Joint Transmission Scheme and Coded Content Placement in Cluster-centric UAV-aided Cellular Networks
Zohreh HajiAkhondi-Meybodi, Arash Mohammadi, Jamshid Abouei, Ming Hou
11 Joint Transmission Scheme and Coded ContentPlacement in Cluster-centric UAV-aided CellularNetworks
Zohreh HajiAkhondi-Meybodi,
Student Member, IEEE , Arash Mohammadi,
Senior Member, IEEE , JamshidAbouei,
Senior Member, IEEE , Ming Hou,
Senior Member, IEEE
Abstract —Recently, as a consequence of the COVID-19 pan-demic, dependence on telecommunication for remote learn-ing/working and telemedicine has significantly increased. Inthis context, preserving high Quality of Service (QoS) andmaintaining low latency communication are of paramount im-portance. Development of an Unmanned Aerial Vehicles (UAV)-aided heterogeneous cellular network is a promising solution tosatisfy the aforementioned requirements. Although an integratedUAV-aided and cluster-centric cellular network can enhanceconnectivity and provide improved QoS, there are key challengesahead for its efficient implementation. On the one hand, it ischallenging to optimally increase content diversity in cachingnodes to mitigate the network’s traffic over the backhaul. On theother hand is the challenge of attenuated UAVs’ signal in indoorenvironments, which increases users’ access delay and UAVs’energy consumption. To address the aforementioned challenges,we incorporate UAVs, as mobile caching nodes, together withFemto Access points (FAPs) to increase the network’s coveragein both indoor and outdoor environments. Referred to as theCluster-centric and Coded UAV-aided Femtocaching (CCUF)framework, a two-phase clustering framework is proposed foroptimal FAPs’ formation and UAVs’ deployment. The proposedCCUF implementation leads to an increase in the cache di-versity, a reduction in the users’ access delay, especially inindoor environments, and significant reduction in UAVs’ energyconsumption. To mitigate the inter-cell interference in edge areas,the Coordinated Multi-Point (CoMP) approach is integratedwithin the CCUF framework. In contrary to existing works,we analytically compute the optimal number of FAPs in eachcluster to increase the cache-hit probability of coded contentplacement. Furthermore, the optimal number of coded contentsto be stored in each caching node is computed to increase thecache-hit-ratio, Signal-to-Interference-plus-Noise Ratio (SINR),and cache diversity and decrease the users’ access delay andcache redundancy for different content popularity profiles.
Index Terms —Cluster-centric, Coded Femtocaching, Coordi-nated Multi-Point (CoMP), Unmanned Aerial Vehicles (UAVs),Users’ Access Delay.
Z. HajiAkhondi-Meybodi is with Electrical and ComputerEngineering (ECE), Concordia University, Montreal, Canada. (E-mail:z [email protected]). A. Mohammadi (corresponding author) is withConcordia Institute of Information Systems Engineering (CIISE), ConcordiaUniversity, Montreal, Canada. (P: +1 (514) 848-2424 ext. 2712 F: +1(514) 848-3171, E-mail: [email protected]). J. Abouei waswith the Department of Electrical, Computer and Biomedical Engineering,Ryerson University, Toronto, ON M5B 2K3, Canada. He is now with theDepartment of Electrical Engineering, Yazd University, Yazd 89195-741,Iran (E-mail: [email protected]). M. Hou is with Defence Research andDevelopment Canada (DRDC), Ottawa, Toronto, ON, M2K 3C9, Canada.(E-mail: [email protected]).This Project was partially supported by the Department of National De-fence’s Innovation for Defence Excellence and Security (IDEaS) program,Canada.
I. I
NTRODUCTION
As a consequence of the COVID-19 pandemic, dependenceon telemedicine and remote learning/working has significantlyincreased due to the exponential rise in the demand for in-home care, remote working, schooling and remote report-ing [1]. Within this context, to preserve real-time qualityand integrity of large-sized multimedia data and provide highQuality of Service (QoS), it is of paramount importance todevelop advanced heterogeneous networking solutions [2]. Inthis regard, caching has emerged as a promising solutionto maintain low latency communication and mitigate thenetwork’s traffic over the backhaul. This, in turn, improvesthe QoS by storing the most popular multimedia contents closeto the end-users [3], [4]. Recently, Unmanned Aerial Vehicle(UAV)-based caching has gained significant attention fromboth industry and academia [5]–[8], due to its high mobility,low cost and easy deployment characteristics. Although theenhanced connectivity that comes by using UAVs will improvethe QoS in outdoor areas, there are key challenges aheadin indoor environments due to attenuation of the receivedsignal [9]. To address this issue and in line with advancementsof 5G networks, the paper focuses on coupling UAVs asaerial caching nodes with Femto Access Points (FAPs) [10],equipped with storage. The ultimate goal is to increase thecaching network’s QoS and its coverage in a heterogeneousenvironment (i.e., integrated indoor and outdoor settings).
Literature Review:
There are several challenges in devel-opment of UAV-aided caching networks, including place-ment optimization [11], [12]; trajectory design [13], [14];resource allocation [15], [16], and; energy consumption man-agement [17], [18]. Existing works [11]–[22] mainly focus onthe deployment of UAVs in outdoor environments. There areseveral benefits that come with utilization of UAVs as aerialcaching nodes such as high probability of establishing Line-of-Sight (LoS) links between UAVs and ground users. Suchbenefits, however, result in several critical challenges espe-cially in indoor environments due to signal attenuation [23].This level of obfuscation in an integrated indoor and outdoorenvironment can be settled by design of UAV-aided cellularnetworks [24].One of the main challenges in UAV-aided cellular networksis to optimally assign caching nodes (UAVs or FAPs) to groundusers to efficiently serve their requests [25], [26]. There areseveral QoS and Quality of Experience (QoE) metrics that a r X i v : . [ c s . I T ] J a n can be considered as the decision criteria for Access Point(AP) selection, including users’ latency; traffic load and energyconsumption of APs; users’ link quality; handover rate, and;Signal-to-Interference-plus-Noise Ratio (SINR). For instance,Athukoralage et al. [25] considered an AP selection frame-work, where ground users are supported by UAV or WiFi APs.In this work, the users’ link quality is utilized to balance theload between UAVs and WiFi APs. Zhu et al. [26] proposeda game theory-based AP selection scheme, where probabilityof packet collision is used to select the optimal AP among allpossible UAVs and Base Stations (BSs). Considering the factthat users’ mobility is one of the inherent features of currentultra-dense wireless networks, movement characteristics, suchas speed of ground users, must be considered as a decisioncriteria for connection scheduling. To date, limited researchhas been performed on UAV-aided cellular networks to addressthe aforementioned problem of connection scheduling betweenFAPs and UAVs in a heterogeneous environment consideringthe speed of users’ movement. The paper addresses this gap.In UAV-aided cellular networks, the main objective is tobring multimedia data closer to ground users and simultane-ously improve users’ QoS and network’s QoE. If the requestedcontent can be found in the storage of one of the availablecaching nodes, this request would be served directly andcache-hit occurs; otherwise, it is known as a cache-miss. Dueto the large size of multimedia contents, however, it is notfeasible to store all contents in the storage of caching nodes.To overcome this problem, coded caching strategies, such asLinear Random Fountain Code (LRFC) [27], and MaximumDistance Separable (MDS) [28] coding schemes have receivedremarkable attention lately. In such coded caching strategies,only specific segments of the most popular multimedia con-tents are stored in the caching nodes. Storing coded popularcontents increases the overall content diversity. In coded fem-tocaching frameworks, however, most early works such as Ref-erence [29] consider homogeneous networks, where the samesegments of the most popular multimedia contents are storedin different caching nodes. To overcome this issue, the mainfocus of recent researchers has been shifted to heterogeneous networks with an emphasis on the mobility of ground usersand overlapped caching nodes [30]. Toward this goal, Chen etal. [31] proposed a cluster-centric small cell network, wherepopular contents are classified into two different categories,i.e., the Most Popular Contents (MPCs), and the Large PopularContents (LPCs). While MPCs are completely stored in thestorage of all Small Base Stations (SBSs), distinct segmentsof LPCs are cached in different SBSs.Cluster-centric cellular networks provide several benefits,such as increased content/cache diversity, which in turn leadsto remarkable growth in the number of requests managedby the caching nodes. However, this comes with the costof experiencing inter-cell interference especially for cell-edgeusers. To tackle this issue, Chen et al. [31] employed theCoordinated Multi-Point (CoMP) approach to mitigate theinter-cell interference and improve the throughput of groundusers located in the cell-edges. Reference [31] developedtwo transmission schemes, namely Joint Transmission (JT)and Parallel Transmission (PT), which are selected based on the popularity of the requested content. Alternatively, Lin etal. [32] proposed a cluster-centric cellular network applyingthe CoMP technique based on the users’ location, where cell-core and cell-edge users are served through Single Transmis-sion (ST) and JT, respectively. Despite all the researches onthe cluster-centric cellular networks, there is no framework todetermine how different segments can be cached to increasethe data availability in a UAV-aided cluster-centric cellularnetwork integrated by the CoMP technology. The paper alsoaddresses this gap. Contribution:
