ebook img

Computing Resource Allocation in Three-Tier IoT Fog Networks: a Joint Optimization Approach Combining Stackelberg Game and Matching PDF

0.27 MB·
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Computing Resource Allocation in Three-Tier IoT Fog Networks: a Joint Optimization Approach Combining Stackelberg Game and Matching

1 Computing Resource Allocation in Three-Tier IoT Fog Networks: a Joint Optimization Approach Combining Stackelberg Game and Matching Huaqing Zhang∗, Yong Xiaok, Shengrong Bu†, Dusit Niyato§, Richard Yu‡, and Zhu Han∗ ∗Department of Electrical and Computer Engineering, University of Houston, Houston, TX, USA. kDepartment of Electrical and Computer Engineering at the University of Arizona, Tucson, Arizona, USA. †School of Engineering, University of Glasgow, UK. 7 §School of Computer Science and Engineering, Nanyang Technological University, Singapore. 1 0 ‡Department of Systems and Computer Engineering, Carleton University, Ottawa, ON, Canada. 2 n a J Abstract—Fog computing is a promising architecture to pro- ControlSignal 4 vide economic and low latency data services for future Internet (cid:258)(cid:258) DataStream 1 of things (IoT)-based network systems. It relies on a set of low- power fog nodes that are close to the end users to offload the DSOd1 DSOd2 DSOdm Virtualized ] services originally targeting at cloud data centers. In this paper, Network T we consider a specific fog computingnetwork consistingof a set G of data service operators (DSOs) each of which controls a set of fog nodes to provide the required data service to a set of data s. servicesubscribers(DSSs).Howtoallocatethelimitedcomputing (cid:258)(cid:258) [c resourcesoffognodes(FNs)toalltheDSSstoachieveanoptimal FNf1 FNf2 FNf3 FNfk and stable performance is an important problem. In this paper, we propose a joint optimization framework for all FNs, DSOs 1 and DSSs to achieve the optimal resource allocation schemes in v 2 a distributed fashion. In the framework, we first formulate a (cid:258)(cid:258) Physical Stackelberg game to analyze the pricing problem for the DSOs Netwotk 2 as well as the resource allocation problem for the DSSs. Under DSSs1 DSSs2 DSSs3 DSSsN 9 3 the scenarios that the DSOs can know the expected amount of Fig. 1: System architecture resourcepurchasedbytheDSSs,amany-to-manymatchinggame 0 is applied to investigate the pairing problem between DSOs and . 1 FNs. Finally, within the same DSO, we apply another layer of 0 many-to-many matching between each of the paired FNs and whichcan be intolerablefor theIoT applicationsthatrequires 7 serving DSSs to solve the FN-DSS pairing problem. Simulation real-time interaction or mobility. In order to overcome these 1 results show that our proposed framework can significantly : improve the performance of the IoT-based network systems. challenges,fogcomputingisproposedasapromisingsolution. v In fog computing, multiple low-power computing devices, Xi IndexTerms—Fogcomputing,Stackelberggame,match- commonly referred to as the fog nodes (FN), at the edge ing theory, Internet of things. r of the networks are deployed to offload the data computing a servicesfromthecloud.With the propertyofsmall-scale,low I. INTRODUCTION construction cost and mobility support, the FNs are generally With the rapid developmentof Internet of things (IoT), the deployed much closer to the DSSs and therefore can provide numberofconnecteddeviceshasincreasedataunprecedented low-latency fast-response and location-awareness service [2]. speed[1].Itisknownthatanalyzingthebigdatageneratedby Fog computing networks can consist of a large number allkindsofIoTdevicesrequiresa largeamountofcomputing of FNs deployed by different DSOs at different locations to resources.In orderto meetthe demandof the data computing provide various data services and applications to the DSSs. services, a large number of large-scale data centers has been When DSSs can choose their DSOs as well as the corre- deployed. In addition, cloud computing has been proposed sponding FNs to further enhance their quality-of-experience, recently to provide flexible and efficient services to the data how to allocate the limited computingresourcesof all the fog service subscribers (DSSs). In cloud computing, the data nodes to the DSSs is still an open problem. In this paper, service operator (DSO) is able to organize a shared pool of we further extend our previous work [3] and focus on the configurable computing resources (such as servers, storage, resource selection and allocation problem between the FNs, networks, applications and services), which can be easily DSOs and DSSs. We propose a joint optimization framework accessed by DSSs on demands. for all FNs, DSOs and DSSs in a distributed fashion. In the Generally speaking, large-scale data centers or clouds are framework,wefirstformulateaStackelberggametomodelthe generally built in remote areas which are far from the DSSs. interactionbetweenDSOsandDSSs,wheretheDSOssettheir This results in high transmission cost and service latency service price first, and the DSSs purchasethe optimalamount 2 TABLE I: List of Notations of computing resource blocks (CRBs) correspondingly. Once the prices of DSOs and the purchased resources of DSSs Symbol Definition havebeenobtained,eachDSScanknowhowmanyCRBs are M TotalnumberofDSOs N TotalnumberofDSSs required and can then try to compete for the CRBs owned by K TotalnumberofFNs the nearby FNs. Thus, we propose a many-to-manymatching Ψ ThesetofDSOs game to investigate the interaction between DSOs and FNs Υ ThesetofDSSs Ω ThesetofFNs where each DSO has a set of CRBs to offload, and each FN µ ServicerateofCRBs hasmanyvacantCRBstosell.AfteralltheDSOsdecidetheir λj Workloadarrival ratefortheDSSsj DSO-FN pairs, FNs will compete with each other to allocate ri Priceofunitvirtualized CRBsetbytheDCOoi Ls Thepreference listofDSSsoverDSOs their CRBs to the DSSs of the DSOs. We also adopt another j layer of many-to-manymatching framework to solve the FN- tj TotalcostduetothedelayofDSSsj hj Cost due to network delay from the physical ser- DSSpairingproblemwithinthesameDSO.Simulationresults viceprovidertoDSSsj show that our proposed framework can significantly improve oj Costduetoqueuingdelayattheservers the performance of the fog computing networks. qj