdius increses nd the STF threshold r becomes higher. As the VIBS cover rdius increses, VIBSs will serve more BSs, which results in the power consumption of bckhul nd cche hitting rises. The USCs lso contribute lot to the power consumption s they re equipped with cche cpcity nd cn directly link to the core network vi bckhul. The higher the threshold r is, the lrger number of USCs is in this network, nd results in the rise of the network power consumption. As the number of USCs is up to both r nd the VIBSs cover rdius, the four lines hve cross spots s expected. As is shown in Figure 6(b), the results of network power consumption with different STF threshold hve little difference. After mgnified the figure, we cn see tht the four lines re not totlly overlpped, which mens the STF threshold hs less impct on power consumptions when compring with the impct of D2D communiction rdius. As the cellulr communiction consumes much more powers thn D2D FIGURE 7. Network verge dely: () The impct of VIBS cover rdius on verge dely; (b) The impct of D2D communiction rdius on verge dely. communiction, the probbility of users tking D2D communiction influences the power consumption lot. When the D2D communiction rdius increses, more users tke D2D communiction nd the totl power consumption goes down. Figure 7 shows the impct of VIBS cover rdius nd D2D communiction rdius on network verge dely. The threshold of STF tkes the vlues of ; m; :4 nd.6. The D2D communiction rdius is 2 meter in Figure 7() nd the VIBS cover rdius is meter in Figure 7(b). The results in Figure 7() show tht when VIBS cover rdius or r decreses, the network verge dely decreses. When user ccesses to VIBS to get the required contents, the dely is smller thn the user ccesses to the NSC, nd the dely of user ccesses to the USC is the lrgest. Thus, the proportion of VIBS nd NSC increse, nd the probbility of user ccesses to VIBS or NSC increses, which results in the decrese of network verge dely. In Figure 7(b), when the D2D communiction rdius increses, the network verge dely decreses. In the proposed network, we ssumed the dely of D2D communiction is neglected. When the D2D communiction rdius increses, the probbility of user tking D2D communiction increses, nd the network verge dely decreses. Figure 8 shows the result of the trffic offloded by D2D communiction with different STF thresholds. We dopt the difference of the throughput of ll the users in our network nd the network without D2D underly to indicte the offloded trffic mount. The impcts of the proportion of three 4 VOLUME 5, NO. 3, JULY-SEPT. 27

11 6635, by the New Str in Science nd Technology of Beijing Municipl Science & Technology Commission (Beijing Nov Progrm: Z53577). FIGURE 8. The trffic offloded by D2D. kinds of BSs re the key fctors which determine the mount of offloded trffic. When r rises, or the VIBS cover rdius decreses, the mount of offloded trffic increses. The numericl results of coverge probbility, network throughput nd power consumption hve shown tht our cching strtegy performs well by tking the socil network of both bse sttions nd terminl users into considertion, nd is promising wy to cope with the exponentilly incresing trffic in the future network. V. CONCLUSION In this pper, we proposed n informtion cching strtegy for two-lyer socil cyberspce by utilizing the socil chrcteristics of BSs lyer nd users lyer. On the bsis of our previous study on BSSN, we proposed new BS selecting policy nd cching strtegy by utilizing the socil reltionship nd physicl loction of the BSs. In our proposed network, D2D communiction is tken s the underly to the cellulr network to improve the network performnce. Indin Buffet Model hs been utilized to model the socil influence mong users. The contents cched in the BSs cn be dynmiclly replced by the ltely populr contents predicted by tking the interest nd behvior of cyber users into considertion. Besides, the trffic lod on the cellulr network cn be efficiently reduced by offloding it to the D2D lyer. The network coverge probbility, the cche hitting probbility, network throughput, power consumption, EE, verge dely nd the trffic offloded by D2D re derived s key