In this paper, we consider an integrated UAV-aided and cluster-centric cellular network to serve groundusers positioned in both indoor and outdoor environments. Ourfirst objective is to increase the content diversity that can beaccessed via the caching nodes. The second goal is to assignground users the best caching nodes to improve the achievableQoS in terms of the users’ access delay and decrease theenergy consumption of UAVs. To achieve the above-mentionedobjectives, the paper proposes a novel Cluster-centric andCoded UAV-aided Femtocaching (CCUF) framework, in whichthe caching node and the transmission scheme are jointly de-termined based on the movement features of ground users andthe CoMP technology. Moreover, to increase content diversity,an analytical solution is derived to allocate distinct segmentsof the multimedia contents in different neighboring FAPs. Insummary, the paper makes the following key contributions: • Indoor penetration loss and deep shadow fading caused bybuilding walls significantly attenuate the UAV’s signalsin indoor environments degradation the network’s QoS.To tackle this issue, we consider two different indoor andoutdoor caching service scenarios for the proposed CCUFframework. More precisely, the indoor area is covered byFAPs, equipped with extra storage. The outdoor, however,is supported by coupled UAVs and FAPs depending onthe movement speed of ground users. • Due to the limited storage of caching nodes, the mostpopular multimedia contents are determined based onthe solution of a formulated optimization problem. Con-sequently, multimedia contents are fragmented into thesame size segments based on the Fountain codes. • With the focus on a dynamic UAV-based femtocachingnetwork, storing distinct contents in neighboring FAPsincreases the resource availability. In contrary to existingcoded caching approaches [31], [32] that lack utilizationof placement strategies to allocate segments of videocontents to different caching nodes, the proposed CCUFscheme determines a solution to store different segmentsof contents in neighboring FAPs. Toward this goal,we consider a cluster-centric cellular network, wheremultimedia contents are classified into three categories,including popular, mediocre, and non-popular contents.While the popular contents are stored completely, distinctsegments of mediocre ones are determined according tothe proposed framework to be stored in the storage ofneighboring FAPs. In this paper, we determine the bestnumber of coded and uncoded contents in each cachingnode to increase the cache-hit-ratio, SINR, and cache diversity while decreasing users’ access delay and cacheredundancy for different content popularity profiles. • To simulate a real wireless network, we consider mo-bile ground users, where the Angle of Arrival (AoA)localization technique is utilized to estimate the initiallocation of ground users. Then, the mobility pattern ispresented by the Difference Correlated Random Walk(DCRW) model. To access a large amount of contentsduring movement of ground users, two scenarios areproposed: ( i ) The whole network is partitioned into sub-networks called inter-clusters. All FAPs in the same inter-cluster save different parts of mediocre contents, whilethe cached contents of different inter-clusters are thesame. In this case, it can be shown that ground userscan acquire more segments during their movements, and; ( ii ) To increase the resource availability, the outdoorenvironment is partitioned into intra-clusters via a K -means clustering algorithm, each covered by a UAV. • To mitigate the inter-cell interference in inter-clusters andprovide more efficient service for edge users, the CoMPtechnique is utilized, including ST and JT schemes. Notethat the transmission type depends on both the popularityof contents and the position of the ground user in the cell.The effectiveness of the proposed CCUF framework is eval-uated through simulation studies in both indoor and outdoorenvironments in terms of cache-hit-ratio, users’ access delay,SINR, cache diversity, cache redundancy, and energy con-sumption of UAVs. According to the simulation results, weinvestigate the best number of coded contents in each cachingnode to increase the cache-hit-ratio, SINR, and cache diversityand decrease users’ access delay and cache redundancy fordifferent content popularity profiles. Moreover, we investigatethe effects of the UAV-aided femtocaching network on theusers’ access delay and energy consumption of UAVs in bothindoor and outdoor environments.The remainder of the paper is organized as follows: InSection II, the network’s model is described and the mainassumptions required for the implementation of the proposedframework are introduced. Section III presents the proposedCCUF scheme. Simulation results are presented in Section IV.Finally, Section V concludes the paper.II. S
YSTEM M ODEL AND P ROBLEM D ESCRIPTION
We consider a UAV-aided cellular network in a residentialarea that supports both indoor and outdoor environments.There exist N f number of FAPs, denoted by f i , for ( ≤ i ≤ N f ), each with the cache size of C f and transmissionrange of R f . There are also N u number of UAVs, denotedby u k , for ( ≤ k ≤ N u ), with equal transmission range of R u . The transmission range of each FAP is significantly lessthan that of a UAV. As it can be seen from Fig. 1, N s < N f number of FAPs in a neighborhood form a cluster, referredto as the inter-cluster. Similarly, N ic number of FAPs in theoutdoor environment form an intra-cluster covered by a UAV.There are N g number of ground users, denoted by GU j , for( ≤ j ≤ N g ), that move through the network with differentvelocities. υ j ( t ) denotes the speed of the ground user GU j at Fig. 1: A typical structure of the proposed UAV-aided cellular network. time slot t . In this work, FAP f i , for ( ≤ i ≤ N f ), operates inan open access mode, i.e., it can serve any ground user GU j ,for ( ≤ j ≤ N g ), located in its transmission range. The FAPswithin each inter-cluster use the CoMP transmission approach(supporting ST and JT schemes) to mitigate the inter-cellinterference in edge areas and manage ground users’ requests.To completely download a requested content, a finite time T isrequired. For ease of exposition, the time T is discretized into N s time slots with time interval δ t , i.e., T = N s δ t . In whatfollows, we present users’ mobility pattern, the content pop-ularity profile, and transmission schemes utilized to developthe proposed CCUF. A. Users’ Mobility Pattern
In the proposed CCUF framework, estimated location ofGUs is required to determine the transmission scheme andselect an appropriate caching node to manage the request.Capitalizing on the reliability and efficiency of the AoAlocalization technique [33], [34], it is utilized to identify theinitial location of GUs as follows x j (0) = d n,i tan θ i,j tan θ i,j − tan θ n,j , (1) y j (0) = d n,i tan θ n,j tan θ i,j tan θ i,j − tan θ n,j , (2)where l j (0) = [ x j (0) , y j (0)] T is the location of ground user GU j at time t = 0 . In addition, d n,i indicates the distancebetween the i th and the n th FAPs, assuming the positions ofFAPs are known. Moreover, the angle between the line fromthe location of the ground user GU j to FAP f i and X -axis isdenoted by θ i,j [34].Given the initial location of ground users, we use theDifference Correlated Random Walk (DCRW) [35] to model Fig. 2:
Zipf distribution. their movement patterns. In this regard, the location of GU j at time slot t , denoted by l j ( t ) , is given by l j ( t ) = l j ( t −
1) + υ j ( t − t, (3)where ∆ t is the time interval between two consecutive esti-mated locations, and υ j ( t ) = [ υ ( x ) j ( t ) , υ ( y ) j ( t )] T denotes theuser’s velocity, obtained as follows [35] d υ j ( t ) = − (cid:18) − log ς θ − θ − log ς (cid:19) ( υ j ( t ) − µ ) dt + J d b t , (4)where ς and ς denote auto-correlation parameters in X -axisand Y -axis, respectively. Terms θ and µ represent the meanturning angle and the mean velocity vector, respectively. Term J represents the velocity shifts covariance, which is a (2 × lower triangular matrix with positive diagonal components.Finally, b t , which is a (2 × vector, determines the standardBrownian motion at time slot t . B. Content Popularity Profile
When the ground user GU j requests content c l from alibrary of C = { c , . . . , c N c } , in which N c is the cardinality ofmultimedia data in the network, this request should be handledby one of the nearest FAPs or UAVs having some segmentsof c l . Regarding the user’s behavior pattern in multimedia ser-vices, the popularity of video contents is determined based onthe Zipf distribution [36], where the probability of requesting l th file, denoted by p l , is calculated as p l = l − γN c (cid:80) r =1 r − γ , (5)where γ represents the skewness of the file popularity. In theproposed CCUF framework, each content c l is fragmented into N s encoded segments, denoted by c ls , for ( ≤ s ≤ N s ),which is the same as the number of FAPs in each inter-cluster.The ground user GU j can download one segment in δ t , whichis equivalent to /N s of a multimedia content. In other words,in each contact, user GU j will leave the transmission area ofthe current FAP. For notational convenience, we assume that P [ n ] ≡ P ( nδ t ) denotes the probability of accessing a newsegment in time slot n , with n ∈ { , . . . , N s } .Without loss of generality and to be practical, we investigatethe probability distribution of a real multimedia data set, i.e.,the YouTube videos trending statistics, shown in Fig. 2. As itcan be seen from Fig. 2, the probability distribution followsZipf, i.e., a small part of the contents are requested with ahigh probability. The majority of contents are not popular,and some contents, are requested moderately. Consequently,in our cluster-centric UAV-aided cellular network, we classifymultimedia contents into three categories, including popular,mediocre, and non-popular [31]. To improve content diversity,the storage capacity of FAPs, denoted by C f , is divided intotwo spaces, where α portion of the storage is allocated tostore complete popular contents, i.e., ≤ l ≤ (cid:98) αC f (cid:99) , where l = 1 indicates the most popular content. Additionally, (1 − α ) portion of the cache is assigned to store different parts of themediocre contents, where (cid:98) αC f (cid:99) + 1 ≤ l ≤ N s ( C f − (cid:98) αC f (cid:99) ) .The optimal value of α is obtained experimentally. The pro-posed model for identifying different segments to be cachedin neighboring FAPs will be discussed later on in Section III. C. Transmission Scheme
In this subsection, we will describe both the connectionscheduling (serving by FAPs or UAVs) and the transmissionscheme depending on the presence of the ground user in indooror outdoor environments.