TotalamountofCRBspurchasedbytheDSSsj lkj Distancebetween theFNfk andtheDSSsj The rest of this paper is organized as follows. We describe Wjs Utility functionofDSSsj the system modelin Section II and formulatethe problemsin Wid Utility functionofDSOdi Section III. Based on the formulatedproblem,we analyze the Wkf Utility functionofFNfk system with the proposed framework in Section IV, where τij ThebooleanvariabledeterminingwhethertheDSO di servesDSSsj ornot. the interaction between DSOs and DSSs is considered in αj,βj,γj Weightfactors intheutility functionofDSSsj Section IV-A, the interaction between FNs and DSOs is tth Themaximumtolerance ofservice delay forDSS analyzed in Section IV-B, and the interaction between FNs sj ckj Transmission cost for unit CRB from FN fk to and DSSs is discussed in Section IV-C. Finally, we evaluate DSSsj theperformanceofourworkinSectionV,showrelatedworks ei Increment of the energy cost in the massive data in Section VI and summarize the paper in Section VII. centerforDSOdi ηkfi NormalizedpreferencefromtheFNfktotheDSO di II. SYSTEMMODEL rptkh RUepnpterofbouunnitdCoRfBtotsaeltdbeylatyhecoFsNt fk Consider a fog computing network where each DSS can qkfjs AmountofCRBsallocated fromtheFNfk tothe DSSsj submit its data computing service to a set of neighboring qifkd AmountofCRBsallocated fromtheFNfk tothe FNs deployed by a set of DSOs as illustrated in Figure 1. DSOdi Accordingly, we consider a three-tier fog network, where Ldif Preference listofDSOdi onallFNs the DSOs locate in the middle layer, managing the FNs Lsjf Preference listofDSSsj onallFNs in the upper layer and serving DSSs in the bottom layer. Lfks Preference listofFNfk onallDSSs Without loss of generality, we assume there are M DSOs, labeled as Ψ = {d1,d2,...,dM} and N DSSs, denoted Assume the DSSs apply real-time interactive applications, as Υ = {s1,s2,...,sN}. Let λj be the workload arrival whereQoSismeasuredbytheservicedelay.Inthispaper,for rate of DSS s , ∀s ∈ Υ. We assume each DSS has a j j normalized preference list, denoted as Ls over all DSOs. DSS sj, ∀sj ∈ Υ, we measure the cost of the service delay j as Moreover, K FNs, labeled as Ω = {f1,f2,...,fK}, locate in the area of consideration. We define the unit amount of computing resources that can be distributed by each FN as tj =hj +oj, (1) the “computing resource block (CRB)” [3], each of which which consists of the cost caused by the queuing delay o at j can provide computing service at the rate of µ. The physical the servers as well as the cost caused by the network delay data transmission network between FNs and DSSs satisfies h from the sensors to the physicalservice providerand from j the SecondNet topology [4], where the network facilities can the physical service provider to DSS s . j providethe guaranteedquality-of-service(QoS)forthe DSSs. According to the queuing delay model in [5], [7], which Accordingly, in order to reduce the risk of potential network can be easily extended to other models, the cost of queuing congestion and achieve real-time fast-response interaction, delay when serving DSS s is j each DSO tries to offload the data services submitted by the DSSstothelarge-scaledatacenterstothelocalFNs.However, λ o = j , (2) as the DSSs cannot have the authorization to access the j µ− λj CRBsdirectly,theDSSsarerequiredtoreceivethevirtualized qj services from the DSOs, and with the management of DSOs, where q is the amount of CRBs purchased by DSS s . j j theCRBs ofthe FNscanfinallybeallocatedtotheDSSs. We Moreover, as the network delay is related to the transmission assume that different DSOs offer data services with different distance, traffic condition in the network and many other requirements. Based on the preference list Ls, the DSS s , unpredictedfactors,inrealsituations,wesupposethenetwork j j ∀s ∈ Υ is required to subscribe to at most one DSO. The delay can be evaluated from training data periodically sent j network architecture is illustrated in Fig. 1. from sensors to the physical service provider and from the 3 physical service provider to the DSS. In this paper, we set the distance between the farthest sensor to the physical serviceproviderplusthedistancebetweenthephysicalservice provider (e.g., FN f ) to DSS s is denoted as l . For k j kj simplicity, we assume the network resourcefrom the physical FNs service providerto DSS sj is sufficient, and the cost incurred Many-to-Many bythenetworkdelayh generallyfollowsalinearfunctionof Matching j the distance from the sensor to the physical service provider DSOs plusthedistancefromthephysicalserviceprovidertotheDSS s , i.e., h =θl , where θ is the scalar. j j kj As the DSSs in the network pay DSOs for the service, following the structure of [3], the utility of DSS s , ∀j ∈ Υ, j can be denoted as the revenue received from the workload Stackelberg data minus both the cost of service delay and payment to the Games DSOs as follows: DSS DSS DSS M Ws = τ (α λ −β q r −γ t ), (3) j ij j j j j i j j Many-to-Many i=1 Matching X Ajointoptimizationframework where r is the price set by DSS d for each unit of the i i virtualized CRB. αj, βj, and γj are weight factors. τij is Fig. 2: A joint optimization framework thebooleanvariabledeterminingwhetherDSO d servesDSS i s or not. If τ = 1, DSS s is served by the DSO d . j ij j i Otherwise, DSS s prefers to be served by other DSOs. The j value of τ follows the preference list Ls of DSS s , and K ij j j Wf = ηf (p −c )qfs, (5) M k ki k kj kj sinceeachDSScanatmostchooseoneDSS,i.e., τij =1, k=1 i=1 X ∀j ∈ Υ. For each DSS, assume there is an upPper bound where qkfjs is the amount of CRBs allocated from FN fk to tth for the service delay. When the service delay is larger DSS sj. than the threshold, the DSS will regard it as an unsuccessful connection. Corresponding, we set qth as the lower bound j III. PROBLEMFORMULATION of CRBs required for DSS s to guarantee the service delay j within the acceptable boundaries. Accordingtothemodeledarchitectureoffognetwork,with Based on the amount of virtualized CRB purchased by tradings between FNs and DSOs and between DSOs and servingDSSs,theutilityofeachDSOistherevenuesreceived DSSs, it is impossible to reach the maximum utilities for from the DSSs minus the payment to the facilities that are all FNs, MDCOs and DSSs