66 China Communications • April 2020ity of service [1]. Evolved from cloud radioaccess networks, fog radio access network hasshifted the resources computing to the edgenodes, named fog devices (FDs), particularly,to the legacy radio units (RUs). Fog architecture is promoted to meet the rising demand forlow latency applications and higher bandwidthin future networks, that is represented by 5G,expected to launch in 2020 [2]. The issue withcloud radio access network is that the user(UE) is required to connect to the cloud centerso as its data is processed. Hence, a possiblenetwork delay can happen due to enlargeddistances that increases the multipath delays/fading, not to mention the delay results frompackets processing. To solve this problem,transferring some functions of the cloud baseband units’ (CBB) that reside in the cloud center to the radio units (RUs) will revoke the delay initiated amongst RUs and the CBB pool.It is worth noting that the traditional cloud radio access network architecture included manyRUs connected to the CBB pool, that is agroup of inter-connected CBBs. Furthermore,the geographical position of the pool itselfshould be optimised to ensure reduced distances to the RUs, which decreases the link delaywhile transmitting the processed data to theRUs, and then to the UEs. Nevertheless, thefar-away placed RUs still witness increaseddelay values. Bringing data processing to theRUs to reduce the delay is not always beneficial, but there are trade-offs, some of themAbstract: The industry of cellular networks isevaluating the new architectures to ensure anenhanced performance. Fog communicationis the new paradigm that presented to unleashedge computing. In this paper, we introduceda mathematical framework to evaluate thetrade-offs of Fog proposal. Specifically, testing the power consumption, delay and energyefficiency in comparison with traditional cloudradio access networks. Although the literaturehas showed that fog radio access networksprovides an enhanced delay performance, thispaper shows that an enlarged amount of poweris consumed, which degrades the energy efficiency in comparison with traditional cloudcounterpart. However, the level of such devolution depends on the number of deployedfog devices that directly influences the powerconsumption. This paper also shows that enhancing the delay by using fog architectureis not a straight forward process, but requiresa particular caring in terms of choosing theappropriate mode while placing/installing fogfunctions within fog devices.Keywords: fog communications; cloud networks; models; modelling; power; energy efficiency; delayI. INTRODUCTIONRecently, fog networks have been proposedto seize the dramatic increase in the trafficdemands as well as providing improved qualReceived: Sep. 12, 2018Revised: Jan. 23, 2019Editor: Shangguang WangEnergy Efficiency and Latency Analysis of FogNetworksRaad S. AlhumaimaDepartment of Communications, College of Engineering, University of Diyala, Diyala, Iraq* The corresponding author, email: [email protected]NETWORKS & SECURITYAuthorized licensed use limited to: Macquarie University. Downloaded on October 06,2021 at 00:51:48 UTC from IEEE Xplore. Restrictions apply.China Communications • April 2020 67and RU is different. The FD consists of RUplus added fog functions from CBB servers[6]. Hence, the term RU is used when referring to traditional cloud network, while FDis used to refer the fog network.4) Coordination: mainly, fog architecturecan be differentiated into two types, centralized and distributed. In the former, theCBB hands over some functions (such asstorage, resources management, and dataprocessing) to the RUs and even the UEs[7]. On the other side, in distributed type,standalone fog nodes are deployed withinthe cell coverage to serve the UEs with necessary call initiations and authentications,while the RUs are limited to only transmit/receive the radio signals and direct them tothe CBB pool [8]. Subsequently, fog nodesare connected to the RUs and the CBB pool toserve the real-time collaboration radio signalprocessing (CRSP), flexible cooperative radioresource management (CRRM) and caching.This process requires new resources management algorithms and optimisation, especiallywhen a decision is made to offload the computation to fog nodes and achieve the neededcooperation with the CBB pool.However, an important factor that advocates fog over tra- ditional cloud networks,or vice versa, is the percentage of power dropfrom the pool that is offloaded to the fog device (FD). This inconstancy of power surelyinfluences the network energy effiency (EE).The functions CRSP and CRRM are the primereason for this consumption, they operate toserve the following functions:1) Compressing and forwarding UEs information to the pool through fronthaul link, offering interference management and bandwidth sharing amongst the FDs. If managing the inter- ference process by local CRSPand CRRM was not efficient, CRSP andCRRM are executed in the pool to retrievedthe traditional cloud architecture [9], andFD is degenerated as traditional RU. Because of this architecture reformulation, thepower consumption can be dramatically enlarged due to operating the same functionsare barely evaluated in the available literature,these trade-offs can be summarised:1) Time: the latency in fog networks will benoticeably decreased by shortcutting pingpong communications amongst UEs-RUsand RUs-CBB pool, to only UEs-RUscommuni- cations. However, the processingtime