1) Indoor Environment:
The transmitted signal by UAVs,propagating in residential areas, becomes weaker due to thepenetration loss and shadow fading effects. Therefore, it isassumed that ground users positioned in indoor areas are onlysupported by FAPs. In the CoMP-integrated and cluster-centriccellular network and as it can be seen from Fig. 1, there are tworegions in each inter-cluster, named cell-edge and cell-core,which are determined based on the SINR value to illustratethe quality of a wireless link. In such a case that the grounduser GU j is positioned in the vicinity of the FAP f i , the SINRfrom f i to GU j , denoted by S i,j , is obtained as follows S i,j ( t ) = P i | ˜ H i,j ( t ) | I f − i ( t ) + N , (6)where P i denotes the transmitted signal power of FAP f i . I f − i ( t ) and N represent the interference from other FAP-ground users, except for the corresponding f i link, and thenoise power related to the additive white Gaussian randomvariable, respectively. Moreover, the path loss and fadingchannel effect between FAP f i and ground user GU j at timeslot t is denoted by ˜ H i,j ( t ) = h i,j ( t ) (cid:112) L i,j ( t ) . In this case, h i,j ( t ) denotes a complex zero-mean Gaussian random variable withunit standard deviation and L i,j ( t ) represents the path lossbetween FAP f i and ground user GU j at time slot t , obtainedas follows L i,j ( t ) = L + 10 η log (cid:0) d i,j ( t ) (cid:1) + χ σ , (7)where η is the path loss exponent. Term χ σ indicates the shad-owing effect, which is a zero-mean Gaussian-distributed ran-dom variable with standard deviation σ . Additionally, d k,j ( t ) represents the Euclidean distance between FAP f i and grounduser GU j at time slot t . Furthermore, L = 20 log (cid:18) πf c d c (cid:19) is the path loss related to the reference distance d where f c and c = 3 × denote the carrier frequency and thelight speed, respectively. Accordingly, the ground user GU j islocated at the cell-core of FAP f i , if S i,j ( t ) > S th ; otherwise, GU j is located at the cell-edge, where S th is the SINRthreshold.The transmission scheme in the proposed CoMP-integratedand cluster-centric cellular network is determined based on twometrics; ( i ) The popularity of the requested content, describedin Subsection II-B, and; ( ii ) The location of the ground userin the cell, i.e., cell-core or cell-edge. The following twodifferent transmission schemes are utilized for developmentof the proposed CCUF framework: • Single Transmission (ST):
In this case, the requestedfile c l , for (1 ≤ l ≤ (cid:98) αC f (cid:99) ) , is popular content, and theground user GU j is located at the cell-core of FAP f i ,i.e., S i,j ( t ) > S th . It means that the content is completelycached into the storage of FAP f i and the high-qualitylink can be established between FAP f i and ground user GU j . Consequently, this request is served only by thecorresponding FAP f i . Moreover, if the requested content,belongs to the mediocre category, i.e., (cid:98) αC f (cid:99) + 1 ≤ l ≤ N s ( C f −(cid:98) αC f (cid:99) ) , regardless of the location of the grounduser, this request served according to the ST scheme. • Joint Transmission (JT):
In this transmission scheme,the requested file c l is popular content, i.e., ≤ l ≤(cid:98) αC f (cid:99) . Consequently, all FAPs have the same completefile. The ground user GU j , however, is located at thecell-edge of FAP f i , i.e., S i,j ( t ) ≤ S th . Therefore, thelink quality between FAP f i and the ground user GU j isnot good enough. In order to improve the reliability ofcontent delivery, the corresponding content will be jointlytransmitted by several FAPs in its inter-cluster. As it canbe seen from Fig. 1, neighboring FAPs in an inter-clustercollaboratively serve cell-edge ground users based on theJT scheme, which is shown by the red color.
2) Outdoor Environment:
The wide transmission rangeof UAVs and the high probability of establishing LoS linkprovide several advantages, including the ability to manage themajority of ground users’ requests, which leads to improvedcoverage in outdoor environments. Due to the limited batterylife of UAVs, however, requests that are handled by UAVsshould be controlled. For this reason, we consider a UAV-aidedfemtocaching network in the outdoor environment. Note thatwe also need to reduce the number of handovers, which canbe frequently triggered by FAPs, if the ground user moves andleaves the current position rapidly. Toward this goal, groundusers are classified based on their velocity into the followingtwo groups: • Low Speed Users (LSUs):
If the speed of ground user GU j , denoted by υ j ( t ) , is less than a predefined threshold υ th , this user can be managed by inter-clusters. Similarto the indoor environments, the transmission scheme isdetermined based on the content popularity profile andthe ground user’s location. • High Speed Users (HSUs):
In this case, the speed ofground user GU j is equal or more than υ th . Therefore,this request should be served by a UAV that covers thecorresponding intra-cluster.This completes our discussion on users’ mobility pattern, thecontent popularity profile, and transmission schemes. Next, wedevelop the CCUF framework.III. T HE CCUF F
RAMEWORK
In conventional femtocaching schemes, it is commonlyassumed that all caching nodes store the same most popularcontents. This assumption is acceptable in static femtocachingmodels, in which users are stationary or move with a lowvelocity. With the focus on a dynamic femtocaching network,in which users can move based on the random walk model,storing distinct contents in neighboring FAPs leads to increas-ing the number of requests served by caching nodes. Despiterecent researches on cluster-centric cellular networks, thereis no framework to determine how different segments shouldbe stored to increase content diversity. Toward this goal, wepropose the CCUF framework, which is an efficient contentplacement strategy for the network model introduced in Sec-tion II. The most remarkable idea behind the CCUF strategyis to identify the most widely requested contents to maximizethe cache-hit-ratio and minimize the users’ access delay. Giventhe common content, an inter-cluster is formed by N s numberof adjacent FAPs, in which distinct segments of coded contentare stored in the FAP’s cache. Consequently, the number ofrequests served by caching nodes drastically increases withoutany growth of the storage of FAPs. Additionally, N ic numberof near FAPs as an intra-cluster is managed by one UAV toprovide low-latency communication for outdoor ground users.The proposed CCUF framework is implemented based on thesteps presented in the following subsections. A. Optimal Content Caching for FAPs and UAVs
Identifying the optimal multimedia contents to be storedin the storage of caching nodes leads to a reduction in theusers’ latency. It is a commonly assumed [3], [29] that thetotal users’ access delay is determined according to availabilityof the required content in the nearby caching nodes. Basedon this assumption, in scenarios where the requested contentcan be served by caching nodes, the cache-hit occurs and theground user will experience no delay; otherwise, the requestis served by the main server resulting in a cache-miss. In thispaper, we relax the above assumption and express the users’access delay as a function of the content popularity profile andthe distance between the ground user and the target cachingnode. In the regard, we propose two optimization models forcontent placement in both FAPs and UAVs to minimize theusers’ latency. Toward this goal, first we describe the delaythat ground users experience when served by FAPs and UAVs.
