Smart Soft-RAN for 5G: Dynamic Resource Management in CoMP-NOMA Based Systems
Mohammad Moltafet, Sepehr Rezvani, Nader Mokari, Mohammad R. Javan, Eduard A. Jorswieck
11 Smart Soft-RAN for 5G: Dynamic ResourceManagement in CoMP-NOMA Based Systems
Mohammad Moltafet, Sepehr Rezvani, Nader Mokari, Mohammad R. Javan, andEduard A. Jorswieck
Abstract
In this paper, we design a new smart software-defined radio access network architecture which isflexible and traffic and density aware for the fifth generation (5G) of cellular wireless networks andbeyond. The proposed architecture, based on network parameters such as density of users and systemtraffic, performs five important tasks namely, dynamic radio resource management (RRM), dynamic BStype selection, dynamic functionality splitting, dynamic transmission technology selection, and dynamicframing. In this regard, we first elaborate the structure of the proposed smart soft-RAN model andexplain the details of the proposed architecture and RRM algorithms. Next, as a case study, based onthe proposed architecture, we design a novel coordinated multi point beamforming technique to enhancethe throughput of a virtualized software defined-based 5G network utilizing the combination of powerdomain non-orthogonal multiple access and multiple-input single-output downlink communication. Indoing so, we formulate an optimization problem with the aim of maximizing the total throughputsubject to minimum required data rate of each user and maximum transmit power constraint of eachmobile virtual network operator and each BS, and find jointly the non-orthogonal set, beamforming,and subcarrier allocation. To solve the proposed optimization problem, based on the network density,we design two centralized and semi-centralized algorithms. Specifically, for the ultra-dense scenario,we use the centralized algorithm while the semi-centralized one is used for the high and moderatedensity scenarios. Numerical results illustrate the performance and signaling overhead of the proposedalgorithms, e.g., taking computational limitations into account the number of supported users is increasedby more than 60%.
Index Terms–
Software-defined radio access network, non-orthogonal multiple access (NOMA),coordinated multi point (CoMP).
M. Moltafet, S. Rezvani, and N. Mokari are with ECE Department, Tarbiat Modares University, Tehran, Iran. M. R. Javanis with the Department of Electrical Engineering, Shahrood University of Technology, Shahrood, Iran. Eduard A. Jorswieck iswith the Dresden University of Technology, Communications Laboratory, Chair of Communication Theory, Dresden, Germany. a r X i v : . [ c s . I T ] A p r I. I
NTRODUCTION
The way of evolution towards fifth generation (5G) has two main branches: evolution ofnetwork architecture and evolution of communications technologies. The network should bedesigned in such a way to be able to dynamically change its architecture and the communica-tions technologies. In such a flexible architecture, huge amount of signaling and computationalresources would be needed to optimally manage the network resources. Even the optimal solutionwould not be achievable and only a low complexity suboptimal solution could be attained.
A. Recent RAN Architectures
Recently, various RAN architectures have been developed from three main perspectives.The first is a new air interface architecture by means of separating signaling and data tohave an efficient and flexible radio resource management (RRM) for capacity boosting andenergy saving. The second one is RAN mode selection with renovating it into massive BSswith a centralized baseband processing. This perspective motivates us to embed the cloud-basedbaseband processing pool with remote radio heads (RRHs) and BS functions virtualization. Thethird one is separating the control plane from data plane in order to have an efficient centralizedRRM with a global view of the network. In this line, software-defined networking (SDN) hasbeen developed in which all network elements are under the control of a central scheduler.In the context of signalling-data separation, the hyper-cellular architecture (HCA) [1] and the“Phantom Cell” concept [2] are developed. In these architectures, the coverage of the network isdivided into two separated layers as control and traffic. Specifically, in HCA, all BSs are turnedinto two different types as control base stations (CBSs) and traffic base stations (TBSs). CBSs areresponsible for control coverage which mainly provides the information broadcasting. Besides,TBSs take care of data servicing to active users. In this architecture, TBSs can be switchedon/off to save energy. On the other hand, CBSs has responsible for globally optimizing the TBSmode selection and RRM. Some research efforts in both the academia and industry study theperformance gains achieved by performing cloud computing technologies in RANs. In this area,the wireless network cloud [3] and cloud-RAN (C-RAN) architectures are the most