of the RUs might be larger than in thecloud architecture. That is because theseRUs have a limited processing capacity,also, they are less efficient than the higher processing capacity found in the CBBservers. Consequently, the processing timeof the RU can grow exponentially with theamount of processed data due to shifting tofog architecture [3]. This matter might advocate against the fog networks, especiallywhen the volume of com- ing traffic is high.To establish a processing cabability withinthe proprietary built-in RUs, the radio frequency (RF) unit that is found in the RUhas to be replaced with programmable electronic boards that are capable of hosting thetransferred CBB/fog functions. However,changing the design of these RUs requiresan investigation about the delay in comparison with traditional counterparts.2) Backhaul: serving the necessary real-timenetwork contents from the RUs means thebackhaul delay congestion is relieved regarding both bandwidth usage and storagecapacity.3) Power: the power consumed to processfog functions in the CBB pool, is nowshifted to the RUs [4]. Such procedure isnot necessarily equalised, i.e. the deductedpower consumption from CBB pool bytransferring some of the functions to theRUs, may or may not be equal to the powerconsumption that is added to the RUs dueto processing these functions. This is because the number of CBBs and RUs is notidentical, as one CBB might serve manyRUs [5]. This means the migrated functionsfrom one CBB will be installed on severalRUs, this magnifies the power consumptionas many as the number of RUs. It is worthmentioning the architectural design of FDIn this paper, we introduced a mathematicalframework to evaluatethe trade-offs of Fogproposal.Authorized licensed use limited to: Macquarie University. Downloaded on October 06,2021 at 00:51:48 UTC from IEEE Xplore. Restrictions apply.68 China Communications • April 2020fog network, traditional cloud or hybrid architecture. This judgement can be based on different parameters such as power consumption,delay or energy efficiency. The above parameters are all objected in the proposed model.The rest of this paper is organised as follow. In Section II, the available literaturereview is discussed. In Section III, the powerconsumption is analysed and modelled for different types of fog architectures, while the EEevaluation of these types is given in SectionIV. The delay of fog scenarios is investigatedin Section V. Furthermore, the results are discussed and analysed in Section VI. Finally, theconclusion and future work are presented inSection VII.II. RELATED WORKMost of the literature comprehensively discussed the delay of fog networks, which isthe main reason why fog was proposed, whilescarce discussion can be found about thepower consumption. For example, in [11], thelatency of fog network is analysed, and theconcept of multiple fog nodes that operate ondifferent computing tasks is introduced. However, the intuitive gain of the delay is compromised with the increase in power consumptionwhere fog functions are abstracted. Similarwork can be found in [12]. The overall performance of fog networks was introduced in[7] regarding spectral and EE. Unfortunately,there was no mathematical model to describehow the optimisation parameters influenceeach others. An offloading technique is proposed in [1] to split the computation powerinto fog and cloud parts to form a hybridoriented fog network design. Moreover, thepower consumption of FD and cloud serversis simply described in [13] as a linear functionof the coming traffic, this work skipped modelling and comparing the numerical resultswith the state of the art consumption. In [14]a channel and power allocation is consideredusing NOMA in fog edge caching. This workshowed that NOMA-based fog networkscan provide an en- hanced utility for mobileat two different places, the pool and FDs.This matter requires speculation about thetriggered power consumption.2) CRSP and CRRM can have similar functionality of small base stations, where coordinated multi-point process is de- ployed to mitigate the inter-tier and intra-tier interferences.Furthermore, the adjacent FDs are interconnected to form a different kind of topologythat implements the local CRSP [9].It is very important when proposing the newFDs that they are practically implementable.The RU unit includes RF unit and power amplifier (PA). In traditional networks, both areun-programmable and propriety built-in devices, which makes updating the new fog functions is dramatically intractable. Hence, thequestion is how to install the new functions?There are three ways: first, combining RF andfog functions within the same board in virtualisation oriented design, second, thanks to thesystem on chip technology, the new functionsare programmed and executed on a separateelectronic board, but still located within theRU. Finally, deploying a geographically separated FDs that are coordinating and collaborating with the legacy RUs and CBB pool [10].However, these cases are all power consumingthat require critical investigation by offeringan effective and measurement based mathematical models, which are discussed in thispaper. In addition, most of the literature onlyexamined the delay in the channel, where theprocessing delay in the FD itself is ignored.This latter can be enlarged due to virtualisingthe functions within the FD, installing newelectronic board, or deploying standalone FDswithin the RUs’ coverage. Hence, this matterrequires new