1) UAVs’ Content Placement:
In order to calculate theusers’ access delay through UAVs, first, we investigate theeffect of distance between ground user GU j and target UAV u k . Serving requests via UAVs leads to establishing air-to-ground links from UAVs to ground users. Due to the obstacles in outdoor environments, the transmitted signal from UAVs isattenuated. To be practical, we consider both LoS and Non-LoS (NLoS) path losses from UAV u k to ground user GU j attime slot t as follows [5] L ( LoS ) k,j ( t ) = L + 10 η ( LoS ) log( d k,j ( t )) + χ ( LoS ) σ , (8) L ( NLoS ) k,j ( t ) = L + 10 η ( NLoS ) log( d k,j ( t )) + χ ( NLoS ) σ , (9)where L = 20 log (cid:18) πf c d c (cid:19) denotes the reference pathloss in distance d , and d k,j ( t ) is the Euclidean distancebetween UAV u k and the ground user GU j at time slot t .In addition, η ( LoS ) , η ( NLoS ) , χ ( LoS ) σ and χ ( NLoS ) σ indicatethe LoS and NLoS path loss exponents and the correspondingshadowing effects, respectively. Consequently, the averagepath loss, denoted by L k,j ( t ) , is obtained as L k,j ( t ) = p ( LoS ) k,j ( t ) L ( LoS ) k,j ( t ) + (1 − p ( LoS ) k,j ( t )) L ( NLoS ) k,j ( t ) , (10)where p ( LoS ) k,j ( t ) is the probability of establishing LoS linkbetween UAV u k and ground user GU j at time slot t , obtainedas [15] p ( LoS ) k,j ( t ) = (1 + ϑ exp ( − ζ [ φ k,j ( t ) − ϑ ])) − , (11)where ϑ and ζ are constant parameters, depending on therural and urban areas. Moreover, φ k,j ( t ) = sin − (cid:18) h k d k,j ( t ) (cid:19) is the elevation angle between UAV u k and the ground user GU j , and h k is the UAV’s altitude. Without loss of generality,altitude h k is assumed to be a fixed value over hoveringtime. If the requested content cannot be found in the storageof UAVs, additional ground-to-air connection is required toprovide UAVs with the requested content through the mainserver. Similarly, the average path loss of the main server-to-UAV u k link is calculated as L m,k ( t ) = p ( LoS ) m,k ( t ) L ( LoS ) m,k ( t ) + (1 − p ( LoS ) m,k ( t )) L ( NLoS ) m,k ( t ) , (12)where L ( LoS ) m,k ( t ) = d − (cid:36)m,k ( t ) and L ( NLoS ) m,k ( t ) = ψ L ( LoS ) m,k ( t ) , inwhich d m,k ( t ) denotes the distance between the main serverand UAV u k . Furthermore, (cid:36) and ψ denote the LoS and NLoSpath loss exponents, respectively [5].As mentioned previously, another parameter that has agreat impact on the users’ access delay is the presence ofthe requested content in the caching node, depending onthe content popularity profile. Therefore, the cache-hit andthe cache-miss probability through serving by UAV u k attime slot t , denoted by p ( h ) u ( t ) and p ( m ) u ( t ) , respectively, areexpressed as p ( h ) u ( t ) = (cid:88) l ∈ C u p l ( t ) ≤ , (13) p ( m ) u ( t ) = 1 − p ( h ) u ( t ) , (14)where C u denotes the cache size of UAV u k , which is assumedto be the same for all UAVs. Consequently, the users’ accessdelay through UAVs is expressed as D u ( t ) = p ( h ) u ( t ) D ( h ) u ( t ) + p ( m ) u ( t ) D ( m ) u ( t ) , (15) where D ( h ) u ( t ) and D ( m ) u ( t ) represent the cache-hit and thecache-miss delays, respectively, calculated as follows D ( h ) u ( t ) = L c R k,j = L c log − (cid:32) P k L k,j ( t ) / I k ( t, u − k ) + N (cid:33) , (16) D ( m ) u ( t ) = L c log − (cid:32) P k L m,k ( t ) / I k ( t, u − k ) + N (cid:33)(cid:124) (cid:123)(cid:122) (cid:125) (cid:44) L MU + L c log − (cid:32) P k L k,j ( t ) / I k ( t, u − k ) + N (cid:33)(cid:124) (cid:123)(cid:122) (cid:125) (cid:44) L UG , (17)where L c and R k,j represent the file size of c l and thetransmission data rate from UAV u k to GU j . Furthermore, P k and I k ( t, u − k ) denote the transmission power of UAV u k and the interference caused by other UAV-user links forthe transmission link between u k and GU j , respectively. Notethat when the cache-miss happens, the content should be firstprovided for the UAV by the main server. Therefore, L MU and L UG in Eq. (17) represent the users’ access delay related tothe main server-UAV and UAV-ground user links, respectively.Given users’ access delay through UAVs, the goal is to placecontents in the storage of UAVs to minimize the users’ accessdelay in Eq. (15). Due to the large coverage area of UAVs, itis not feasible to move through areas supported by differentUAVs frequently. Therefore, we assume that contents arecached completely (either popular or mediocre ones). Towardthis goal, the cached contents are selected as the solution of thefollowing optimization problem to minimize the users’ accessdelay: min x l N c (cid:88) l =1 (cid:16) N g (cid:88) j =1 (cid:0) − p ( j ) l ( t ) (cid:1) D ( j ) u ( t ) (cid:17) x l (18)s.t. C1. x l ∈ { , } , C2. N c (cid:88) l =1 x l ≤ C u , where p ( j ) l ( t ) denotes the probability of requesting content c l by the ground user GU j at time slot t , which is obtained ac-cording to the request history of ground users [3]. Furthermore, D ( j ) u ( t ) is the delay that the ground user GU j may experience,which is calculated based on Eq. (15). In the constraint C1 , x l is an indicator variable, which is equal to when content c l exists in the cache of UAV u k . Moreover, the constraint C2 represents that the total contents cached in the storage of u k should not exceed the storage capacity of UAV u k .
2) FAPs’ Content Placement:
Serving requests by FAPsleads to a ground-to-ground connection type between FAPsand ground users. Similarly, the users’ access delay throughFAP connections is calculated as D f ( t ) = p ( h ) k ( t ) D ( h ) f ( t ) + p ( m ) k ( t ) D ( m ) f ( t ) , (19)where D ( h ) f ( t ) , as the cache-hit delay, is expressed as D ( h ) f ( t ) = L c log − (1 + S i,j ( t )) , (20) Fig. 3: (a) A typical hexagonal cellular network, where seven FAPs forman inter-cluster, (b) Real coverage area of FAPs in a practical model. with S i,j ( t ) denoting the SINR from f i to GU j defined inEq. (6). The optimal coded contents to be stored in the storageof FAPs are determined according to the solution of thefollowing optimization problem: F ( y , z ) = min y l ,z l (cid:98) αC f (cid:99) (cid:88) l =1 (cid:16) N g (cid:88) j =1 (cid:0) − p ( j ) l ( t ) (cid:1) D ( j ) f ( t ) (cid:17) y l (21) + N s ( C f −(cid:98) αC f (cid:99) ) (cid:88) l = (cid:98) αC f (cid:99) +1 (cid:16) N g (cid:88) j =1 (cid:0) − p ( j ) l ( t ) (cid:1) D ( j ) f ( t ) (cid:17) z l , s.t. C1. y l , z l ∈ { , } , C2. (cid:98) αC f (cid:99) (cid:88) l =1 y l ≤ αC f , C3. N s C f − (1+ N s ) (cid:98) αC f (cid:99) ) (cid:88) l =1 z l ≤ (1 − α ) N s C f , where F ( y , z ) is the cost function associated with users’access delay, experienced by serving the request through FAPs.By assuming that N p = (cid:98) αC f (cid:99) and N a = N s ( C f − (cid:98) αC f (cid:99) ) are the cardinality of popular and mediocre contents, re-spectively, y = [ y , . . . , y N p ] T is an indicator vector forpopular contents, where y l would be 1 if l th content is storedin the cache of FAPs, otherwise it equals to 0. Similarly, z = [ z , . . . , z N a ] T is an indicator variable for mediocrecontents. According to the optimization problem, despite pop-ular contents that are stored completely, just one segment ofmediocre contents are cached. Similarly, y l and z l in constraint C1 illustrate the availability of content c l in the cache of FAP f i . Finally, constraints C2 and C3 indicate the portion of cacheallocated to popular and mediocre contents, respectively. B. Content Placement in Multiple Inter-Clusters
After identifying popular and mediocre contents, we needto determine how to store different segments of mediocrecontents within an inter-cluster. Lack of prior research studieson the coded content placement in a cluster-centric cellularnetwork (e.g., [30], [31]) motivates us to propose the CCUFframework. Without loss of generality, we first consider a sim-ple hexagonal cellular network including N s = 7 FAPs as oneinter-cluster (Fig. 3(a)). Given the vector z = [ z , . . . , z N a ] T that determines the mediocre contents, in this phase, we need to indicate which segment of the mediocre content c l , denotedby c ls for (1 ≤ l ≤ N a ) and (1 ≤ s ≤ N s ) , should be cachedin FAP f i for (1 ≤ i ≤ N s ) . In this regard, we form an N a × N s indicator matrix of FAP f i , denoted by Z f i . Notethat the l th row of Z f i indicates segments of file c l is storedin the cache of FAP f i , where z ls = 1 means that s th segmentof file c l is stored in the cache of FAP f i . Consequently, thecached contents of other FAPs in an inter-cluster is determinedas follows z f i ,l z Tf j ,l = 0 , i = 1 , . . . , N s , j = 1 , . . . , N s , i (cid:54) = j, (22)where z f i ,l denotes the l th row of Z f i . After allocatingmediocre contents to FAPs inside an inter-cluster, the samecontent as FAP f i is stored in FAP f k in the neighboringinter-cluster, where k is given