popular. Theconcept of C-RAN, sometimes referred to as centralized-RAN, was first introduced by ChinaMobile Research Institute in 2009 in Beijing, China [4], where multiple RRHs are distributed overa geographical location and are centrally controlled by a pool of baseband processing units (BBU)which are shared among cell sites [3], [5]. This architecture causes more reduction in the networks cost by lowering the energy consumption compared to the traditional architecture. However, thefull centralized manner of C-RAN entails more signaling between the RRHs and BBU pool andimposes more pressure on the fronthaul connections which increases the latency and decreasesthe throughput [6], [7]. In order to overcome the constrained fronthaul and backhaul capacitiesin ultra-dense heterogeneous networks (HetNets) and C-RANs, a new architecture known as theheterogeneous cloud radio access network (H-CRAN) is presented in [8], [9]. H-CRAN takes fulladvantages of both the HetNets and C-RANs. In H-CRAN, low-power RRHs cooperate with eachother to achieve more gains of cooperation. In addition, the BBU pool is interfaced with highpower nodes (HPNs) to mitigate the cross-tier interference between RRHs and HPNs. HPNs arealso responsible to guarantee the backward compatibility and coverage while low power nodes(LPNs) are mainly deployed to support the throughput [9]. In this system, a few functionalities areconfigured in RRHs while most important functionalities are processed in the BBU pool and othercommunication functionalities from the physical to network layers are left to HPNs [8], [9]. Fog-RAN is another cloud-based architecture which extends the traditional cloud computing paradigmto the edge of the network to overcome the disadvantages of C-RAN and H-CRAN such as highround-trip-time of delay-sensitive applications in dense HetNets with confined channel capacities[10]. In this system, the traditional RRHs in C-RANs evolved to the fog-computing-based accesspoint which is equipped with a collaboration radio signal processing, cooperative radio resourcemanagement, and considerable caching and computing capabilities [10]. The emergence SDNtechnology enables separation of the control plane from the data plane, centralized controllingby means of connected switches and routers to all networks elements and software applicationsprogramming interfaces to RANs. In this line, SoftRAN is a software defined (SD) centralizedcontrol plane for RAN which abstracts all BSs in a virtual big BS consisting of a central controllerand radio elements (individual physical BSs) [6], [11]. Recently, some research works investigatethe integration of above trends in future RANs. Open-RAN is another software-defined RANarchitecture via virtualization which is firstly proposed in [12]. This architecture is a virtualizedprogrammable system which makes RAN more open, controllable, and flexible. CONCERT [13]is another RAN architecture which converges the cloud computing and cellular systems based ondata-control decoupling. Moreover, the concept of software-defined fronthaul in SD-based cloud-RANs is proposed in [14] and the combination of SoftRAN and the data-signaling separation isproposed in [15]. SoftAir is a SD-based system proposed in [16] for 5G wireless networks. Inthe SoftAir system, the control plane which is placed in the networks server is responsible for network management and optimization while the data plane consists of software-defined BSs inRAN and software-defined switches in the core network. This architecture also takes the novelideas of cloud-based network function centralization and network function virtualization andprovides a scalable and flexible network management and mobility-aware traffic load balancing[16]. Software-defined hyper-cellular architecture design is based on the integration of cloudRAN, SDN, and air interface separation [17]. This system is divided into three subsystemsknown as: 1) RRH network; 2) fronthaul network; 3) virtual BS cloud [17]. RRHs are merelyresponsible for RF transmission/reception, or some baseband processing functions. They canalso be dynamically configured as control BS, traditional, or put into sleep mode based on thenetwork status and their capabilities. A novel architecture is proposed in [18] which is based onthe deep integration of software defined and virtualized RANs with fog computing which is agood solution for real-time data services. The SDN controller can operate in three models as:centralized, distributed, and hybrid models based on the network status. In order to deal withthe high latency of Soft-RAN systems, the hierarchical software-defined RAN is developed.Against the virtualizing all BSs as a single centralized control BS in Soft-RAN, this architecturehas multiple clusters of BSs where in each cluster there is a virtual local controller which isresponsible for the located BSs in the cluster. In addition, BSs and the centralized controller ineach cluster are connected via the fronthaul links [19].