evaluations, which are also modelled in this paper. The proposed model in thispaper can also help understanding how the different parameters that are concurrent with fogtechnology affect each other in the future, andwhat is the most effective parameter withinthis development. Hence, the service providersor network operators are obligated to test thetrade-offs due to fog adaptation, after, a decision can be made whether it is worth to adaptAuthorized licensed use limited to: Macquarie University. Downloaded on October 06,2021 at 00:51:48 UTC from IEEE Xplore. Restrictions apply.China Communications • April 2020 69RUs and FDs. The FD is modelled using threedifferent ways:1) Virtualised Fog: fog functions arevirtualised/coexisted within the RF unit. Ifwe assume the traditional RF total powerconsumption is Prf. Hence, the virtualised fogfunctions might increase this consumption linearly or exponentially. In [3], and [22], it wasmentioned that, once processing the arrivingpackets/resource blocks, the power consumption grows linearly. However, due to virtualisation process, this dynamic traffic is able togrow the power consumption exponentially[23]. This behaviour is constructed becauseeach individual virtual machine (VM) functionis linearly proportional to the dynamic load,when all VMs operate at the same time, thepower consumption will be enlarged exponential. This means when all the VMs try toaccess the limited resources of the RF board,an exhaustive data processing is triggered thatoperates the device at full load. Nevertheless,both linear and exponential behaviours are included within one model. That is, the additionto RF power usage is denoted by P P crsp crrm + ,where Pcrsp and Pcrrm denote the power consumption of CRSP and CRRM, respectively.The increasing in CRSP’s power consumption while increasing the number of resourceblocks (RBs) is given by dP dRB Pcrsp / = κ crsp,when solving this equation, it yieldsP P ecrsp crsp =int RB κ×,(1)(2) where Pcrspint in the initial/static power consumption of CRSP when no RBs are processed. Thesame style is used to model CRRM function,whereP P ecrrm crrm =int RB λ×, (3)and Pcrrmint denotes the initial power consumption of CRRM function, κ and λ are the increasing constants.The reason why exponential expression isused in this modelling is that once the constants κ or λ approaches 0, the model tend tobe linear, hence, this pattern joins both linearand exponential at the same time. In addition,networks. The interference, resources optimization and mobility management have beendiscussed in [15]. This article showed that fognetworks are able to reduce signaling cost inthe backhaul. However, the processing delay isignored. To meet the social requirements (highbit rate and less delay) of the UEs, the work in[16] proposed an algorithm that facilitate device-to-device communications, by doing so,the network’s throughput and fronthaul burdenare improved. In addition, [17] analysed thedelay and transmission rate of fog networkin comparison to tradition cloud, however,the power consumption, EE and different fogdesigns are not discussed. Controlling the distributed FDs using software defined controllerto meet the required scalability is proposed in[18]. Moreover, an integration of Internet ofEverything (IoE) and Fog Computing (FC)paradigms that produces Fog of Everything(FoE) has been proposed in [19], where theperformance of the time-energy is analysedand evaluated.In [20], an algorithm is proposed to manipulate the case of collapse in fog networks viacompressing the bandwidth, which improvesthe network performance and ensures the network’s survivability. Finally, software definednetworks and network function virtualisationtechniques are integrated in one architecture toserve and escalate the utility of fog computing.This method was presented in [21] without realising the cost of the virtualisation regardingthe power consumption. The latter can simplyreach tens of Watts in the target and sendingservers, hence, careful investigation is demanded.III. FOG COST MODELLINGMainly, the proposed model included twofundamental pillars, these are delay and powerconsumption:3.1 Fog network powerconsumptionIn this section, we modelled the fog networkpower consumption, including the CBB pool,Authorized licensed use limited to: Macquarie University. Downloaded on October 06,2021 at 00:51:48 UTC from IEEE Xplore. Restrictions apply.70 China Communications • April 2020resented by loss factors γ dc f for DC-DC consumption, and γ mc f for AC-DC consumption.Hence, the model is updated toP P P P PFD rf crsp crrm amp dc mc1= + + + × ( ) / ( ) γ γ f f .(9)2) Separated Fog: fog functions are implemented on a separate board, rather than RFunit, but still within the RU. The consumptionof the extra board is jointly added to the power consumption of RF unit. It was assumedthat the power consumption of the board isPadd, which is based on its type, for example,FPGA ones consume about 30 W. However, this consumption is traffic, performing,and operating frequency based. System onchip boards consume much less power, butthey are very expensive compared to FPGAboards. However, we generally assumed theinitial power consumption of any type isPaddint, that is driven up to the maximum due toupgraded number of processed RBs. Hence,the power consumption can be expressed asdP dRB P eadd / = add int RB ζ × , where ζ denotes thepower increasing constant. Finally, the powerconsumption of second type of FDs is formulated asP P P PFD rf add amp dc