by Z f k = Z f i if k = w + wz + z , (23)where w and z represent the number of FAPs required to reachanother FAP storing similar contents, in two different direc-tions. For example, in Fig. 3(a) N s = 7 , w = 2 , and z = 1 forstarting in a FAP including c and reaching the similar FAPin a neighboring inter-cluster. The main idea behind the codedplacement scheme in our proposed CCUF framework comesfrom the frequency reusing technique in cellular networks [37].In order to increase the resource availability and minimuminter-cell interference, the same radio frequencies are used indifferent cells, where the distance between two cells with thesame spectrum bandwidth is determined based on Eq. (23). Remark 1 : As shown in Fig. 3(b), the coverage area of FAPsin a practical scenario is influenced by path loss and shadowingmodels. Location p = ( x, y ) is placed within the transmissionarea of FAP f i , if the strength of the received signal at point p ,denoted by RSSI p , is higher than the threshold value RSSI th ,where RSSI p is calculated as RSSI p ( dB ) = RSSI ( d ) + 10 η log ( dd ) + X σ , (24)where d and d denote the distance between FAP and point p in the boundary of transmission area of FAP, and the referencedistance is set to m, respectively. Moreover, η represents thepath loss exponent, which is dB or dB, and X σ is azero-mean Gaussian with standard deviation σ that representsthe effect of multi-path fading in our UAV-based femtocachingscheme [38]. C. Cache-Hit Probability in the Proposed CCUF Framework
To quantify the benefits of the proposed CCUF strategy,we form the probability of finding a new segment by user GU j at time slot t under the following two scenarios: ( i ) Uncoded cluster-centric, and; ( ii ) Coded cluster-centric UAV-aided femtocaching network, denoted by p uc , and p cc , respec-tively. Concerning the nature of mobile networks, ground usersmove and leave their current positions. In this paper, it isassumed that low-speed ground users can obtain one segmentin each contact, i.e., T = N s δ t is required to completelydownload content c l . First, we consider a simple mobilitypattern, where ground users are positioned in the transmission area of a new FAP in each time slot t . Eventually, in T = N s δ t ,the whole content c l will be downloaded. Then, we generalizethe mobility pattern to the DCRW model, where ground usersare moving randomly.
1) Simple Mobility Pattern:
Regarding the uncoded cluster-centric UAV aided femtocaching framework, content c l con-sisting of c ls , for ( ≤ s ≤ N s ) segments, is stored completelyin all FAPs. Consequently, the probability of downloading n = N s segments of content c l in T = N s δ t depends onthe probability of requesting file c l , denoted by p l . Since thestorage capacity of each FAP is equal to C f , the cache-hitprobability, denoted by p uc , is obtained as follows p uc [ n = N s , t = T ] = C f (cid:88) l =1 p l . (25)Note that, since contents are completely stored in uncodedcluster-centric framework, ground users can access the wholecontent in each contact. By considering the size of contentand the speed of ground users, however, it is not possible todownload more than one segment in each contact. Therefore,storing the whole of contents in all FAPs is not beneficial. Onthe other hand, the cache-hit probability in the coded cluster-centric UAV-aided femtocaching network is obtained as p cc [ n = N s , t = T ] = N p (cid:88) l =1 p l + N s ( C f −(cid:98) αC f (cid:99) ) (cid:88) l = N p +1 p l . (26)To illustrate the growth rate of the cache-hit probability in thecoded one, we rewrite p uc in Eq. (25) as follows p uc [ n = N s , t = T ] = N p (cid:88) l =1 p l + C f (cid:88) l = N p +1 p l . (27)As it can be seen from Eqs. (26), and (27), the first termrelated to the popular content is the same. The second term,however, illustrates that the number of contents that can beserved through FAPs in the coded cluster-centric network is κ times greater than the uncoded one, where κ is given by κ = N s ( C f − (cid:98) αC f (cid:99) ) C f . (28)Accordingly, due to the allocation of different segments in thecoded cluster-centric network, more segments of the desiredcontents are accessible. Therefore, more requests can beserved in comparison to the uncoded cluster-centric UAV-aidedfemtocaching networks.
2) Generalizing to the DCRW Mobility Pattern:
Fig. 3(a)provides an overall overview of the proposed network, con-sisting of numerous inter-clusters, where ground users moverandomly in all directions. In other words, it is possible forground users to return back to the previous coverage area ofFAPs (which is not considered in the simple mobility pattern).Moreover, in some cases, ground users may be positioned inthe transmission area of a FAP, which stores the same segmentof the required content that the ground user has alreadydownloaded. Consequently, the cache-hit probability of thecoded cluster-centric will not be the same as the previousscenario. If the requested content is the popular one, regardless of the location of the ground user within an inter-cluster, theground user can download one segment of the required contentwith the probability of N p (cid:80) l =1 p l at each contact. Despite this partof the cache-hit probability, which is constant, the successfulprobability of downloading a new segment of a mediocrecontent in each contact depends on the current and all previouslocations of the ground user. Therefore, we first determinethe successful probability of achieving a new segment of amediocre content, denoted by p ns ( n = n , t = n δ t ) , for( ≤ n ≤ N s ) . Then, we calculate the cache-hit probability ofa coded cluster-centric network based on the DCRW mobilitypattern.As it can be seen from Fig. 3(a), regardless of the locationof GU j , this user can download one segment successfully inthe first contact (i.e., n = 1 ). Therefore, we have p ns ( n =1 , t = δ t ) = 1 . Similarly, where n = 2 , the ground user GU j can download a new segment without considering its location.Therefore, the probability of downloading two segments aftertwo contacts is p ns [ n = 2 , t = 2 δ t ] = 1 . More precisely, in thesecond contact, the ground user can be positioned in the cellof ( N s − number of FAPs, where the probability of beingin the cell of FAP f i is p ( f = f i ) = 1( N s − . Therefore, wehave p ns [ n = 2 , t = 2 δ t ] = N s − (cid:88) i =1 p ns [ n = 2 , t = 2 δ t | f = f i ] p ( f = f i )=( N s − × × N s −
1) = 1 . (29)Accordingly, the probability of finding a new segment in thethird contact is p ns [ n = 3 , t = 3 δ t ] = N s − (cid:88) i =1 p ns [ n = 3 , t = 3 δ t | f = f i ] p ( f = f i )= ( N s −
2) 1( N s − × N s − × N s − N s − , (30)where ( N s − FAPs have different segments, whereas if GU j returns to the FAP at t = δ t , the ground user can find a similarsegment. Considering the fact that N s = 7 in our proposedwireless network, p ns [ n = 3 , t = 3 δ t ] would be . . It meansthat user GU j can find a new segment of the desired contentin the third contact with a probability of . . Similarly, itcan be proved that the probability of finding a new segmentin n > is obtained as p ns [ n = n , t = n δ t ] = N s − (cid:88) i =1 p ns [ n = n , t = n δ t | f = f i ] × p ( f = f i ) = ( N s − n − ( N s − n − . for n > (31)Taking into account the unequal likelihood of finding newsegments of mediocre contents in different contacts, we re- Algorithm 1
Proposed CCUF Strategy Initialization:
Set α , λ , N s , and C f . Input: p ( j ) l ( t ) , L ( t ) k , and L ( t ) j . Output: x l , y l , and z l . Content Placement Phase: for u k , k = 1 , . . . , N u , do min x l N c (cid:88) l =1 (cid:16) N g (cid:88) j =1 (cid:0) − p ( j ) l ( t ) (cid:1) D ( j ) u ( t ) (cid:17) x l s.t. C1. and
C2. in Eq. (18). end for for f i , i = 1 , . . . , N f , do min y l ,z l (cid:98) αC f (cid:99) (cid:88) l =1 (cid:16) N g (cid:88) j =1 (cid:0) − p ( j ) l ( t ) (cid:1) D ( j ) f ( t ) (cid:17) y l + N s ( C f −(cid:98) αC f (cid:99) ) (cid:88) l = (cid:98) αC f (cid:99) +1 (cid:16) N g (cid:88) j =1 (cid:0) − p ( j ) l ( t ) (cid:1) D ( j ) f ( t ) (cid:17) z l , s.t. C1. C3. in Eq. (21). end for z f i ,l z Tf j ,l = 0 , i = 1 , . . . , N s , j = 1 , . . . , N s , i (cid:54) = j, Z f k = Z f i if k = w + wz + z , Transmission Phase: for GU j , j = 1 , . . . , N g , do if GU j is in indoor environment then if GU j is an edge-user and requests popular content then The request should be handled according to the
JT scheme. else
The request should be handled according to the
ST scheme. end if else if υ j ( t ) ≥ υ th then The request is served by UAV u k . else Similar to lines to . end if end if end for calculate p cc as follows p cc [ n = N s , t = T ] = N p (cid:88) l =1 p l + N s (cid:88) n =1 ( N s − n − ( N s − n − (cid:16) N s ( C f −(cid:98) αC f (cid:99) ) (cid:88) l = N p +1 p l (cid:17) . (32) D. 2-D Deployment of UAVs in Intra-clusters