B. Communications Technologies
In addition to the RAN architecture, the transmission technologies used in RAN has importanteffect on the efficient resource management. Various technologies such as non-orthogonal multi-ple access (NOMA) [20], [21], CoMP, and (massive) multiple input multiple output (MIMO) areproposed to address the existing challenges of 5G networks. NOMA techniques are introduced asa promising candidate in which the same spectrum can be used by more than one user in a non-orthogonal way. The non-orthogonal use of the frequency band introduces an extra interferencecompared to the orthogonal multiple access (OMA) scheme; however, if the resulting interferenceis controlled in an appropriate way, the penalty of non-orthogonal usage is reduced while therewould be an increase in spectral efficiency. Power domain NOMA (PD-NOMA) is introducedas a multiple access technique for 5G of the cellular networks. In this technique, the transmitterapplies superposition coding (SC) meaning that the transmitted signal is superimposed of thesignals of users sharing that frequency bands. The users in the same band are sorted basedon some criteria (e.g. channel quality or receive SNR), and each user applies the successive interference cancellation (SIC) to the received signal to cancel the interference of worse userswhile it treats the signals of better users as noise. Moreover, one of the main limitations ofcellular networks is the interference produced by reusing the same frequency band among users.Therefore, advanced techniques are needed to mitigate this interference. One of the promisingschemes is multiple antenna transmission leading to massive MIMO systems whose advantage isthe use of spatial diversity and multiplexing. Note that in HetNets, there are many BSs each withdifferent capability which share the same spectrum. This implies that the transmission of each ofthem affects the quality of the others, and the system is strongly coupled. To alleviate the effect ofinterference, coordination among transmitting points is important. In this scheme, called CoMP,multiple transmitters are coordinating to implement the distributed antenna systems. Sometimes,the coordinating points construct a distributed antenna system. These transmitters could performbeamforming to decrease the harmful effect of interference.
C. Related Literature
The architectural evolution goes towards more flexibility. The network should be able todynamically adopt the proper resource management strategies from centralized to the fullydistributed manners. The functions provided by the network could be available in a centralentity or some of them could be relegated to the BSs. Some BSs could be switched off for thesake of energy conservation. In addition, the choice for the adopted transmission technologycould take into account the dynamics of the environment as well as the users density and theirtraffic volume. In the following we review the recently published works in these areas.
1) Dynamic RRM:
RAN Architectures:
In SD cellular networks, multiple centralized RRMalgorithms are proposed based on the received global information of the networks [22], [23].In [22], the authors propose a centralized RRM algorithm in multicell downlink orthogonal fre-quency division multiple access (OFDMA) systems to maximize the throughput of the network.They also jointly consider the carrier aggregation and coordinated multi point (CoMP) techniqueswhich can significantly improve the performance. Another algorithm is also proposed in [24]to find both the user association and bandwidth allocation and cache refreshment strategiesin SD-based virtualized information centric network. The integration of the device-to-devicecommunication, SDN, and network function virtualization are investigated in [25]. In [26], aninformation centric virulalization network in SDN is considered and the data delivery path isestablished based on the effective capacity maximization. The optimal power allocation forcontent caching in SDN wireless networks by considering the effective capacity as the objective is obtained in [23] and the effect of the delay-quality of service (QoS) on power allocationand the gain from content caching are evaluated. Due to the various services developed for Gnetworks, a new RAN architecture is needed to be designed which is smart and flexible enoughto undergo necessary changes when are required. None of the presented RAN architectures haveenough flexibility and smartness to meet the G demands. G Technologies:
Recently, several works are published with the aim of combining CoMPwith PD-NOMA in the downlink of wireless networks [27]–[31]. In [27], the authors utilizeAlamouti code to improve the cell-edge users throughput in the downlink of a CoMP-NOMAcellular network consisting of two coordinated BSs where each cell has two users. In [27], eachcell performs SIC in which the cell-center users are assumed to be non-CoMP. The authorsin [28] investigate the design of an opportunistic CoMP-NOMA scheme to improve both theusers throughput and the outage probability. Specifically, they investigate the design of the jointmulti-cell power allocation algorithm. In [29], the authors propose a power allocation algorithmto maximize the energy efficiency in downlink of CoMP-NOMA systems. In this work, thenetwork throughput is evaluated under three transmission schemes: 1) coordinated BSs transmitsignals to all users; 2) Coordinated BSs transmit signals only to the cell-edge users; 3) thereceived signal at each user is transmitted by only one BS. In [30], a CoMP-NOMA system isinvestigated in which each transmitter and receiver have multiple antennas. Furthermore, in [31],a distributed power allocation is evaluated for the downlink of CoMP-NOMA cellular systemswith the view of spectral efficiency in which the power allocation is adapted independentlyat each cell for the active users. The authors in [32] study the problem of precoding in thedownlink of a multiple input single output (MISO) system with the objective of maximizing thesum rate while simultaneously satisfying the NOMA constraints. In [33], the authors proposea single cell MIMO based system in which by applying the PD-NOMA technology, multipleusers can send their signals simultaneously with the objective of maximizing the system sumrate. To solve the proposed problem, they used two methods, namely, 1) suboptimal solutionwith low complexity, 2) optimal solution with high complexity. The authors of [34] study arobust PD-NOMA scheme for the MISO system to maximize the worst case achievable sum ratewith a total transmit power constraint. In [35], the authors propose a user clustering and zeroforcing beamforming scheme for a downlink communication of single carrier PD-NOMA basedsystems. The authors of [36] investigate a power allocation problem to maximize sum rate in aMIMO based system considering the PD-NOMA technology. In [37], the authors study Massive
MIMO technology in a PD-NOMA based system. The authors of [38] present a comparisonstudy between PD-NOMA and sparse code multiple access from the throughput and complexityaspects. Dynamic RRM and user association based on the population density and traffic status ofthe network is an efficient method to improve the system performance. Moreover, dynamic RRMneeds the network structure with high flexibility, and therefore, to implement it, the SD-basednetworks should be exploited. None of the previous works use the SDN technology to managethe radio resources based on the network density and traffic status of the network.
2) Dynamic BS Type Selection:
In [39]–[41], the dynamic BS mode selection are studied.In [39], the authors propose a dynamic BS sleeping scheme, where BSs dynamically are turnedinto sleep mode based on the traffic status of the network, under a SD-based central controller.The main purpose of the scheme is reducing the energy consumption of the network. Besides,the authors of [40] propose a dynamic BS switching scheme in cellular wireless access networksbased on the received traffic profile to have an efficient energy saving scheme for reducing thesystems cost. In the proposed scheme, they believe that the utilization of all BSs can be veryinefficient during off-peak time. Moreover, some dynamic BS on/off switching strategies with theaim of minimizing the energy consumption in wireless cellular networks is investigated in [41],where the authors formulate a combinatorial optimization problem with a high computationalcomplexity and signaling overhead. Then, to reduce the computational complexity, they proposea distributed manner and three heuristic algorithms with low signaling overheads. Althoughthe prior works can significantly improve the energy efficiency of the system, they neglect theconsideration of different BS type selection based on the network conditions. As discussedabove, in the BS type selection, each BS can be turned into CBS, data BS, traditional BS orother existing types and also can be turned into the sleep mode, based on the network conditions.
3) Dynamic Framing:
In 5G, wireless communications will be highly heterogeneous in someaspects as service types, propagation environments, and device types. To tackle this heterogeneityin physical layer, the network should be reconfigurable in frame design based on the diverseservice requirements and the degrees of freedom for control signaling elements [42]–[45]. In LTEand other existing wireless communication systems, the multicarrier modulations are restricted toa single predefined subcarrier spacing. Specifically, LTE applies 15 kHz subcarrier spacing with1 ms transmission time interval (TTI) [42]–[45]. Although this short range subcarrier spacingworks well for LTEs propagation environment, it is very difficult to make it work in numerousphysical properties of 5G operating in very high frequency, e.g., millimeter wave (mmWave) [44], [45]. Besides, the frequency drift (like Doppler shift) happens in frequency operation [43],[45]. For example, if a high frequency range, e.g., 26 GHz, is used, it is more than several tensof KHz [45]. In addition, 5G is supporting velocities up to 500 Km/h which can not be handledwith LTE pilot density in time [43]. Accordingly, 5G needs different numerologies in the sameOFDM modulation with much larger subcarrier spacing than that of current LTE [42], [43].