mc2= + + × ( ) / ( ) γ γ f f . (10)3) Heterogeneous Fog: fog functions useda separate device and located within RUscoverage. At this level, it is required to modelthe RU and FD individually because the FDis physically and geographically separatedthan RU. This means it will be dealt with asstandalone device that possess its own powerconsumption and conversion losses. To modelthat, the FD is represented as Pfde. This consumption however, is susceptible to the dynamic load. Hence, it was assumed thatP P efde fde = + int ∑K kπ *RBk, (11)where Pfdeint denotes the static power consumption of the device when no load is existed, RBkindicates the number of RBs that are processedby k-th FD. In addition, the legacy/traditionalRU power consumption is formulated asP P Pru rf amp dc mc = + × / ( ) γ γ f f . (12)this expression permits to exhibit the staticscale of the consumption, represented by Pcrsp intand Pcrrmint to reference the usage of power whenno load is processed. The other part of the FDis the power amplifier, whose consumptionPamp is given asP P f amp max = / , (4)where Pmax denotes the maximum output power of the antenna, and f is its efficiency [22].Usually, each electronic board requires proper DC voltage to operate. First the mains supplyis converted to the DC voltage (AC-DC), after,the DC is converted to another DC level that issuitable to each unit within the FD. Therefore,these conversions are power consuming. Asthe process of DC-DC or AC-DC conversion isnot 100% efficient, there will be a conversionloss. If we assume that Ldc is DC-DC conversion loss, it is represented as a function of theconverter efficiency ( ) ηdc . Hence, loss functionis modelled as decay function with exponentialdecay constant (q), i.e.,L L e ( ) η = o -ηq, (5)where Lo is the initial loss of time (t = 0). Thismodel means the more efficient the conversion is, the lower losses take place. Moreover,the constant q can vary amongst differentdevices as this matter is complied to the quality of manufacturing. Intuitively, the powerconsumption of DC conversion is linearlyproportional to the consumption of other unitswithin the FD. Hence, the DC consumption ismodelled asP L P P P Pdc dc dc rf crsp crrm amp = + + + ( )( η ). (6)By using the same logic, AC-DC is modelledasP L P P P Pmc mc mc rf crsp crrm amp = + + + ( )( η ). (7)Hence, the power model of the first scenario of FD is given asP P P P P P PFD rf crsp crrm amp mc dc1= + + + + + , (8)where Prf denoted the RF unit power consumption. It is worth noting that RF unit is trafficindependent [24]. To simplify the model, theDC-DC and AC-DC consumption are repAuthorized licensed use limited to: Macquarie University. Downloaded on October 06,2021 at 00:51:48 UTC from IEEE Xplore. Restrictions apply.China Communications • April 2020 71P Ant P PCC = + × × ( ( )) / ( CB O dc mc cc p p p p γ γ γ ), (13)where Ant is the number of serving antennas. It is worth mentioning that anamount of power that is consumed withinthe FDs, including all scenarios, has beendropped from the CBB pool. This amountis equal to P P crsp crrm + . Subsequently, thepool power consumption PCC is updated to:Ant P P P P × + – – × × ( CB O crsp crrm DC MC CC p ) / (γ γ γ p p p ) .Subsequently, the total power consumption offog network can be modelled by adding thepower consumption of FD to the pool, as follows:P G P K N PFN CC = × + ∨ × ( ) FD. (14)where G and K denote the number of CBBsand FDs, respectively. Note that, K denotesthe number of separate FDs of the third scenario, and N is the number of RU that operatesas FD. In addition, PFD might be equal to PFD 1 ,PFD2, or PFD 3 according to the tagged scenario.IV. FOG EE ANALYSISThe criteria of measuring network EE is superior over bare capacity or spectral efficiencyevaluation because EE shows the power indicator, which gives additional dimension whileassessing the network performance. To evaluate the EE, we have assumed a fog networkthat contains total number of FD(N) and totalnumber of UEs(U). The small scale fadingbetween FD (n) and the UE(u) is denoted asHn u , , and it was assumed Rayleigh fading. Thepower received by the UE(u) from FD(n) isgiven asP P H rn u FD n u n u , , , =t, (15)where PFDt denotes the transmitted power fromthe FD, r d n u n u , , = -α denotes the path loss between FD(n) and UE(u), α is the path loss exponent, while Hn u , represents the channel gainfrom n-th FD to u-th UE. Furthermore, dn u , isthe straight line distance between n-th FD andu-th UE, which is given asd x x y y n u n u n u , = – + – ( ) ( ) 2 2 , (16)Hence, the total consumption of thethird type of fog network is modelled asP P PFD ru fde3= + .On the other side, the pool contains manyCBBs whose consumption vary linearlywith the incoming traffic. Hence, the powerconsumption of a single CBB is linearly proportional to the traffic load and number ofantennas [22]. This concept can be mappedas the change in the amount of CBB’s powerconsumption is proportional to the change inthe number of processed RBs (dP dN CB RB / ) isequivalent to δ PCB, where δ is the linear increment factor due to increasing the processedRBs.It is worth noting that this behaviour is originally triggered due to increasing the transmitted RBs increases the power consumption inthe CBB as more RBs are processed, and eachRB demands a power usage, which increasesthe power consumption of the CBB.Also, once the number of transmitted RBs isincreased, it requires more transmission power(Pmax) to send them all. This results in increasing