simultaneously. Accordingly, we able to provide the physical CRBs. Each DSO prefers to consider a sequential decision making process. During the offload its services to the FNs nearby. However, if there are process, the DSOs first predict the total amount of CRBs not sufficient amount of CRBs from the FNs which can meet purchased from their servings DSSs and set their service the requirements of all DSSs. Some DSSs will be served by prices to their DSSs based on the prediction. Observing the the massive data centers, which are located far away from the behaviors of the DSOs, each DSS determines the optimal DSSs. We suppose the increment of the energy cost in the amount of CRBs to purchase to achieve the maximum utility. massive data center is e . Therefore, for the DSO d , if qfd Furthermore,knowingthe totalamountof requiredCRBs, the i i ik amountofCRBsareoffloadedtotheFNf ,andqo amountof DSOs try to allocate CRBs to the FNs in the neighborhood. k i CRBs are still served by the massive data centers, the utility Finally, all FNs selected by the same DSO competes for the function of DSO d , ∀i∈Ψ, can be denoted as DSSs. i We can write the optimization problem can be formulated N K as follows. Wd = τ (r q )− p qfd−e qo. (4) i ij i j k ik i i Xj=1 kX=1 mqajx Wjs(qj|r), ∀j ∈Υ, p is the price set by the FN f , which is determined by the k k M cost and current traffic of FN fk. τij =1, ForFNfk inthenetwork,theutilityisthepaymentreceived  i=1 (6) farsomthethteranDsSmOisssiomnincuosstthfeortreaancshmiusnsiiotnCRcoBs,t.wWhiechsehtasckaj s.t. qτtPjij≤≥αjt0λt,hj,≤τij(βjqjri+γjtj), linear relationship with the distance l . Moreover, we set j kj cηokfinsaisdethriengnotrhmeaplirzeefedrepnrecfeerteondceiffteorethnet DDSSOOsd,it.hAecdcioscrdoiunngtleyd, where r is thepricτiinjg∈p{ro0fi,1le},of a∀lliD∈SΨO,s observed by DSS utility function for each DSO is s . j 4 PredictingthebehaviorsoftheDSSs, eachDSO isrequired toleranceandsettingpricesofDSOs,theoptimalamount tosettheservicepriceto eachservingDSSandcompetewith of CRBs should be determined for high utilities. other DSOs to choose FNs in the neighborhood. Thus, the 2) PricingproblemfortheDSOs:Inthedataservicewith formulated problem of the DSO is fog, the DSOs are requiredto providevirtualized CRBs to the DSSs and try to rent the CRBs from the FNs to max Wd(r |q∗,p,r∗ ), ∀i∈Ψ, serve DSSs in the physical network. Therefore, how to i i −i ri do the pricing for the DSOs is a problem. Considering N K τ (r q )≥ p qfd+e qo, (7) the announcedrent from all FNs, the DSOs need to set s.t. j=1 ij i j k=1 k ik i i apricewhichcanbringprofitsforthemselves.However,  rPi ≥0, P if the price is set too high, the serving DSSs will not where q∗ denotes the optimal amount of CRBs purchased by purchasea largeamountofCRBs. Therefore,predicting allDSSs. p istheprofileofrentforallFNs. r∗ istheprofile the reactions of CRBs and observing the rent of FNs, −i the DSOs are required to determine its service price so of optimal service prices set by other DSOs. as to receive the maximum revenues. Moreover, each FN in the network would like to choose 3) DSO-FN pairing problem: As FNs may be private its preferred DSOs and serve its DSSs with low distance. computingdevices, which are small-scale and unableto Thus, competing with other FNs, it is required to determine communicatewithDSSsdirectly,theFNsareaccessible the amountof CRBs allocated to DSOs and its serving DSSs, totheDSOsonly.Inthemulti-DSOscenario,astheFNs respectively. The optimization problem is denoted as are accessible to all DSOs, it is a problem for all DSOs in the network to pair FNs distributedly so as to serve max Wkf(qifkd,qkfjs|q∗,qfi−d∗k,qf−sk∗j,p−k), ∀k ∈Ω, theirDSSswithlowlatency.Withdifferentrelationsand qfd,qfs trading history, each FN has different preference on all ik kj M DSOs. Observing the rent of all FNs, each DSO also qifkd ≤qkf−th, hasa preferencesoneach FN. Based on thepreferences s.t. jiPP=NK=11qqkfifkjsd≤≤qqkfid−−tthh,, ∀i∈Ψ, 4) oDsFwfNSiaO-tlcDl-hFSDNSitSsOppcaasui’irrrirainenngngdtrpeFpsaNruiorslibt’ns,l,geiwmtrehi:sseuArrleetfqtafeuonriryrteahDdehStipoOgahrioereirrancuFghtNiblaietiystsw.taaebbellnee k=1 DSOs and FNs, each FN has allocated its CRBs to all whereqfkdPPK=1iqskftjshe≤opqjsti−mthal,amo∀ujn∈toΥfC, RBsrenttoDSOd (f8o)r DfdoiSsrtOathnse.ceHFNobwestewtvoeeear,nllwoeciatahctehintFhoNenieraCDnRdSBOes,acittohiasDllsStDSillSiSassp.vrAaorbsiolteuhmse, i−k i with a longer transmission distance, the FNs have to all other FNs, qfs∗ is the optimal amount of CRBs allocated −kj pay more on the transmission cost. Thus, based on the to DSS s for all other FNs. During the service, the total j transmission distance, each FN has a preference over CRBs distributed to all DSOs or all DSSs cannot exceed its all DSSs. Moreover, each DSS also have preference total available CRBs qf−th for FN f . Furthermore, the total k k over FNs based on the rent. Therefore, following the CRBs purchased from DSO d or DSS s should not exceed i j preference of all FNs’ and DSSs’, a stable FN-DSS its demand qd−th or qs−th, respectively. i j pairing result should be achieved. In summary, following the relationships among all FNs, According to the formulated problems, all FNs, DSOs and DSOs and DSSs, we focus on the following problems: DSSs are rational and autonomous individuals, who observe 1) Resource purchasing problem for the DSSs: In the the behaviors of others and make decisions to improve their network, as the DSSs can only access to the DSOs in own utilities. Therefore, in order to reach the optimal and a virtualized fashion, they are required to purchase the stable solutions for all FNs, DSOs and DSSs, we model a optimalamountofCRBs fromtheDSOs. Followingthe three-stage joint optimization framework, as shown in Fig. systemmodel,asdifferentDSSshavedifferenttolerance 2. In the framework, we first model the Stackelberg games of servicedelay,whenthe upperboundof servicedelay to solve the pricing problem for the DSOs and resource is high, the DSSs are able to purchase a small amount purchasing problem for the DSSs. When the DSOs know the of CRBs to achieve satisfying services. However, when expectedamountofresourcepurchasedbythe DSSs, a many- the upper bound of service delay is low, the DSSs have to-manymatchingisproposedbetweentheDSOsandtheFNs to purchase