performnce indictors. Theoreticl nd numericl results show tht the proposed network chieves high coverge probbility, nd the network throughput cn chieve the gin of 44 nd 6 percent by optimizing the VIBS cover rdius nd the D2D communiction rdius respectively. Menwhile, the results lso show tht lrge mount of trffic cn be offloded by D2D lyer in our two-lyer socil cyberspce. Therefore, the informtion cching strtegy will be promising pproch to coping with the exponentilly incresing trffic dt in the future network. ACKNOWLEDGMENTS This work is supported by the Ntionl Science Foundtion of Chin (NSFC) under grnt 65754, nd REFERENCES [] J. G. Andrews, et l. Wht will 5G be?, IEEE J. Sel. Ares Commun., vol. 32, no. 6, pp , Jun. 24. [2] X. Ge, S. Tu, G. Mo, C. X. Wng, nd T. Hn, 5G ultr-dense cellulr networks, IEEE Wireless Commun., vol. 23, no., pp , Feb. 26. [3] J. Zhng, X. Zhng, nd W. Wng, Cche-enbled softwre defined heterogeneous networks for green nd flexible 5G networks, IEEE Access, vol. 4, pp , Jul. 26. [4] B. Perbthini, E. Bstug, M. Kountouris, M. Debbh, nd A. Conte, Cching t the edge: A green perspective for 5G networks, in Proc. IEEE Int. Conf. Commun. Workshop, 25, pp [5] S. Wng, X. Zhng, Y. Zhng, L. Wng, J. Yng, nd W. Wng, A survey on mobile edge networks: Convergence of computing, cching nd communictions, IEEE Access, vol. PP, no. 99, pp., 27. [6] D. M. Boyd nd N. B. Ellison, Socil network sites: Definition, history, nd scholrship, IEEE Eng. Mnge. Rev., vol. 38, no. 3, pp. 6 3, Aug. 2. [7] S. Wng, X. Zhng, J. Zhng, J. Feng, W. Wng, nd K. Xin, An pproch for sptil-temporl trffic modeling in mobile cellulr networks, in Proc. 27th Int. Teletrffic Congr., pp , 25. [8] C. H. Yu, K. Doppler, C. B. Ribeiro, nd O. Tirkkonen, Resource shring optimiztion for device-to-device communiction underlying cellulr networks, IEEE Trns. Wireless Commun., vol., no. 8, pp , Aug. 2. [9] F. Benevenuto, T. Rodrigues, M. Ch, nd V. Almeid, Chrcterizing user behvior in online socil networks, in Proc. 9th ACM SIGCOMM Conf. Internet Mes. Conf., 29, pp [] M. Motni, V. Srinivsn, nd P. S. Nuggehlli, Peoplenet: Engineering wireless virtul socil network, in Proc. th Annu. Int. Conf. Mobile Comput. Netw., 25, pp [] Y. Zhng, E. Pn, L. Song, W. Sd, Z. Dwy, nd Z. Hn, Socil network wre device-to-device communiction in wireless networks, IEEE Trns. Wireless Commun., vol. 4, no., pp. 77 9, Jn. 25. [2] F. Nzir, J. M, nd A. Senevirtne, Time criticl content delivery using predictble ptterns in mobile socil networks, in Proc. Int. Conf. Comput. Sci. Eng., 29, pp [3] M. Ch, H. Kwk, P. Rodriguez, Y. Y. Ahn, nd S. Moon, I tube, you tube, everybody tubes: Anlyzing the worlds lrgest user generted content video system, in Proc. 7th ACM SIGCOMM Conf. Internet Mes., 27, pp. 4. [4] B. Bi, L. Wng, Z. Hn, W. Chen, nd T. Svensson, Cching bsed socilly-wre D2D communictions in wireless content delivery networks: A hypergrph frmework, IEEE Wireless Commun., vol. 23, no. 4, pp. 74 8, Aug. 26. [5] M. M. Alwkeel nd V. A. Alo, A teletrffic performnce study of mobile LEO-Stellite cellulr networks with gmm distributed cll durtion, IEEE Trns. Veh. Technol., vol. 55, no. 2, pp , Mr. 26. [6] J. Guo, F. Liu, nd Z. Zhu, Estimte the cll durtion distribution prmeters in GSM system bsed on K-L divergence method, in Proc. Int. Conf. Wireless Commun. Netw. Mobile Comput., 27, pp [7] X. Zhng, et l., Socil computing for mobile big dt, Computer, vol. 49, no. 9, pp. 86 9, 26 [8] J. M, et l. Modelling socil chrcteristics of mobile rdio networks, in Proc. IEEE Int. Conf. Commun. Workshop, 25, pp [9] J. Zhng, X. Zhng, Z. Yn, Y. Li, Y. Zhng, nd W. Wng, Socil-wre cche informtion processing for 5G ultr-dense networks, in Proc. Int. Conf. Wireless Commun. Signl Process., 26, pp. 5. [2] T. L. Giffiths nd Z. Ghhrmni, The indin buffet process: An introduction nd review, J. Mch. Lern. Res., vol. 2, pp , 2. [2] M. C. Gonzlez, C. A. Hidlgo, nd A. L. Brbsi, Understnding individul humn mobility ptterns, Nture, vol. 453, no. 796, pp , 28. [22] G. Auer, et l. How much energy is needed to run wireless network?, IEEE Wireless Commun., vol. 8, no. 5, pp. 4 49, Oct. 2. [23] Z. Zhou, M. Dong, K. Ot, J. Wu, nd T. Sto, Energy efficiency nd spectrl efficiency trdeoff in device-to-device (D2D) communictions, IEEE Wireless Commun. Lett., vol. 3, no. 5, pp , Oct. 24 VOLUME 5, NO. 3, JULY-SEPT. 27 4