the power consumption of PA, which increases the total network consumption.Subsequently, the DC-DC conversion power consumption increases linearly with theconsumption of all the devices within the pool,including PCB and POp. That is, the incrementin DC-DC power consumption to the changeof PCB and POp is dP d P P P dc CB O dc / ( ) + = p η ,where η denotes the linear increment constant. The same method is used to modelAC-DC and cooling power consumption,which yields dP d P P P mc CB O mc / ( ) + = p ξ anddP d P P Pcc CB O cc / ( ) + = ∧ p f o r A C – D C a n dcooling power consumptions, respectively,where ξ and ∧ are the increment constant factors. To simplify this, the DC, AC and coolingconsumption are updated to loass factors γ dc p ,γ pmc and γ pcc to represent the DC-DC, AC-DCconversions and cooling losses in the pool,respectively. Hence, the power consumption inthe cloud centre is formulated asAuthorized licensed use limited to: Macquarie University. Downloaded on October 06,2021 at 00:51:48 UTC from IEEE Xplore. Restrictions apply.72 China Communications • April 2020Finally, the third case is when fog functionsare implemented on a standalone device. ItsEE is evaluated as follows: EE3 =BW log P(1 ) 2 , , + Γ n u n uK Pfde ( ) ( ) G P N P × + × +CC ruγ γdc mcs s× ×, (21)where γ dc s γ mc s indicate the DC-DC and AC-DCpower conversion loss factors of the separatefog device, respectively.V. FOG NETWORKS DELAYIt was assumed the delay from FD to the poolisτn o n o , , = d c / , (22)where (o) denotes the origin of the pool’s geographical position, c is the speed of light incase wireless link is found amongst the pooland FDs. In case of optical fiber links,τn o n o opt , , = d c / , (23)where (c c ind opt = / ) is the speed of light inside the optical fiber, and (ind) is its refractiveindex. Subsequently, the delay from the userto the FD is denoted asτn u n u , , = d c / , (24)this link is only wireless. Clearly, if the UEis served from the FD rather than the pool,the delay gain will be equivalent to τ n o , as theUEs are no longer connected to the pool allthe time, rather, to RUs. However, once theservice type and authentication are establishedto the UE, the latter can return to the pool toprovide the persistent resources allocation.Nevertheless, when fog functions are shiftedfrom the pool to the FD, the latter witnessesan enhanced processing delay due to virtualisation. The work in [3] has mentioned that theexecution/processing time is linearly proportional to the processed RBs and modulationcoding scheme (MCS) that is used to transmitthese RBs. Nevertheless, due to virtualisation,this time can grow exponentially as one virtual machine might take 5 times more cycleto process a packet than a bare-metal device[25]. Therefore, a model is required to combine both concepts. If we assume τ device is thewhere xn, yn, xu, yu indicate the Cartesian xand y axes of the FDs and UEs, respectively.In the first scenario, we have removed theamount of power (P P crsp crrm + ) from CBBpool and place it in the FD, the power effectof these functions on both sides of the network appears neutralised, is it? As the numberof deployed FDs is usually larger than thenumber of CBBs (N G > ), hence, the consumption of fog network becomes more thantraditional cloud, the power difference between the two networks (P*) can reach up to:[( )( )] / ( ) N G P P – + × crsp crrm dc mc γ γ f f .However, the direct calculation of fog EEcan be given as EE =BW log P (1 ) 2 , ,+ Γn u n u,(17) PFNthis expression is not a comparison-wise withtraditional cloud networks. Hence, we haveshifted this EE model to an expression that is,able to compare both networks, encloses thetraditional cloud, and shows the power divergence with the inherited fog. If we assume RUpower consumption isP P Pru rf amp dc mc = + × ( ) / ( ) γ γ f f , (18)hence, the fog EE of the first scenario can bemodelled asEE1 =( ) ( ) G P N P PBW log P× + × +CC ru(1 ) 2 , , + Γ n u n u* , (19)where Γ =m u ,h rm u m uB N, ,odenotes the signal tonoise ratio(SNR). This formulation shows bothcloud and fog representatives, such as numberof number of traditional CBBs (G), numberof RUs or FDs of the first type (N), and theirpower consumptions variation (P*).The second case is more straight forward,the functions are implemented on a separateboard, whose power consumption is Padd.Hence, the EE can be modelled by adding thisamount from RU’s power consumption. Ifwe assume that the number of these separateboards is equal to S, the EE can be modelled as:EE 2 =( ) ( ) G P N P × + × +BW log PCC ru(1 ) 2 , , + Γ n u n uγ γ f f dc mcP×add. (20)Authorized licensed use limited to: Macquarie University. Downloaded on October 06,2021 at 00:51:48 UTC from IEEE Xplore. Restrictions apply.China Communications • April 2020 73power consumption formula, as in PFD 1 and PFD 2 .However, this means that DC-DC conversionincreases the total consumption of the units(FD, RU, CBB, etc) by 100/92. In the firstscenario, we have assumed that each of thefunctions CRRM and CRSP increases the RFpower consumption by 2.5W after processing100 RBs. Hence the power consumption of FDis increasing from the value 12.9 to about 17.9.The same amount of power has been reducedfrom the CBB pool where previously thesefunctions reside. Figure 1 shows the powerconsumption comparison of fog network withtraditional cloud when processing 100 RBs fordifferent number of CBBs, RUs and FDs. ThisFigure indicates that the more RUs or FDs areincluded in the comparison, the