a large amount of CRBs to guarantee the to deal with the DSO-FN pairing problem. Finally, within service delay is within the tolerated region. Moreover, the same DSO, we apply another many-to-many matching the service priceset by the DSOs also affectsthe utility between its paired FNs and serving the DSSs to solve the of DSSs. When the price is in high value, even though FN-DSS pairing problem. thelargeamountofCRBs isabletoimprovethequality ofdataservices,theDSSshavetomakealargepayment IV. SYSTEMANALYSIS to the DSOs for their services. The revenue may not In this section, we analyze the optimal strategies for FNs, be satisfying. Therefore, considering both the delay DSOs and DSSs. Based on the analysis of the formulated 5 framework, in the following sub-sections, we first investigate the interactions between the DSOs and DSSs to determine N K λ β λ β how many CRBs are required during the service. Given the Wd = τ ( j jr + j jr )− p qfd−e qo. optimized behaviors of the DSOs and DSSs, we analyze the i j=1 ij µ sγj i µ sγj i k=1 k ik i i X X interactions between the FNs and DSOs based on different f (14) preferences. Finally, with the obtained results, we discuss the In the formulated problem (13), we take the first derivative interactionsbetween the FNs andDSSs within the same DSO of Wid with respective to ri and discover it is a monotonous for better services. increasing function with respective to ri. Furthermore, as the servfice delay cannot surpass tth for the DSSs. The CRB purchased by DSSs has the following low bound A. The Interaction between DSOs and DSSs λ t In the virtualized network, the DSOs provides CRBs for q ≥ j th . (15) j the DSSs. Following the formulated problems for both DSOs µtth−λj andDSSs,thereisaStackelberggame,wheretheDSOsactas Thus, followingthe relation in (9), the maximumand optimal leadersandDSSsactasfollowers.Inthegame,whenallDSSs price for the DSO d to DSS s is, ∀i∈Ψ, ∀j ∈Υ, i j choose their serving DSOs with their preferences, the DSO 2 first sets the service price. Then, based on the price all DSSs γ µt −λ r = j th j . (16) determineoptimalamountof CRBs to purchase.Accordingly, i β λ j(cid:18) j (cid:19) considering the optimization problem of DSSs, we have the following lemma. B. The Interaction between FNs and DSOs Lemma 1. InthemodeledStackelberggamebetweenDSOd Accordingto the predictedamountofCRBs orderedbythe i andDSSs ,whentheDSOannouncesitsservicepricer ,the DSSs, the DSOs try to offload the services from the massive j i optimal amount of CRBs q purchased by the DSS is data centers to the DSSs nearby.Observing the service prices j set by all FNs, DSO d , ∀i ∈ Ψ, has a preference list Ldf = i i λ λ qj∗ = µ rjiβγjj + µj. (9) phLredif1fe,r.t.o.,cLhodiKfosieotnhealFlNFNs swiinththaelonweigphrbicoer,howoed.seAtstheDSOs q Proof: Accordingto the utility functionof DSS sj 3, the Ldf =−p , ‘∀i∈Ψ, ∀k ∈Ω. (17) second derivative of Ws with respect to q is ik k j j Furthermore, each FN also has different preferences over ∂2Wjs =− 2λ2jµ . (10) DSOs. Thus, we set the preference list of FN fk, ∀k ∈ Ω, ∂qj2 (µqj −λj)3 on all DSOs as Lfkd = Lf1kd,...,LdMfk , satisfying, As ∂2Wjs < 0, Ws is a quasi-concave function with respect Lfd =ηf , h ‘∀i∈Ψi, ∀k ∈Ω. (18) ∂q2 j ik ki to q . Fjurthermore, the first derivative of Ws with respect to j j Consideringthepreferencelistsof FNsandDSOs, i.e., Ldf q is ik j and Lfd, respectively, we design a many-to-many matching ik ∂Ws λ 2 β algorithm for the DSO-FN pairing problem. As shown in j = j −r j. (11) Algorithm 1, after the preference lists are constructed, we set i ∂q µb −λ γ j (cid:18) j j(cid:19) j apointerforeachFNinitspreferencelist.Initially,wesetthe We setthefirstderivativeequaltozeroandobtaintheoptimal pointer at the first DSO in the list. During the each round of amount of CRBs to purchase so as to achieve the maximum matching, each FN first propose to the DSO in the pointer of utility, i.e, the preferencelist. Based on behaviorsof the FNs, each DSO chooses its most preferred FNs in its preference list until the λ λ q∗ = j + j. (12) requiredCRBs are all supplied by the FNs. If the FNs supply j µ r βj µ more CRBs than the DSO requires, the DSO will reject the iγj CRBs from less favourite FNs. If the FNs supplies less CRBs q than the DSO requires, the DSO will choose massive data Therefore,consideringthereactionsof theDSSs, we adjust centers for the rest of the services. At the end of each round, the optimization problemfor the DSO di, ∀i∈Ψ, as follows. if all of the CRBs from the FN have been allocated to the DSOs,thepointersoftheFNwillkeepunchanged.Otherwise, max Wd(r |q∗,p,r∗ ), ∀i∈Ψ, the pointers of the rejected FNs will move to the next DSO i i −i ri in the preference list. In the next round, the FNs which still N K haveavailableCRBs willproposeto thenewDSOs according f τ (λj βjr + λj βjr )≥ p qfd+e qo, s.t. rjP=i1≥i0j, µ qγj i µqγj i kP=1 k ik i i tboyththeeirDpSoiOnteinrs.thSepelcaisfitcraolluyn,di,fbthuet CdiRscBasrdoefdFiNn tahree ccuhrorseennt (13) round, we suppose the pointer of the FN doesn’t change its where position, considering the pointed DSO in the current round 6 Algorithm 1 Many-to-Many Matching Algorithm for DSO- whichismorepreferredthanFNf .Inthefuturerounds,when k FN pairing problem. there are new FNs proposing to DSO d , if the DSO would i like to change its current accepted FNs, the DSO can only 1: for FN fk do choose the FNs that is evenbetter than the FNs in the current 2: ConstructthepreferencelistLfdonallDSOsaccording accepted list. Therefore, for FN f which has been rejected k k to (18); once from DSO d , it is impossible for it to be accepted by i 3: One pointer is set as the indicator pointing at the first the same DSO in the future rounds. DSO in the preference list. Lemma 3. Following the Algorithm 2, the DSO-FN pairing 4: end for problem will ultimately converge and achieve a stable match- 5: for DSO di do 6: Construct the preference list Ldf on all FNs according ing result. i to (17); Proof:AsprovedintheLemma2,thepointeroftheFNs 7: end for can only move in one direction. Therefore, in the perspective 8: We set flagk, ∀k ∈ Ω, as the indicator to show if the of the FNs, when the pointer of each FN has moved to the CRBsofFNfk werechosenbytheDSOinthelastround, end of the preference list, the FN has distributed