D. Contributions
The contributions of this paper are twofold. First, we propose a new software-defined RANarchitecture which is traffic and density aware. The proposed RAN architecture is smart andflexible enough to take the traffic and density of users in the network into consideration forchoosing the appropriate resource management approach which can improve the energy effi-ciency and spectral efficiency with controllable complexity. Second, for our proposed SD-RANarchitecture, we consider the case of multi infrastructure providers (InPs) and multi virtualnetwork operators (MVNOs), and design a novel optimization problem with a novel CoMP-NOMA model considering the MISO technology. The work presented in this paper is the firststepping-stone towards several potential research directions.
1) The Proposed
Smart Soft-RAN
Architecture:
We propose a flexible RAN architecturewhich is smart and is able to make the necessary changes in response to the dynamics of thenetwork. The proposed
Smart Soft-RAN architecture is able to perform the following five tasks: • Dynamic RRM: This task is based on the network density and selects one of the three typesof RRM as, Software-Defined Centralized Resource Management (SD-CRM), Software-Defined Semi-Centralized Resource Management (SD-SCRM), and Software-Defined LocalResource Management (SD-LRM). • Dynamic functionality splitting: It can be implemented in the cloud-based networks. Withthis task, in order to balance the processing load of BSs and decrease delay, based onthe network situation and users’s demand, the functionality of the cloud can be abstractedamong BSs. • Dynamic BS type selection: In this task, based on the network situation, each BS isdynamically scheduled to turn into a certain mode of operation, such as control BS, RRH,data BS, traditional BS or sleep mode. • Dynamic technology selection: This task, based on the network conditions, selects theappropriate access, fronthaul and backhaul technologies such as 1) multiple access type, e.g., OFDMA, PD-NOMA, sparse code multiple access (SCMA), 2) connectivity modee.g., dual connectivity or multi connectivity, 3) relay type e.g., decode and forward andamplify and forward, 4) MIMO type, etc. • Dynamic framing: In this task, based on the target service type, the predefined QoS,users speeds, and the serving environment characteristics, flexible numerology with varioussubcarrier spacing, TTI, etc is deployed. Via this method which is first introduced in Release15 of the 5G standard, the spectrum and energy efficiency can be significantly improved andthe QoSs of new emerging 5G multi-service systems can be practically satisfied. Moreover,this technology can be regarded as a key solution for rapid traffic variations, specifically indense deployments with a small number of users per-BS.We provide the description of the proposed RAN and the functionality of each part. We providedetails on how the above tasks are performed by the proposed RAN in an integrated manner.
2) Resource Management in a CoMP-NOMA Based Network:
After elaborating on the pro-posed RAN architecture, we evaluate its performance for a communication network scenario.In other words, as a case study in the context of the proposed RAN, we consider a cellularmulticarrier HetNet in which NOMA is used as the transmission technology and the BSscoordinate with each other for interference management. • We design a novel density-aware beamforming, subcarrier assignment, and user associationscheme for the downlink of the considered network which maximizes the system sum ratewith a constraint on the minimum requested rate of each MVNO and constraints on thetransmit power and subcarrier allocations. In such a system, based on the traffic and densityinformation of the network, centralized or semi-centralized resource allocation is adopted. • We consider a virtualized case where the physical resources provided by several InPs aredivided into several virtual resources each of which could be used by one mobile MVNO. • To improve the performance of the considered system, we utilize the CoMP technologyin a PD-NOMA based system with MISO communication. In the proposed system modelbecause of the PD-NOMA technique there are various challenges in modeling the CoMPtechnology such as CoMP set, SIC ordering of PD-NOMA, and user association. To tacklethese challenges, we design a new CoMP model in the PD-NOMA based systems. • To solve the proposed optimization problem, based on the network density we propose twosolution algorithms as centralized and semi centralized methods. In the proposed solutionsthe functionality of each elements of networks are determined and investigated. Moreover, the signaling overhead of each solution is investigated.We note that in the case study section, from the mentioned five tasks where the proposed smart soft-RAN is able to perform, we focus on the dynamic RRM task and leave the study ofother tasks as future works.II. D ESCRIPTION OF THE P ROPOSED
Smart Soft-RAN