more degradation occurs in the power performance (morepower consumption) of fog network. This happened because when increasing the number ofRUs or FDs, the number of fog functions thatexecution time of the device without usingvirtualisation, whereτ τdevice = + int ( * ) mod RB , (25)where τ int represents the initial device delaydue to other functions rather than MCS, thelatter is denoted by the constant factor (mod).Due to virtualisation cost, τ device itself is driven up by a value eΩ×RB that is responsible toincrease the value of τdevice by 1.7, i.e.,τ τdevice device 2 = × eΩ×RB, (26)where Ω indicates the increasing constant.Hence, the total delay of all functions is formulated as τ τ 1 = , 1FD VM∑∑fd vm 1vm fd2,(27) 1devicewhere τdevicevm fd , 12 denotes the execution time offunction vm within the tagged FD fd1. It is worthmentioning that this modelling is only valid forthe first type of FDs, where the functions are installed on the same board with the RF unit. If fogfunctions are found on separate board/device,there will be another formulation asτ τ τ 2 2 = +FD∑fd 22device brdfd, (28)where τbrd indicates the board delay, which isalso traffic based delay, i.e.τ τbrd brd =int RB *eω× , (29)where ω denotes the increasing constant.On the other side, the delay drop fromCBBs, whose functions are moved, is represented by τ drop = × RB. Hence, the total delayat of the first scenario is equivalent to τ τ τ11 1= – drop.(30)This logic holds true for second scenario,where the total delay is indicated byτ τ τ22 2= – drop(31) for second scenario.VI. RESULTS AND ANALYSISTo produce fair results, the DC and AC consumption for all fog scenarios are assumedidentical. For example, the DC-DC consumption factor in Table 1, is equivalent to 0.92, thisratio can be found in the denominator of theTable I. Model parameters.Factor Traditional(Generated) Unitγ dc 0.9250 –γ mc 0.910 –γ cc 0.90 –κ 0.002 –λ 0.002 –mod 0.014 –ω 0.01 –Ω 0.0021 – 0.0023 –ζ 0.0179 –π 0.0179 –Ant 1 –ind 1.3 –α 3 –τint 50 µsecτbrdint 20 µsecτrlvd 5 µsecPrf 12.9 WPamp 29.6528 WPCB 29.4 WPaddint 8 WAuthorized licensed use limited to: Macquarie University. Downloaded on October 06,2021 at 00:51:48 UTC from IEEE Xplore. Restrictions apply.74 China Communications • April 2020CRRM and CRSP functions is changed whilekeeping the same behaviour.In the second scenario, it was assumed thatthe initial power consumption of the separateboard is 8W, hence, the amount eζ ×RB increasesthis initial by about 6W, which adds up to 14Wwhen processing the maximum number of RBs.Subsequently, Figure 2 shows the amountof power consumed for traditional cloud andfog networks. It is clear that this scenario offog has produced more power consumptionthan first scenario as the added board consumes more power than the virtualised RF.In addition, the third scenario is the casewhere fog network relies on separate devicesdeployed within the coverage of RUs. This casehas also produced a noticeable amount of power consumption in comparison to traditionalcloud, as shown in Figure 3. Although the maximum deployed FDs in this Figure is 40, thepower consumption is still prominent. The reason of such increasing is two fold: first the FDsare now consume the power separately thanthe RUs, which requires more power to operatethem. Second, once the FDs are separated, theyare no longer share the DC and AC sourceswith other units, in contrary to the previous scenarios. This status adds extra/separated powerconsumption to the network due to requiringadditional AC and DC power conversions.Subsequently, the EE analysis is shown inFigure 4. Due to the fact that cloud networkhas produced less power consumption thanfog network, the former has scored higher EE(bit/sec/W) than the other three fog scenarios. Moreover, to analyse the delay, we haveassumed only the first two scenarios of fognetwork, where the fog functions are virtualised or run on the separate board within theFD. The third scenario is assumed identicalto the second, where the delay of the separateFD is similar to the separate board. First wepresented the delay that is inherited from MCSprocess. In the first scenario, it was assumedthat virtualising the functions increases thisdelay by 1.7, thanks to the advancement andsmart virtual machines hyperVisors (HV). Theare required to be deployed within the FDs isincreased too. Such matter urges the power consumption of fog network to increase while thenumber of CBBs are fixed. More explanationcan be found in Section IV, specifically in (2).However, the values of Figure 1 can be slightly changed if the amount of power added byFig. 1. Power consumption comparison of cloud and fog networks of the first scenario.Fig. 2. Power consumption comparison of cloud and fog networks of the secondscenario.15 20CBBs countRUs count 50 00 5 101001506000800010000020004000Consumed power (W)Cloud (40 RUs)Cloud (80 RUs)Cloud (150 RUs)Fog (40 FDs)Fog (80 FDs)Fog (150 FDs)15 20CBBs count50 00 5 10RUs count100150200060000800010000120004000Consumed power (W)Cloud (40 RUs)Cloud (80 RUs)Cloud (150 RUs)Fog (40 FDs)Fog (80 FDs)Fog (150 FDs)Authorized licensed use limited to: Macquarie University. Downloaded on October 06,2021 at 00:51:48 UTC from IEEE Xplore. Restrictions apply.China Communications • April 2020 75within the RU, finally, the functions are installedon a separate device. Although the presentedmodel was converged to specific data. However,it can be generalised to fulfil different deviceswith various manufacturing properties and