its available but discarded in the current round. Initially, flagk =1; CRBstotheDSOsinanoptimizedway.Inotherwords,theFN 9: while the pointers of all FNs have not scanned all the is unable to change its pairing results unilaterally for higher DSOs in their preference list do utilities. Furthermore, in the perspective of the DSOs, when 10: FNs propose to DSOs with their service price; thepointersofallFNshavemovedtotheendofthepreference 11: forFN fk whostill haveavailableFNsto purchasedo lists, each DSO hasevaluated all proposalsfromthe available 12: if flagk =1 then FNs. Therefore, it also cannot unilaterally change its pairing 13: The pointer keep current position in the list; results, get accepted by the FNs and achieve a higher utility 14: else for itself. Furthermore, according to [6], when every agents 15: The pointer moves to the next position in the list; preference list is substitutable, a pairwise stable matching 16: end if always exists. Based on the above, the DSO-FN pairing will 17: The FN proposes to pointed DSO in its preference ultimately converge and achieve a stable matching result in list with its available FNs; Algorithm 2. 18: We set flagk =0; 19: end for C. The Interaction between FNs and DSSs 20: DSOs determines which FNs to choose; 21: for DSO di do When the CRBs from FNs have been rent to all DSOs, 22: if The total available amount of CRBs proposed by within each DSO, how to allocate the CRBs to all DSSs still the FNs exceed the requirements then remainsa problem.Accordingto the rent of all FNs, DSS sj, 23: TheDSOdi choosesthemostpreferredamountof ∀j ∈ Υ, has a preference list Lsjf = Lsj1f,...,LsjKf on all CRBs from FNs, and rejects the rest; FNs in the neighborhood,satisfying, h i 24: For CRBs of the FN fk which is chosen by the DSO in the last round, but rejected in the current Lsf =−p , ∀j ∈Υ, ∀k ∈Ω. (19) kj k round, we set flag =1; k 25: end if Furthermore, according to the utility function of the FN, 26: end for the distance between the FN and its serving DSS affect the 27: end while revenues the FN received. With a longer distance, the FN has to pay more for the data transmission in the network. Therefore,we set a preferencelist Lfs = Lsf,...,Lsf for i j1 jK may need more CRBs from the FN. The matching repeats in FN fk, ∀k ∈Ω, over all DSSs, i.e., h i circulations until all the pointers of the FNs have moved out their preference list. According to the algorithm, we have the Lfs =−l , ∀k ∈Ω, ∀j ∈Υ. (20) kj kj following lemmas. Consideringthe preferencelists of DSSs andFNs, i.e., Lsf Lemma2. ForeachFNinthematchingalgorithm,thepointer kj and Lfs, respectively, we design a many-to-many matching oftheFNinitspreferencelistmovesinonedirection.Inother kj algorithm for the FN-DSS pairing problem within DSO d , words, when its pointer has moved to the next DSOs in the i ∀i∈Ψ.AsshowninAlgorithm2,afterthepreferencelistsare preference list, whatever the matching results of other FNs, constructed,wesetapointerforeachDSSinitspreferencelist. the FN cannot achieve a higher utility by moving the pointer Initially, we set the pointer at the first FN in the list. During back. the each round of matching, each DSS first proposes to the Proof: As shown in Algorithm 2, when the DSOs deter- FN in the pointer of the preference list. Based on behaviors mines which FNs to choose, they choose the CRBs from the of the DSSs, each FN chooses its most preferred DSSs in its mostpreferredamount.IfsomeCRBsfromFNf isdiscarded preferencelistuntilthe maximumCRBs availablein theDSO k by DSO d , the current accepted CRBs belong to the FNs d are reached. If the DSSs request more CRBs than the FN i i 7 can supply, the FN will reject the less favourite DSSs. At the Algorithm2Many-to-ManyMatchingAlgorithmforFN-DSS end of each round, if the DSSs have been allocated all of its pairing problem. requested CRBs from the FNs, the pointers of the DSS will keep unchanged. Otherwise, the pointers of the rejected DSS 1: for DSS sj do will move to the next FN in the preference list. In the next 2: ConstructthepreferencelistLsf onallDSOsaccording j round,the DSSs which requireCRBs will proposeto the new to (19); FNsaccordingtotheirpointers.Specifically,ifsomeCRBs of 3: One pointer is set as the indicator pointing at the first FN are allocated to the DSS in the last round, but changed FN in the preference list. to other DSSs in the current round, we suppose the pointer 4: end for of the DSS in the next round doesn’t change its position, 5: for FN fk do considering the pointed FN in the current round may be able 6: Construct the preference list Lfs on all FNs according k to supply more CRBs to the DSS. The matching repeats in to (20); circulations until all the pointers of the DSSs have moved 7: end for out their preference list. Following Lemma 2 and Lemma 3 8: We set flagj, ∀j ∈ Υ, as the indicator to show if the in a similar way, the FN-DSS pairing problem can ultimately FNs allocate CRBs to the DSS sj in the last round, but achieve a stable matching result. adjusted in the current round. Initially, flagj =1; 9: while the pointers of all DSSs have not scanned all the V. SIMULATION RESULTS ANDDISCUSSIONS FNs in their preference list do In this section, we present simulation results to evaluate 10: DSSs propose to FNs for their services; ourproposedframeworkwithMATLAB.InthesimulatedIoT 11: forDSSsj whichhasnotbeenallocatedrequiredCRBs scenario, without specific explanation, there are 120 DSSs, 4 do DSOs and 20 FNs allocated randomlyin a circle district with 12: if flagj =1 then adiameterof10kilometers.AswefocusontheIoTscenarios 13: The pointer keep current position in the list; where DSSs are closely located with its sensors, without loss 14: else of generality for the methods in this paper, we suppose the 15: The pointer moves to the next position in the list; sensorsof each DSS are locatedat the same positionwith the 16: end if DSS. We follow the settings in [7] and set the service rate of 17: The DSS proposes to pointed FN in its preference each computingresourceblockis 0.1 (ms)−1. For each DSS, list for its data services; the workload arriving rate is a random number averaged 0.5 18: We set flagj =0; (ms)−1.Thedatatransmissionspeedis50km/ms.Thedelay 19: end for tolerance of all DSSs is set to 60 ms. Furthermore, for each 20: Each FN determines which DSSs to choose; FN, we set its preference to each DSO as a random number 21: for FN fk do satisfyingtheuniformdistributionbetween[0,1].Basedonthe 22: if The total available CRBs requested by the DSSs usage of its computing resources and its service cost, we set exceed the available volume then theannouncedrentasarandomnumbersatisfyingtheuniform 23: TheFNfk allocatetheCRBstothemostpreferred distributionbetween(0,10),andtheamountofavailableCRB DSSs, and rejects the rest; as a random number satisfying uniform distribution between 24: ForCRBsallocatedtotheDSSsj inthelastround, (0,100). The weight factors α, β and γ are set as 50, 0.01 but adjusted in the current round, we set flagj = and 0.001, respectively. 