To increase the resource availability for ground users, theoutdoor environment is partitioned based on an unsupervisedlearning algorithm, each partition is covered by a UAV. Con-sidering a Gaussian mixture distribution for ground users, we have a dense population of ground users in some areas. Themain goal is to deploy UAVs in such a way that groundusers can experience high QoS communications even in adense area. Note that the distance between UAVs and groundusers is a critical factor that can significantly impact the QoSfrom different perspectives such as the energy consumption ofUAVs and the users’ access delay. Our goal is to partition N g ground users into K intra-clusters, where the sum of Euclideandistances between the ground user GU j , for (1 ≤ j ≤ N kg ) ,and UAV u k is minimized. In this case, N kg is the cardinality ofground users positioned in the intra-cluster related to the UAV u k . Therefore, the UAVs’ deployment is obtained according tothe following optimization problem min l k ( t ) N u (cid:88) k =1 N kg (cid:88) j =1 || l j ( t ) , l k ( t ) || , (33)where l k ( t ) denotes the location of the UAV u k at time slot t , defined as the mean of the coordinates of all ground usersinside the corresponding intra-cluster as follows l k ( t ) = N kg (cid:80) j =1 l j ( t ) N kg , k = 1 , . . . , N u . (34)To solve the above optimization problem, we utilize theK-Means clustering algorithm [39], which is known as anefficient unsupervised learning framework. In the first step,a set of points, denoted by P = { P , . . . , P N u } , is generated,where P k for ( ≤ k ≤ N u ) should be within the pre-specifiedenvironment. Then, the set of ground users in the vicinity of P k is determined as follows u j ∈ N kg if || l j ( t ) , P k || < || l j ( t ) , P r || , ∀ k (cid:54) = r. (35)Given the set of ground users belonging to each intra-cluster,UAVs’ locations are determined according to Eq. (34). In thesecond step, by moving ground users from one intra-clusterto another, the Euclidean distances between ground users andUAVs are calculated to update the location of UAVs accordingto Eq. (33). The K-Means algorithm is terminated when thereis no change in the ground users belonging to an intra-cluster over several iterations. This completes our discussionon development of the CCUF scheme. The pseudo-code of theproposed CCUF framework is summarized in Algorithm 1 .IV. S
IMULATION R ESULTS
To demonstrate the advantage of the proposed CCUF frame-work, we consider a macro cellular network consisting ofone MBS with the radius R = 1000 m, N f = 180 FAPs,and N u = 10 UAVs, where each inter-cluster compromisesof N s = 7 FAPs. Fig. 4 illustrates a typical × m area, where ground users are randomly distributed and theirlocations are determined according to the AoA localizationmethod. It can be shown that the Root Mean Square Error(RMSE) between the estimated and the actual location of theground users is about . m, which is acceptable in comparisonto the transmission range of FAPs. Fig. 4:
Typical location estimation results based on the AoA localizationscheme.
Fig. 5 depicts an integrated heterogeneous network, whereyellow and red areas determine indoor and outdoor environ-ments, respectively. Fig. 5 also shows the deployment of UAVsin the intra-clusters within the network, which is generated bypartitioning ground users according to the K-means clusteringalgorithm. As a result of the Gaussian mixture distributionfor clients, we have a dense population in some areas, whichcan be changed over time by the movement of ground users.Therefore, the location of N u = 10 UAVs and the formationof intra-clusters in this paper is varying, depending on theuser density distribution. The general simulation parametersare summarized in Table I. In order to find the optimum valueof α , three types of caching strategies are considered: • Uncoded UAV-aided Femtocaching (UUF):
Without cod-ing and clustering, popular contents are stored completelyinto FAPs and UAVs. In this case, the value of α , whichindicates the percentage of contents stored completely,would be one (i.e., α = 1 ). • Proposed Cluster-centric and Coded UAV-aided Femto-caching (CCUF):
In this case, the uncoded popular andthe coded mediocre contents are stored in the cachingnodes, where < α < . According to the simulationresults, the best value of α is obtained. • The Conventional Cluster-centric and Coded UAV-aidedFemtocaching (Conventional CCUF):
In this framework,regardless of the content popularity profile, all contentsare stored partially in this framework. Consequently, thevalue of α is equal to zero.These three strategies are evaluated over the cache-hit-ratio,cache diversity, cache redundancy, SINR, and users’ accessdelay to determine the best value of α . Moreover, to illustratethe effect of considering a UAV-aided femtocaching frame-work in an integrated network, we compare the users’ accessdelay and energy consumption of UAVs, by serving users inboth indoor and outdoor areas. Cache-Hit-Ratio:
This metric illustrates the number of re-quests served by caching nodes versus the total number ofrequests made across the network. The high value of cache-hit-ratio shows the superiority of the framework. Since we assume Fig. 5:
Deployment of UAVs in intra-clusters within an integrated network,where “yellow” and “red” colors indicate indoor and outdoor environments,respectively.
TABLE I: List of Parameters.
Notation Value Notation Value N g η ( LoS ) , η ( NLoS ) . , N f h k m N u (cid:36) , ψ , N s L c . MB N c τ p − s R u m P k dBm R f m χ ( LoS ) σ , χ ( NLoS ) σ . , P T ( t ) , P R ( t ) 0 . , . W N − dBm Fig. 6:
The cache-hit-ratio versus the popularity parameter γ for differentvalues of α . that ground users can download one segment in each contact,we evaluate the cache-hit-ratio in terms of the number offragmented contents served by caching nodes. Fig. 6 comparesthe cache-hit-ratio of the UUF ( α = 1 ), the proposed CCUF( < α < ), and conventional CCUF ( α = 0 ) frameworksversus the value of γ . As previously mentioned, parameter γ shows the skewness of the content popularity, where γ ∈ [0 , .Note that the large value of γ indicates that a small numberof contents has a high popularity, where a small value of Fig. 7:
The cache-hit-ratio versus the α percentage of contents that arestored completely. γ illustrates an almost uniform popularity distribution forthe majority of contents. As it can be seen from Fig. 6,depending on the popularity distribution of contents, γ , theconventional CCUF framework results in a higher cache-hit-ratio. The most important reason is that given a constant cachecapacity, the coded content placement of the conventionalCCUF strategy leads to a remarkable surge in the contentdiversity. In contrast, for a high value of γ , where a smallnumber of contents is widely requested, the UUF and theproposed CCUF frameworks have better results comparedto the conventional CCUF. By considering the fact that thecommon value of γ is about . ≤ γ ≤ . (e.g., see [4],[29], [40]), we define CHR th as the threshold cache-hit-ratio,which is the average of cache-hit-ratio of different values of α for a specific γ . As it can be seen from Fig. 6, the proposedCCUF framework with < α ≤ . and the UUF schemeoutperform other schemes from the aspect of cache-hit-ratio.Fig. 7 shows the cache-hit-ratio versus different values of α when the popularity parameter γ changes in the range of . too . Accordingly, for . ≤ γ ≤ . , by increasing the valueof α , the cache-hit-ratio decreases drastically. In the following,we also investigate the impact of α on the users’ access delayto determine the best value of α . Users’ Access Delay:
Users’ access delay depends on threeparameters, i.e., the availability of the content in cachingnodes, the distance between the ground user and the corre-sponding caching node, and the channel quality, known as theSINR. Figs. 8 and 9 compare the users’ access delay of theaforementioned frameworks, which is obtained according toEq. (19). By utilizing the CoMP technology in the proposedCCUF, serving edge-users according to the JT scheme has agreat impact on the SINR, where users’ access delay decreaseby increasing the SINR. As can be seen from Table II, theSINR of edge-users improves by increasing the value of α .Note that JT scheme can be performed if the same contentsare stored in the neighboring FAPs. Therefore, by increasingthe value of α , the users’ access delay will decrease. Withthe same argument, we define D th , which is the average ofusers’ access delay of different values of α for a specific γ , Fig. 8: The users’ access delay in the indoor environment versus differentvalue of γ . Fig. 9:
The users’ access delay in the indoor environment versus differentvalues of α . TABLE II:
The SINR experienced by edge-users for different values of α and γ . γ = 0 . γ = 0 . γ = 0 . γ = 0 . γ = 0 . γ = 1 α = 0 16 .