12 XING ZHANG is full professor with the School of Informtion nd Communictions Engineering, Beijing University of Posts nd Telecommunictions (BUPT), Chin. His reserch interests re minly in 5G wireless communictions nd networks, green communictions, cognitive rdio nd coopertive communictions, big dt nd Internet of Things. He is the uthor/couthor of two technicl books nd more thn ppers in top journls nd interntionl conferences nd filed more thn 3 ptents. He hs served s generl co-chirs of the 3rd IEEE Interntionl Conference on Smrt Dt (SmrtDt-27), s TPC co-chir/tpc member for number of mjor interntionl conferences. He received the best pper wrds in the 9th Interntionl Conference on Communictions nd Networking in Chin (Chincom 24), the 7th Interntionl Symposium on Wireless Personl Multimedi Communictions (WPMC 24), nd the 8th IEEE Interntionl Conference on Wireless Communictions nd Signl Processing (IEEE WCSP 26). He is senior member of the IEEE nd IEEE ComSoc, Member of CCF. YONGJING LI received the BE degree in communiction engineering from Beijing University of Posts nd Telecommunictions, Beijing, Chin, in 25 nd she is currently working towrd the ME degree in the key lbortory of universl wireless communictions, School of Informtion nd Communiction Engineering. Her reserch interests include green communiction, 5G network rchitecture, nd mobile edge computing technologies. YAN ZHANG received the PhD degree in School of Electricl & Electronics Engineering, Nnyng Technologicl University, Singpore. He is fulltime full professor t University of Oslo, Norwy. He is lso chief scientist t Simul Reserch Lbortory, Norwy. He is n ssocite editor or on the editoril bord of number of well-estblished scientific interntionl journls. He is currently serving the book series editor-in-chief for the book series on Wireless Networks nd Mobile Communictions (CRC Press). He serves s orgnizing committee chirs nd technicl progrm committee for mny interntionl conferences. He hs received the 8 Best Pper Awrds. His current reserch interests include: Wireless networks leding to 5G, nd cyber-physicl systems (e.g., smrt grid, helthcre, trnsport). He is senior member of the IEEE, IEEE ComSoc, Computer Society, PES nd IEEE VTS. He is fellow of IET. JIAXIN ZHANG received the BE degree in informtion engineering from Beijing University of Posts nd Telecommunictions, Beijing, Chin, in 22 nd he is currently working towrd the PhD degree in the Key Lbortory of Universl Wireless Communictions, School of Informtion nd Communiction Engineering. His reserch interests include green communiction, 5G network rchitecture, nd mobile edge computing technologies. HAILING LI received the MAEng degree in Chinese Acdemy of Sciences, nd she is working towrd the PhD degree in University of Chinese Acdemy of sciences nd working in CNCERT/CC. Her reserch interests re in the res of network behvior nlysis, network security event monitoring nd emergency responses. SHUO WANG received the BE degree in informtion engineering from Nnjing University of Aeronutics nd Astronutics, Nnjing, Chin, in 2 nd received the ME degree in communiction nd informtion systems from Chin Acdemy of Spce Technology, Beijing, Chin, in 24. He is currently working towrd the PhD degree in the key lbortory of universl wireless communictions, School of Informtion nd Communiction Engineering, Beijing University of Posts nd Telecommunictions, Beijing, Chin. His reserch interests include 5G network technology, green communiction, mobile edge cching nd cloud computing. DANYANG WANG received the BE degree in communiction engineering from Beijing University of Posts nd Telecommunictions, Beijing, Chin, in 26 nd she is currently working towrd the ME degree in the key lbortory of universl wireless communictions, School of Informtion nd Communiction Engineering. Her reserch interests include green communiction, 5G network rchitecture, nd mobile edge computing technologies. 42 VOLUME 5, NO. 3, JULY-SEPT. 27