1; As shownin Fig.3, we evaluatethe utility ofall FNswhen 25: end if the number of DSSs increases. When the number of FNs is 26: end for fixed, we discover with the number of DSSs increasing, the 27: end while utilityofallFNsgenerallyincreases,andtheincreasingspeed first increases then gradually decreases to zero. The reason is that when the number of DSSs increases, but the number of the performanceof theDSSs in ourproposedframeworkwith FNs is fixed, the FNs will be able to serve more favourable the performance of the DSSs which is served by the massive DSSs with a low transmission distance. However,when all of data centers only. With the same amount of workload, we the available CRBs of FNs are allocated to the DSSs nearby discover that when the number of DSSs increases, the utility and the number of DSS keep increasing, the DSS will be of DSS genrally increases, and the utility with FNs achieve a allocated with CRBs fromthe massive data centers. Thus, the highervaluethantheonewithoutFNs.Furthermore,whenthe total utility of the FNs stop increasing. Nevertheless, when number of DSSs is fixed and the average workload for each we increase the number of FNs, we discover that with more DSS increases, the DSSs are able to receive more revenues FNs, the utility can converge to a higher bound ultimately. from the services. Thus, the utility of DSSs increases. Furthermore, because of the competition between FNs, when In Fig. 5, we analyze the relation between the utility of the number of DSS is small, with more FNs, the increasing FNs and average workload arrive rate for DSSs. As shown in speed is smaller. the figure, when the number of FNs is fixed and the average In Fig. 4, we consider the utility of all DSSs with the valueofworkloadλincreases,theutilityofFNsfirstincreases number of DSSs increasing. In the simulation, we compare then gradually converge to a fixed value. The reason is that 8 6000 6000 Numer of FNs = 15 Number of FNs = 30 Numer of FNs = 20 5500 Number of FNs = 25 5000 Numer of FNs = 25 Number of FNs = 20 Numer of FNs = 30 5000 4000 Ns Ns4500 F F Utility of 3000 Utility of 34500000 2000 3000 1000 2500 0 2000 40 60 80 100 120 140 160 180 200 220 240 1 1.5 2 2.5 3 3.5 4 4.5 5 Number of DSSs Average Value of λ (ms)-1 Fig. 3: The utility of all FNs versus the number of DSSs Fig. 5: The utility of all FNs versus the average value of λ ×106 18000 14 Small duty, without FNs Number of DSSs =240 16000 SMmedailul mdu dtyu, twy,i twh itFhNosut FNs 12 NNuummbbeerr ooff DDSSSSss ==118200 14000 Medium duty, with FNs Heavy duty, without FNs 10 Heavy duty, with FNs 12000 Utility of DSSs108000000 Utility of DSOs 68 6000 4 4000 2 2000 0 0 40 60 80 100 120 140 160 180 200 220 240 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 Number of DSSs µ (ms)-1 Fig. 4: The utility of all DSSs versus the number of DSSs Fig. 6: The utility of all DCOs versus the value of µ delay,the utility ofDSS generallydecreaseswith the valueof when the workload of all DSSs increases, the FNs are able t increasing. to allocate more of its CRBs to the DSSs nearby. However, th whenalltheavailableCRBsofFNsareallocatedtotheDSSs, the utility of the FNs stops improving and converges to one VI. RELATEDWORKS specific value. When the number of FNs increases, with the In the literature, fog computing has been advocated to be same value of λ, as the FNs are able to provide more CRBs the promising future of the cloud. The concept of pulling to the DSSs, the converged value is higher. the cloud closer to the users has been widely considered in InFig.6,weobservetheutilityofallDSOswhenthevalue previous work. In [8], the authors put forward the concept of µ increases. As shown in the figure, when µ increases, of mist computing, aiming to distribute the cloud and its each DSS is able to be served with a less amount of CRBs. benefits deeply into the network. In [9], the deployment of Therefore, the DSO is able to set a higher price to the DSSs edge cloud was proposed. From the DSSs’ side, the edge and receive high revenues. Moreover, when the number of cloud was able to surrogate the requirements and simplify DSSsincreases,astheDSO isableto servemoreDSSsatthe the management of the network. From the servers’ side, the same time, the total utility of DSOs also increases. edge cloud can exploit content and support service delivery In Fig. 7, we evaluate the relationship of utilities and the in an efficient way. Without deploying massive data centers tolerance service delay t of DSSs. As shown in the Fig. 7a, with high cost and latency, [10] on the other hand strength- th whenthevalueoft increases,eachDSSisabletobeserved ened the importance of small distributed data center designs. th with less CRBs. Thus, the FNs will supply less CRBs in the The authors took email delivery as an example and showed network,andtheutilityoftheFNsgenerallydecreases.InFig. the advantages of geo-diversity characteristics of micro data 7a,astheDSOisablesetahighpriceforitsservices,withthe centers. In [11], a novel and distributed computing platform, valueoft increasing,theutilityofDSOgenerallyincreases. called nano data centers was proposed. The authors evaluated th Moreover, in Fig. 7c, even though the DSS is able purchase the energy consumption of nano data centers and showed a less CRBs with high t , the price of CRBs set by the DSOs significant improvement on energy efficiency, compared with th alsoincreases.Furthermore,astheDSSsuffersalotwithhigh the traditional data centers. In [12], considering the existing 9 580 6×105 1900 560 5 1800 540 520 4 1700 