37 16 .
37 16 .
37 16 .
37 16 .
37 16 . α = 0 . .
55 18 .
12 18 .
89 19 .
84 20 .
88 21 . α = 0 . .
01 18 .
65 19 .
46 20 .
40 21 .
38 22 . α = 0 . .
62 19 .
30 20 .
11 21 .
00 21 .
90 22 . α = 0 . .
06 19 .
75 20 .
53 21 .
37 22 .
20 22 . α = 0 . .
42 20 .
09 20 .
85 21 .
64 22 .
42 23 . α = 1 19 .
72 20 .
38 21 .
11 21 .
86 22 .
58 23 . shown in Fig. 8. Therefore, the best value of α would be α ≥ . . Consequently, the cache-hit-ratio and users’ accessdelay of the proposed CCUF framework would be efficient if α ∈ [0 . , . . Cache Diversity:
This metric illustrates the diversity of con-tents in an inter-cluster, which is defined as the number ofdistinct segments of contents, expressed as follows CD = N a N s C f = 1 − (cid:98) αC f (cid:99) C f . (36)As stated previously, we have N a = N s ( C f − (cid:98) αC f (cid:99) ) . Asit can be seen from Fig. 10, the value of CD would be one, Fig. 10:
The percentage of the cache diversity and the cache redundancyversus different values of α . Fig. 11:
The maximum cache capacity, required to achieve the maximumcache diversity, versus different values of α . if α = 0 , which means that all cached contents are different.The cache diversity, however, linearly decreases by increasingthe value of α , and reaches the lowest value zero, when allcontents are cached completely (i.e., α = 1 ). Cache Redundancy:
This metric indicates the number ofsimilar contents that ground users meet during their randommovements. As it can be seen from Fig. 10, the cache redun-dancy increases by storing the entire contents. By consideringthe coded content placement, even in the proposed CCUFframework, ground users that move randomly through thenetwork, can meet a similar coded contents during theirmovements (see Fig. 3).
Maximum Required Cache Capacity:
Given a specific numberof contents through the network, denoted by N c , the storagecapacity of caching nodes is determined by C f = βN c . In thiscase, parameter β indicates the percentage of contents that canbe stored in caching nodes. In the coded content placement,since only one segment of the contents is cached, it is fairlylikely that the total number of possible segments that can Fig. 12: The users’ access delay experienced through UAVs in both indoorand outdoor versus different values of β . be cached exceeds the total number of contents. Therefore,the maximum required cache capacity, denoted by β max , fordifferent values of α is obtained as β max ≤ N c N s N c − ( N s − αN c = 1 α (1 − N s ) + N s , (37)where the remainder of the storage would be occupied byredundant contents if β > β max . As it can be seen fromFig. 11, the maximum cache capacity β max increases by thevalue of α . Consequently, in smaller values of α , we need asmaller cache capacity to have the maximum cache diversity. Users’ Access Delay through UAVs and UAVs’ Energy Con-sumption:
Finally, we evaluate the users’ access delay andthe energy consumption of UAVs in Figs. 12 and 13, whenthe corresponding ground user is located in both indoorand outdoor environments. According to Eqs. (11), and (16),and (17), the probability of establishing a LoS connectionbetween the ground user GU j and the UAV u k has a greatimpact on the users’ access delay, served by UAVs. To tacklethis problem, we assume that ground users in indoor areasare served by inter-clusters. With the same argument, Fig. 13illustrates the energy consumption of UAVs, calculated asfollows [8] E ( LoS ) u k ( t ) = L c P T ( t ) τ p + L c P R ( t ) τ p + P ( LoS ) j ( t )( τ f − τ p ) , (38) E ( NLoS ) u k ( t ) = L c P T ( t ) τ p + L c P R ( t ) τ p + P ( NLoS ) j ( t )( τ f − τ p ) , (39)where P T ( t ) and P R ( t ) represent the power consumed fortransmission and reception powers of Mb file, respectively.Moreover, P j ( t ) , τ f , and τ p denote the received power atground user GU j , and the flyby and the pause times of UAV u k , respectively. V. C ONCLUSIONS
In this paper, we developed a Cluster-centric and CodedUAV-aided Femtocaching (CCUF) framework for an integratedand dynamic cellular network to maximize the number ofrequests served by caching nodes. To increase the cache Fig. 13:
The energy consumption of UAVs in both indoor and outdoorenvironments in different time slots. diversity and to store distinct segments of contents in neigh-boring FAPs, we employed a two-phase clustering techniquefor FAPs’ formation and UAVs’ deployment. In this case,we analytically formulated the cache-hit probability of theproposed CCUF framework. Moreover, in the cluster-centriccellular network, multimedia contents were coded based ontheir popularity profiles. In order to benefit the CoordinatedMulti-Point (CoMP) technology and to improve the inter-cellinterference, we determined the best value of the number ofcontents that should be stored completely. According to thesimulation results and by considering the optimum value of α , the proposed CCUF framework results in an increase inthe cache-hit-ratio, SINR, and cache diversity and decreaseusers’ access delay and cache redundancy. Going forward,several directions deserve further investigation. First, it is ofinterest to introduce a Reinforcement Learning (RL)-basedmethod for outdoor environment, where ground users can beautonomously served by UAVs or FAPs, based on the dynamicpopulation of their current locations and their speeds. Second,the optimum number of ground users to be served by a UAVin the proposed network needs to be analyzed.R EFERENCES[1] V. Chamola, V. Hassija, V. Gupta and M. Guizani, “A ComprehensiveReview of the COVID-19 Pandemic and the Role of IoT, Drones, AI,Blockchain, and 5G in Managing its Impact,”
IEEE Access , vol. 8, pp.90225-90265, May 2020.[2] H. Hui, Y. Ding, Q. Shi, F. Li, Y. Song, and J. Yan, “5G network-basedInternet of Things for Demand Response in Smart Grid: A survey onApplication Potential,”
Applied Energy , vol. 257, pp. 113972-113987,Jan. 2020.[3] Z. Hajiakhondi-Meybodi, J. Abouei, and A. H. F. Raouf, “CacheReplacement Schemes Based on Adaptive Time Window for Video onDemand Services in Femtocell Networks,”
IEEE Transactions on MobileComputing , vol. 18, no. 7, pp. 1476-1487, July 2019.[4] Z. HajiAkhondi-Meybodi, J. Abouei, M. Jaseemuddin and A. Moham-madi, “Mobility-Aware Femtocaching Algorithm in D2D NetworksBased on Handover,” accepted for publication in
IEEE Transactionson Vehicular Technology , June 2020.[5] B. Jiang, J. Yang, H. Xu, H. Song, and G. Zheng, “MultimediaData Throughput Maximization in Internet-of-Things System Based onOptimization of Cache-Enabled UAV,”
IEEE Internet of Things Journal ,vol. 6, no. 2, pp. 3525-3532, Apr. 2019. [6] F. Cheng, G. Gui, N. Zhao, Y. Chen, J. Tang, and H. Sari, “UAV-Relaying-Assisted Secure Transmission With Caching,”
IEEE Transac-tions on Communications , vol. 67, no. 5, pp. 3140-3153, May 2019.[7] N. Zhao, F. Cheng, F. R. Yu, J. Tang, Y. Chen, G. Gui, and H. Sari,“Caching UAV Assisted Secure Transmission in Hyper-Dense NetworksBased on Interference Alignment,”
IEEE Transactions on Communica-tions , vol. 66, no. 5, pp. 2281-2294, May 2018.[8] V. Sharma, I. You, D. N. K. Jayakody, D. G. Reina, and K. K. R. Choo,“Neural-Blockchain-Based Ultrareliable Caching for Edge-Enabled UAVNetworks,”