specifications. This research showed that the fog network cost regarding the power consumption islatter is responsible for managing fog functions/VMs in a timely limited manner, and ensures that these VMs do not disrupt each otherwhile sharing the device’s resources. Hence,the more advanced HV, the less time the VMspends to process its packet.Compared to the second scenario, the separate board produces less delay when processingthe RBs because the delay cost of virtualisationis higher than when the VMs installed separately on a separate device. In the latter, due toless sharing of the FD’s resources, less delay isproduced, as shown in Figure 5. The RBs basedFD initial delay of the first scenario starts from80 µsec, till it reaches about 200 µsec. Thevirtualisation process increases the initial of FDby 1.7 when two VMs are running at the sametime. However these values can be changedaccording to the different data that are acquiredfrom various devices, yet the model accommodates such variation. In the second scenario, theinitial delay of the board is assumed much lessthan the first case, this delay is added to the initial to construct the board and original RF unitdelay within the same FD. It is worth mentioning that, since the values of link delay are muchless than the effect of the processing delay inthe FD itself. A critical fog mode (type of fognetwork) selection is a vital to lessen the effectof this metric up on the network performance.However, the total delay, shown in Figure6, was generated by combining the channeldelay and the above mentioned processingdelays. It is worth mentioning that the link delay of fog network is the average delays of allUEs to the FDs, while in traditional cloud, thechannel delay comprises both UEs-RUs andRUs-CBB pool. This eventually advocates fogagainst the legacy cloud.VII. CONCLUSION AND FUTURE WORKThis paper discussed and compared the case of5G fog networks with traditional cloud accessnetworks. Three scenarios are tackled, first, thefog functions are virtualised within RU unit, second, the functions reside on a separate board, yetFig. 3. Power consumption comparison of cloud and fog networks of the third scenarioFig. 4. EE analysis of cloud, first, second and third scenarios, with 20 CBBs, 40RUs, and 100 processed RBs.15 20Number of CBBs50 00 5 10Number of RUs150 100200060000800010000120004000Consumed power (W)Cloud (40 RUs)Cloud (80 RUs)Cloud (150 RUs)Fog (40 RUs)Fog (80 RUs)Fog (150 RUs)4030Number of RUs2010100050Processed RBs01.211.40.80.20.41.61.800.6EE/HzCloud EEFog EE,scenario1Fog EE,scenario2Fog EE,scenario3Authorized licensed use limited to: Macquarie University. Downloaded on October 06,2021 at 00:51:48 UTC from IEEE Xplore. Restrictions apply.76 China Communications • April 2020consumption. Hence, a hybrid and adaptabledesign of both legacy cloud and fog is worth toexamine where a decision can be made priorserving the UEs. This is to decide whether theUEs are requesting on-line or off-line services.Accordingly, the UE’s packets are directed to theCBB pool if the request is off-line, or to the FDsif the request is on-line. In this case, some of thepower cost will be dropped as the fog networkcan be shifted to a pure cloud whose power consumption is less in some periods of time. Nevertheless, the process of making the decision is aping-pong communications amongst UEs, RUs,FDs, and the pool, this situation adds anotherpower overhead in the backhaul and fronthauldue to transmitting/receiving processes. Thisissue imposes more speculations to evaluate thetotal cost of such paradigm. earch opens another discussion about the case of transferring fogfunctions on demand once the decision is made,using live migration techniques. This method entails calculating the cost of migration, as migrating a single virtual machine costs about 10W inthe receiving side, while the cost of the transmitting side varies according to the speed of migration. This matter requires a holistic comparisonwith the proposed model, whose functions arestatic within the FDs.ACKNOWLEDGEMENTThis work was supported by University ofDiyala, college of Engineering, department ofcommunications, Diyala, Iraq.References[1] K. Liang, L. Zhao, X. Zhao, Y. Wang, and S. Ou,“Joint resource allocation and coordinatedcomputation offloading for fog radio accessnetworks,” China Communications, vol. 13, pp.131–139, N 2016.[2] J. Liu, B. Bai, J. Zhang, and K. B. Letaief, “Cacheplacement in fog- rans: From centralized to distributed algorithms,” IEEE Transactions on Wireless Communications, vol. 16, pp. 7039–7051,Nov 2017.[3] S. Bhaumik, S. P. Chandrabose, M. K. Jataprolu,G. Kumar, A. Mu- ralidhar, P. Polakos, V. Srinivasan, and T. Woo, “Cloudiq: a framework forprocessing base stations in a data center,” inProceedings of the 18th annual internationala noteworthy. This issue however, degrades theenergy efficiency of these designs while reducing the delay. The latter is a paramount factor toconsider when the network is serving a real timecommunications. Accordingly, this research advocates against using fog network when servingoff-line applications, as that costs an extra powerFig. 