Utility of FNs458000 Utility of DSOs3 Utility of DSSs1600 460 2 1500 440 1 1400 420 400 0 1300 40 50 60 70 80 90 100 40 50 60 70 80 90 100 40 50 60 70 80 90 100 t (ms) t (ms) t (ms) th th th (a) The utility of all FNs versus t (b) The utility of all DSOs versus t (c) The utility of all DSSsversus t th th th Fig. 7: The utilities versus t th telecommunicationand Internetservice providers,the authors [26], the authors consider a Stackelberg game between data showed that it was required and beneficial to leverage the center and buses in the smart city, where each buses collect existing infrastructure and provide value-added services with data along its route and compete with other buses for the FNs. In[13], theauthorsoutlinedthevisionoffogcomputing reward forwarding to the data center. In the game, following and overviewed the important features of fog computing. the proposed heuristic algorithm, the Stackelberg equilibrium In [14], the network optimization with fog computing was isshowntobeachievedwherethedatacenterandeachbusare considered.Asthedatacenterswereawareofthelocationsof able to reach maximum utility. In [27], the authors formulate DSSs with fog computing, dynamic adaptation of computing aStackelberggameforpowerallocationofdatacentersinthe resources to the DSSs’ conditions was proposed. In [15], the cloud.Inthegame,thepowergridcontrolleractsastheleader authorscomparedthecloudcomputingwiththefogcomputing and sets prices of the provided energy based on the current and showed some significant characteristics of fog, which amount of renewable energy and costs. Observing the prices, was required for current data services. In [16], the authors the cloud controller, i.e., the follower, determines the optimal elaboratedtheroleoffogcomputinginsiximportantscenarios amount of enery to purchase and do resource allocation for and surveyed the security issues with fog computing. its data centers. With backward induction, the near-optimal strategiesofbothplayersinthegamecanbeachieved.In[28], Moreover, fog computing has been widely considered to the authors model the interaction between the monopolistic be beneficial for the IoT. In [17], the authors overview the data center opertor and the customers as a Stackelberg game. opportunitiesandchallengesoffog,especiallytheapplications In the game, the pricing strategies of the monopolistic data offogcomputinginIoT.In[18],theauthorsdevisethemethod center operator the corresponding behavior of data service of MEdia FOg Resource Estimation (MeFoRE), to provide customers is detailedly analyzed in both homogeneous and resource estimation based on the service give-up ratio and heterogeneouscustomerscenarios.In[29],theauthorsadopted to enhance QoS based on the previous QoE. [19] addresses the Stackelberg game to solve the problem of minimizing the utility based pairing problem between the fog nodes and energy consumption in the data center networks. In [30], the IoT deviceswith the Irving’smatchingalgorithm.In [20], the authors proposed a multi-leader multi-follower Stackelberg authors propose a distributed dataflow programming model gametoaddressandcooperationproblemsamongWi-Fi,small for IoT devices to optimize resource allocation on computing cell and macrocell networks. In [31], the authors combined infrastructuresacrossthefogandthecloud.In[21],theauthors the Stackelberg game and the bargainingto design a resource consider issues of resource prediction, customer type based allocation problem in a multi-tier LTE unlicensed network. resource estimation and reservation, advance reservation, and Furthermore, the auction mechanisms are also powerful tools pricing in the fog computing for IoTs. [22] considers the to solve the problem. In [32]–[34], the resource management requirements of mobility, scalability, reliable control and ac- problem could be perfectly optimized, but it requires high tuation in some challenging scenarios of IoT to show the communicationandcomputationoverhead.In[35],theauthors benefits and significance of fog computing. Considering the adopted the generic game theoretic framework to identify advantages of fog computing, the authors in [23] discuss and importantedgesinthecontextofk-edgeconnectivitybetween propose a procedure to be implemented in smart phones for certain pairs of nodes in a general given network. In [36], UV measurement. the graphical game was put forward to analyze the optimized In order to solve the resource management problems in behaviors of each node in a general graph. a network system with a distributed fashion, game theory has been shown as a powerful tool [24]. In the literature, VII. CONCLUSIONS most of the cases, the network system is normally modeled In this paper, we proposed a joint optimization framework as a bipartite or a multi-tier graph. Based on this model, in the multi-FN, multi-DSO and multi-DSS scenario for IoT in [25], a Stackelberg game theoretic model was shown for fogcomputing.Inthe framework,we first modeledthe Stack- dynamic bandwidth allocation between virtual networks. In elberg games to solve the pricing problem of the DSOs and 10 resource purchasing problem of the DSSs. Then a many-to- [19] S. F. Abedin, M. G. R. Alam, N. H. Tran, and C. S. Hong, “A Fog manymatchingwas proposedbetweenthe DSOs and the FNs BasedSystemModelforCooperative IoTNodePairingUsingMatching Theory,”in201517thAsia-PacificNetworkOperationsandManagement todealwiththeDSO-FNpairingproblem.Finally,weapplied Symposium (APNOMS),pp.309-314,Busan,Korean,Aug.2015. another many-to-many matching between its paired FNs and [20] N.K.Giang,M.Blackstock,R.Lea,andV.C.M.Leung,“Developing serving DSSs to solve the FN-DSS pairing problem within IoTApplicationsintheFog:ADistributedDataflowApproach,”in2015 5th International Conference on the Internet of Things (IOT), pp. 155- the same DSO. For each stage of the problem,all participants 162,Seoul,Korean,Oct.2015. were able to achieve the equilibrium or stable results where [21] M. Aazam, and E. N. Huh, “Fog Computing Micro Datacenter Based no one was able to change its behavior unilaterally for a DynamicResourceEstimationandPricingModelforIoT,”in2015IEEE 29thInternationalConferenceonAdvancedInformationNetworkingand higher utility. Simulation results showed that all FNs, DSOs Applications, pp.687-694,Gwangiu, Korean,Mar.2015. and