IEEE Transactions on Industrial Informatics , vol. 15, no.10, pp. 5723-5736, Oct. 2019.[9] J. Lee, K. Kim, M. Kim, J. Park, Y. K. Yoon and Y. J. Chong,“Measurement-Based Millimeter-Wave Angular and Delay DispersionCharacteristics of Outdoor-to-Indoor Propagation for 5G Millimeter-Wave Systems,”
IEEE Access , vol. 7, pp. 150492-150504, Oct. 2019.[10] R. Avanzato and F. Beritelli, “A Smart UAV-Femtocell Data SensingSystem for Post-Earthquake Localization of People,”
IEEE Access , vol.8, pp. 30262-30270, Feb. 2020.[11] Z. Wang, L. Duan and R. Zhang, “Adaptive Deployment for UAV-Aided Communication Networks,”
IEEE Transactions on WirelessCommunications , vol. 18, no. 9, pp. 4531-4543, Sept. 2019.[12] X. Liu, Y. Liu and Y. Chen, “Reinforcement Learning in Multiple-UAVNetworks: Deployment and Movement Design,”
IEEE Transactions onVehicular Technology , vol. 68, no. 8, pp. 8036-8049, Aug. 2019.[13] M. Samir, S. Sharafeddine, C. Assi, T. M. Nguyen, and A. Ghrayeb,“Trajectory Planning and Resource Allocation of Multiple UAVs forData Delivery in Vehicular Networks,”
IEEE Networking Letters , vol.1, no. 3, pp. 107-110, Sept. 2019.[14] S. Li, B. Duo, X. Yuan, Y. Liang and M. Di Renzo, “Reconfigurable In-telligent Surface Assisted UAV Communication: Joint Trajectory Designand Passive Beamforming,”
IEEE Wireless Communications Letters , vol.9, no. 5, pp. 716-720, May 2020.[15] M. Chen, W. Saad, and C. Yin, “Liquid State Machine Learning forResource and Cache Management in LTE-U Unmanned Aerial Vehicle(UAV) Networks,”
IEEE Transactions on Wireless Communications , vol.18, no. 3, pp. 1504-1517, Mar. 2019.[16] Z. Yang, C. Pan, K. Wang and M. Shikh-Bahaei, “Energy Efficient Re-source Allocation in UAV-Enabled Mobile Edge Computing Networks,”
IEEE Transactions on Wireless Communications , vol. 18, no. 9, pp.4576-4589, Sept. 2019.[17] B. Ji, Y. Li, B. Zhou, C. Li, K. Song, and H. Wen, “PerformanceAnalysis of UAV Relay Assisted IoT Communication Network EnhancedWith Energy Harvesting,”
IEEE Access , vol. 7, pp. 38738-38747, Mar.2019.[18] L. Zhang, Z. Zhao, Q. Wu, H. Zhao, H. Xu, and X. Wu, “Energy-AwareDynamic Resource Allocation in UAV Assisted Mobile Edge ComputingOver Social Internet of Vehicles,”
IEEE Access , vol. 6, pp. 56700-56715,Oct. 2018.[19] B. Li, Z. Fei, and Y. Zhang, “UAV Communications for 5G and Beyond:Recent Advances and Future Trends,”
IEEE Internet of Things Journal ,vol. 6, no. 2, pp. 2241-2263, Apr. 2019.[20] M. Chen, W. Saad, and C. Yin, “Echo-Liquid State Deep Learning for360 Content Transmission and Caching in Wireless VR Networks WithCellular-Connected UAVs,”
IEEE Transactions on Communications , vol.67, no. 9, pp. 6386-6400, Sept. 2019.[21] H. Wu, X. Tao, N. Zhang, and X. Shen, “Cooperative UAV Cluster-Assisted Terrestrial Cellular Networks for Ubiquitous Coverage,”
IEEEJournal on Selected Areas in Communications , vol. 36, no. 9, pp. 2045-2058, Sept. 2018.[22] M. Chen, M. Mozaffari, W. Saad, C. Yin, M. Debbah, and C. S.Hong, “Caching in the Sky: Proactive Deployment of Cache-EnabledUnmanned Aerial Vehicles for Optimized Quality-of-Experience,”
IEEEJournal on Selected Areas in Communications , vol. 35, no. 5, pp. 1046-1061, May 2017.[23] S. Suman, S. Kumar and S. De, “Path Loss Model for UAV-AssistedRFET,”
IEEE Communications Letters , vol. 22, no. 10, pp. 2048-2051,Oct. 2018.[24] Z. Hu, Z. Zheng, L. Song, T. Wang and X. Li, “UAV Offloading: Spec-trum Trading Contract Design for UAV-Assisted Cellular Networks,”
IEEE Transactions on Wireless Communications , vol. 17, no. 9, pp.6093-6107, Sept. 2018.[25] D. Athukoralage, I. Guvenc, W. Saad and M. Bennis, “Regret BasedLearning for UAV Assisted LTE-U/WiFi Public Safety Networks,”
IEEEGlobal Communications Conference (GLOBECOM) , Washington, Feb.2017 pp. 1-7. [26] S. Zhu, L. Gui, N. Cheng, F. Sun and Q. Zhang, “Joint Design of AccessPoint Selection and Path Planning for UAV-Assisted Cellular Networks,” IEEE Internet of Things Journal , vol. 7, no. 1, pp. 220-233, Jan. 2020.[27] E. Recayte, F. Lazaro, and G. Liva, “Caching at the Edge withFountain Codes,” in Proc.
Advanced Satellite Multimedia Systems Con-ference and the Signal Processing for Space Communications Workshop(ASMS/SPSC) , Berlin, Oct. 2018, pp. 1-6.[28] D. Ko, B. Hong, and W. Choi, “Probabilistic Caching Based on Maxi-mum Distance Separable Code in a User-Centric Clustered Cache-AidedWireless Network,”
IEEE Transactions on Wireless Communications ,vol. 18, no. 3, pp. 1792-1804, Mar. 2019.[29] K. Shanmugam, N. Golrezaei, A. G. Dimakis, A. F. Molisch and G.Caire, “FemtoCaching: Wireless Content Delivery Through DistributedCaching Helpers,”
IEEE Transactions on Information Theory , vol. 59,no. 12, pp. 8402-8413, Dec. 2013.[30] D. Ren, X. Gui, K. Zhang, and J. Wu, J., “Hybrid CollaborativeCaching in Mobile Edge Networks: An Analytical Approach,”
ComputerNetworks , vol. 158, pp.1-16, July 2019.[31] Z. Chen, J. Lee, T. Q. S. Quek and M. Kountouris, “Cooperative Cachingand Transmission Design in Cluster-Centric Small Cell Networks,”
IEEETransactions on Wireless Communications , vol. 16, no. 5, pp. 3401-3415,May 2017.[32] P. Lin, Q. Song, J. Song, A. Jamalipour and F. R. Yu, “CooperativeCaching and Transmission in CoMP-Integrated Cellular Networks UsingReinforcement Learning,”
IEEE Transactions on Vehicular Technology ,vol. 69, no. 5, pp. 5508-5520, May 2020.[33] Z. HajiAkhondi-Meybodi, M. S. Beni, K. N. Plataniotis, and A. Mo- hammadi “Bluetooth Low Energy-based Angle of Arrival Estimationvia Switch Antenna Array for Indoor Localization,”
InternationalConference on Information Fusion , July 2020.[34] Z. HajiAkhondi-Meybodi, M. S. Beni, A. Mohammadi, and K. N.Plataniotis, “Bluetooth Low Energy-based Angle of Arrival Estimationin Presence of Rayleigh Fading,” Accepted in
IEEE InternationalConference on Systems, Man, and Sybernetics , 2020.[35] C. M. Albertsen, “Generalizing the First-Difference Correlated RandomWalk for Marine Animal Movement Data,”
Scientific Reports , vol. 9,no. 1, pp. 4017–4031, Mar. 2019.[36] L. Qiu and G. Cao, “Popularity-Aware Caching Increases the Capacity ofWireless Networks,”
IEEE Transactions on Mobile Computing , vol.19,no. 1, pp. 173–87, Jan. 2019.[37] Y. Choi, C. S. Kim and S. Bahk, “Flexible Design of Frequency ReuseFactor in OFDMA Cellular Networks,”
IEEE International Conferenceon Communications , Istanbul, 2006, pp. 1784-1788.[38] A. M. Al-Samman, T. A. Rahman, T. Al-Hadhrami, A. Daho, M. N.Hindia, M. H. Azmi, K. Dimyati, and M. Alazab, “Comparative Studyof Indoor Propagation Model Below and Above 6 GHz for 5G WirelessNetworks,”
Electronics , vol. 18, no. 1, Jan. 2019.[39] T. Kanungo, D. M. Mount, N. S. Netanyahu, C. D. Piatko, R. Silvermanand A. Y. Wu, “An efficient k-means clustering algorithm: analysis andimplementation,”
IEEE Transactions on Pattern Analysis and MachineIntelligence , vol. 24, no. 7, pp. 881-892, July 2002.[40] T. Zhang, Y. Wang, Y. Liu, W. Xu and A. Nallanathan, “Cache-enablingUAV Communications: Network Deployment and Resource Allocation,”