5. Delay analysis of first and second scenarios while processing differentnumber of RBs.Fig. 6. Total delay analysis of first, second, cloud wireless and cloud optical channel while processing different number of RBs.Resource Blocks0 10 20 30 40 50 60 70 80 90 100Time delay (µsec)050100150200250300350Initial delayVM delay, fisrt scenarioBoard delayInitial+board, second scenarioResource Blocks0 10 20 30 40 50 60 70 80 90 100Time delay (µsec)50100150200250300350400450τ11τ22C-RAN,opticalC-RAN,wirelessAuthorized licensed use limited to: Macquarie University. Downloaded on October 06,2021 at 00:51:48 UTC from IEEE Xplore. Restrictions apply.China Communications • April 2020 77al Conference on Ubiquitous Wireless Broadband(ICUWB), pp. 1–4, Oct 2016.[17] S. Jia, Y. Ai, Z. Zhao, M. Peng, and C. Hu, “Hierarchical content caching in fog radio access networks:ergodic rate and transmit latency,” China Communications, vol. 13, pp. 1–14, December 2016.[18] P. K. Sharma, M. Chen, and J. H. Park, “A software defined fog node based distributed blockchain cloud architecture for IoT,” IEEE Access,vol. 6, pp. 115–124, 2018.[19] E. Baccarelli, P. G. V. Naranjo, M. Scarpiniti, M.Shojafar, and J. H. Abawajy, “Fog of everything:Energy-efficient networked computing architectures, research challenges, and a case study,”IEEE Access, vol. 5, pp. 9882–9910, 2017.[20] H. Yang, W. Bai, A. Yu, Q. Yao, J. Zhang, Y. Lin,and Y. Lee, “Band- width compression protection against collapse in fog-based wireless andoptical networks,” IEEE Access, vol. 6, pp. 54760–54768, 2018.[21] K. Liang, L. Zhao, X. Chu, and H. Chen, “Anintegrated architecture for software definedand virtualized radio access networks with fogcomputing,” IEEE Network, vol. 31, pp. 80–87,January 2017.[22] R. S. Alhumaima, M. Khan, and H. S. Al-Raweshidy,“Component and parameterised power model forcloud radio access network,” IET Communications,vol. 10, no. 7, pp. 745–752, 2016.[23] R. S. Alhumaima and H. S. Al-Raweshidy, “Modelling the power consumption and trade-offs ofvirtualised cloud radio access networks,” IET Communications, vol. 11, no. 7, pp. 1158–1164, 2017.[24] R. S. Alhumaima and H. S. Al-Raweshidy, “Evaluating the energy efficiency of software defined-based cloud radio access networks,” IETCommunications, vol. 10, no. 8, pp. 987–994, 2016.[25] R. Shea, H. Wang, and J. Liu, “Power consumption of virtual machines with network transactions: Measurement and improvements,” inINFO- COM, 2014 Proceedings IEEE, pp. 1051–1059, IEEE, 2014.BiographyRaad S. Alhumaima, receivedhis B.E. degree in Communications Engineering from theUniversity of Diyala in 2003, hisM.Sc. degree in Laser/ Electronics and Communicationfrom the University of Baghdadin 2011, and his PhD in Communications and Electronics from Brunel UniversityLondon-UK in 2017. He has Published several IET,IEEE and conference papers. He is currently a lecturerin Diyala University,College of Engineering, Department of Communications, Diyala, Iraq. His current research interests cover next generation networks, including SDN, C-RAN, NFV, Fog-RAN and Quantummobile communication.conference on Mobile computing and networking, pp. 125–136, ACM, 2012.[4] S. Hung, H. Hsu, S. Lien, and K. Chen, “Architecture harmonization between cloud radio accessnetworks and fog networks,” IEEE Access, vol. 3,pp. 3019–3034, 2015.[5] K. Liang, L. Zhao, X. Zhao, Y. Wang, and S. Ou,“Joint resource allocation and coordinatedcomputation offloading for fog radio accessnetworks,” China Communications, vol. 13, pp.131–139, N 2016.[6] J. Kang, O. Simeone, J. Kang, and S. S. Shitz,“Control-data separation with decentralizededge control in fog-assisted uplink communications,” IEEE Transactions on Wireless Communications, vol. 17, pp. 3686– 3696, June 2018.[7] M. Peng and K. Zhang, “Recent advances in fogradio access networks: Performance analysisand radio resource allocation,” IEEE Access, vol.4, pp. 5003–5009, 2016.[8] Y. Ku, D. Lin, C. Lee, P. Hsieh, H. Wei, C. Chou,and A. Pang, “5G radio access network designwith the fog paradigm: Confluence of communications and computing,” IEEE CommunicationsMagazine, vol. 55, pp. 46–52, April 2017.[9] M. Peng, S. Yan, K. Zhang, and C. Wang,“Fog-computing-based radio access networks:issues and challenges,” IEEE Network, vol. 30,pp. 46– 53, July 2016.[10] Y. Yu, T. Chiu, A. Pang, M. Chen, and J. Liu, “Virtualmachine place- ment for backhaul traffic minimization in fog radio access networks,” in 2017IEEE International Conference on Communications (ICC), pp. 1– 7, May 2017.[11] T. C. Chiu, A. C. Pang, W. H. Chung, and J. Zhang,“Latency-driven fog cooperation approach infog radio access networks,” IEEE Transactions onServices Computing, pp. 1–1, 2018.[12] G. M. S. Rahman, M. Peng, K. Zhang, and S.Chen, “Radio resource allocation for achievingultra-low latency in fog radio access networks,”IEEE Access, vol. 6, pp. 17442–17454, 2018.[13] R. Deng, R. Lu, C. Lai, and T. H. Luan, “Towardspower consumption- delay tradeoff by workload allocation in cloud-fog computing,” inCom- munications (ICC), 2015 IEEE InternationalConference on, pp. 3909– 3914, IEEE, 2015.[14] H. Zhang, Y. Qiu, K. Long, G. K. Karagiannidis, X.Wang, and A. Nallanathan, “Resource allocationin NOMA-based fog radio access networks,”IEEE Wireless Communications, vol. 25, pp.110–115, JUNE 2018.[15] H. Zhang, Y. Qiu, X. Chu, K. Long, and V. C. M.Leung, “Fog radio access networks: Mobilitymanagement, interference mitigation, and resource optimization,” IEEE Wireless Communications, vol. 24, pp. 120–127, Dec 2017.[16] C. Zhang, Y. Sun, Y. Mo, Y. Zhang, and S. Bu,“Social-aware content downloading for fogradio access networks supported device-to-device communications,” in 2016 IEEE InternationAuthorized licensed use limited to: Macquarie University. Downloaded on October 06,2021 at 00:51:48 UTC from IEEE Xplore. Restrictions apply.