DCOs were able to reachoptimalutilities for themselves, [22] M. Yannuzzi, R. Milito, R. Serral-Graci, D. Montero, and M. Ne- and high performance of the proposed framework could be mirovsky, “Key Ingredients in an IoT Recipe: Fog Computing, Cloud Computing,andmoreFogComputing,”in2014IEEE19thInternational achieved compared with the data services without fog nodes. Workshop on Computer Aided Modeling and Design of Communication LinksandNetworks (CAMAD),pp.325-329,Athens,Greek,Dec.2014. REFERENCES [23] B. Mei, W. Cheng, and X. Cheng, “Fog Computing Based Ultravio- let Radiation Measurement via Smartphones,” in in 2015 Third IEEE [1] J.Gubbi,R.Buyya,S.Marusic,andM.Palaniswami,“InternetofThings Workshop on Hot Topics in Web Systems and Technologies (HotWeb), (IoT): A vision, architectural elements, and future directions,” Future WashingtonD.C.,Nov.2015. Generation Computer Systems,vol.29,nol.7,pp.1645-1660,2013. [24] Z. Han, D. Niyato, W. Saad, T. Basar, and A. Hjorungnes, “Game [2] I. Goiri, K. Le, J. Guitart, J. Torres, and R. Bianchini, “Intelligent Theory inWireless and Communication Networks: Theory, Models and Placement of Datacenters for Internet Services,” in Proc. IEEE ICDCS, Applications,” inCambridgeUniversity Press,UK,2011. pp.131-142,Minneapolis, MI,Jun.2011. [25] C.Wang,Y.Yuan,C.Wang,X.Hu,andC.Zheng,“VirtualBandwidth [3] H.Zhang,Y.Xiao,S.Bu,D.Niyato,R.Yu,andZ.Han,“FogComputing Allocation Game in Data Centers,” in Proc. 2012 IEEE International inMulti-tierDataCenterNetworks:AHierarchical GameApproach,”in ConferenceonInformationScienceandTechnology,pp.682-685,Hubei, Proc.IEEEICC’16,KualaLumpur,Malaysia, May2016. China,2012. [4] C. Guo, G. Lu, and H. J. Wang, “Secondnet: a Data Center Network [26] M.Wu,D.Ye,S.Tang,andR.Yu,“CollaborativeVehicleSensinginBus Virtualization Architecture with Bandwidth Guarantees,” in Proc.of the Networks:AStackelbergGameApproach,”2016IEEE/CICInternational 6thInternational Conference. ACM,pp.15,Graz,Austria,Sep.2010. ConferenceonCommunicationsinChina(ICCC),Chengdu,China,Aug. [5] A.Gandhi,M.Harchol-Balter, R.Das,andC.Lefurgy,“OptimalPower 2016. Allocation in Server Farms,” ACM SIGMETRICS Perform. Eval. Rev., [27] Y. Wang, X. Lin and, M. Pedram, “A Stackelberg Game-Based Opti- vol.37,no.1,pp.157-168,Jun.2009. mizationFrameworkoftheSmartGridWithDistributedPVPowerGen- [6] A.E.Roth,“TheEvolutionoftheLaborMarketforMedicalInternsand erationsandDataCenters,”inIEEETransactionsonEnergyConversion, Residents:aCaseStudyinGameTheory,”JournalofPoliticalEconomy, vol.29,no.4,pp.978-987,Dec.2014. vol.92,no.6,pp.991-1016, Dec.1984. [28] H.Wang,Y.Zhao,andH.Guan,“OnPricingSchemes inDataCenter [7] Z.Liu,M.Lin,A.Wierman,S.H.Low,andL.L.Andrew,“Geographical NetworkwithGameTheoreticApproach,”201423rdInternationalCon- Load Balancing with Renewables,” ACM SIGMETRICS Perform. Eval. ferenceonComputerCommunicationandNetworks(ICCCN),Shanghai, Rev.,vol.39,no.3,pp.62-66,Dec.2011. China,Aug.2014. [8] B.Ahlgren,P.Aranda,P.Chemouil,S.Oueslati,L.Correia,H.Karl,M. [29] B. Yang, Z. Li, S. Chen, T. Wang, and K. Li, “Stackelberg Game Sollner, and A. Welin, “Content, Connectivity, and Cloud: Ingredients Approach for Energy-aware Resource Allocation in Data Centers,” in fortheNetworkoftheFuture,”IEEECommun.Mag.,vol.49,no.7,pp. IEEETransactions onParallelandDistributedSystems ,toappear. 62-70,Jul.2011 [30] H. Zhang, M. Bennis, L. A. DaSilva, and Z. Han, “Multi-leader [9] S.IslamandJ.-C.Gregoire,“NetworkEdgeIntelligencefortheEmerging MultifollowerStackelbergGameamongWi-Fi,SmallCellandMacrocell Next-Generation Internet,” Future Internet, vol. 2, no. 4, pp. 603-623, Networks,”inProc.IEEEGlobecom’14, Austin,TX,Dec.2014. Dec.2010. [31] H. Zhang, Y. Xiao, L. X. Cai, L. Song, N. Dusit, and Z. Han, [10] K.Church,A.Greenberg,andJ.Hamilton,“OnDelivering Embarrass- “HieraticalCompetitionforLTEUnlicensedusingStackelbergGameand ingly Distributed Cloud Services,” in Proc. ACM HotNets, pp. 55-60, Bargaining,” inProc.IEEEGlobecom’15, SanDiego,CA,Dec.2015. Calgary,Canada, Oct.2008. [32] L. M. Ausubel, “An Efficient Dynamic Auction for Heterogeneous [11] V. Valancius, N. Laoutaris, C. Diot, P. Rodriguez, and L. Massoulie, Commodities,” AmericanEconomicReview, vol.96,no.3,pp.602-629, “GreeningtheInternetwithNanoDataCenters,”inProc.ACMCoNEXT, Jun.2006. pp.37-48,Rome,Italy,Dec.2009. [33] G. Demange, D. Gale, and M. Sotomayor, “Multi-item Auctions,” [12] M. B. Mobley, T. Stuart, and Y. Andrew, “Next-Generation Managed JournalofPolitical Economy,vol.94,no.4,pp.863-872,Aug.1986. Services: A Window of Opportunity for Service Providers,” CISCO [34] D. Mishra and R. Garg, “Descending Price Multi-item Auctions,” Technical Report,2009. Journal of Mathematical Economics, vol. 42, no. 2, pp. 161-179, Apr. [13] F.Bonomi,R.Milito, andJ.Zhu,“FogComputingandits Roleinthe 2006. InternetofThings,”inProc.ofthefirsteditionoftheMCCworkshopon [35] R.Narayanam,“AGameTheoreticApproachtoIdentifyCriticalCom- Mobilecloudcomputing.ACM,pp.13-16,Helsinki,Finland,Aug.2012. ponents in Networked Systems,” in Proc. 2012 Annual SRII Global [14] Z.Jiang,D.S.Chan,M.S.Prabhu,P.Natarajan,H.Hao,andF.Bonomi, Conference, pp.515-524,SanJose,CA,2012. “ImprovingWebSitesPerformanceUsingEdgeServersinFogComput- [36] M.Kearns,“GraphicalGames,”AlgorithmicGameTheory,bookchapter, ingArchitecture,” inServiceOrientedSystemEngineering(SOSE),2013 CambridgeUniversity Press,pp.159-180,2007. IEEE7thInternationalSymposiumon,pp.320-323,RedwoodCity,CA, Mar.2013. [15] M. Firdhous, O. Ghazali, and S. Hassan, “Fog Computing: Will it be theFutureofCloudComputing?”inProc.ThirdInternationalConference onInformaticsandApplications(ICIA2014),pp.8-15,KualaTerengganu, Malaysia, Jul.2014. [16] I.Stojmenovic, andS.Wen,“TheFogComputingParadigm:Scenarios andSecurityIssues,”inProc.ComputerScienceandInformationSystems (FedCSIS),2014FederatedConference on,Warsaw,Porland,2014. [17] M. Chiang, and T. Zhang, “Fog and IoT: An Overview of Research Opportunities,” ininIEEEInternetofThingsJournal, tobepublished. [18] M. Aazam, M. St-Hilaire, C. H. Lung, and I. Lambadaris, “MeFoRE: QoE Based Resource Estimation at Fog to Enhance QoS in IoT,” in 2016 23rd International Conference on Telecommunications (ICT), Thessaloniki, Greek,May2016.

See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.