With the advent of the new RAN architectures and emerging transmission technologies,resource management become more challenging. Selecting appropriate RAN architecture toresponse the time variant density and traffic volume of the network can be a major and crucialquestion. We may need a new RAN architecture which is flexible , traffic, and density aware .The architecture and the role of the network nodes as well as their capabilities should evolve inresponse to change in traffic and user density of the network. In Fig. 1, an example of the mainstructure of the proposed RAN architecture is shown. The physical radio access infrastructure,denoted by radio access, is provided by several infrastructure providers. Note that these physicalinfrastructures, which are mainly the BSs and their backbone connection, are the connectingpoints of wireless users to the network. As shown in Fig. 1, in our proposed model, there areseveral types of BSs in the network which provide signalling and data coverage from the users,a pool of BBUs providing centralized base band processing, SDN controller which controls thenetwork operation by properly programming the network elements’ functionalities, hypervisorswhich are responsible for virtualization of the networks, and applications running on top ofthe network. For fully centralized case, all the base band processing is performed in the BBUpool and the BS is called RRH. However, some functions of BBU could be shifted to the BSsresulting in the so-called remote radio systems (RRSs). Note that in the figure, we have RRS1and RRS2 which means these BSs have different functionalities. Depending on the amount ofthe functionality in BSs, their abilities and roles could be different. It is the role of the SDNcontroller, which separates the data and control plane, to change the capabilities of the network’selements and make the network programmable. Indeed, it is the SDN controller’s responsibility toprogram the BSs and change their functionalities. On the other hand, a hypervisor is responsibleto virtualize the network into several virtual networks each of which is allocated to one virtualnetwork provider. Note that both the forwarding infrastructure, i.e., BSs, and processing andcontrol infrastructures, BBUs and SDN controller, could be virtualized as shown in Fig. 1. TheBSs abilities could change from just performing radio transmission in ultra dense scenarios(BBU pool is responsible for base band processing) and providing data forwarding to perform I n P N I n P Data BS(RRH) Traditional BS MobileUser Base Band
Processing Server
Hypervisor U l t r a D e n s i t y R a d i o A cce ss E nd U s er D a t a P l a n e C o n t r o l P l a n e V i r t u a li z a t i o n L ay er Virtual SDNController L o w D e n s i t y M o d er a t e D e n s i t y H i g h D e n s i t y I n P Data BS(RRS 1) Data BS(RRS 2) Offline
Mode VN V VN U l t r a D e n s i t y L o w D e n s i t y M o d er a t e D e n s i t y H i g h D e n s i t y VN VN V C o n t r o l P l a n e VN C o n t r o l P l a n e VN C o n t r o l P l a n e Control BS (CBS) SDNController BB U S DN C o n t r o ll er N e t w o r k A pp li c a t i o n NetworkApplication P r o ce ss i ng M a n a g e m e n t C l oud ( P M C ) C o n t r o l A cce ss Fig. 1: The main structure of the proposed smart soft-RAN model. base band processing and control coverage (low density situations). Since a BS is virtualized toseveral virtual BSs, each virtual BS could have different properties. Note that, the BS’s abilitycould change over time or even it could be turned off. This smart way of the network changingthe network architecture and functionalities could be under the control of the SDN directly oras an outcome of the resource allocation problem. In the proposed traffic/density-aware smartSofRAN architecture, the SDN controller is responsible for network RAN architecture selectionas well as the RRM type decision. In other words, the SDN controller firstly takes the trafficload status of the network and density of users. Then, it decides on the RAN architecture, i.e., Traffic Status Number of Active Users QoS requirements Service Type Available Radio and Physical Resources
SDN Controller Density of Active Users Signaling Overhead Computational Complexity Capacity of Fronthaul and
Backhaul Links Operation Cost RRM Type Selection
Network Information BBU/BS RRM BS Type Selection Technology Selection Function Splitting Frame Type Selection
Preprocessing
Fig. 2: The process of joint network RAN architecture and RRM type selection. low density, moderate density, high density, and ultra density architectures, based on the severalproperties such as the signalling load between BSs and cloud, constrained capacity of fronthaullinks, and the delay of queuing and SDN controller. Besides, it also makes a decision whichRRM type, i.e., centralized, semi-centralized, or distributed, should be applied based on theirperformance, computational complexities, and the selected RAN architecture type. The SDNcontroller considers the criteria and applies a pre-processing on the input parameters, and basedon the obtained results, selects an appropriate RAN architecture and RRM type. The process ofjoint network RAN architecture and RRM type selection by the SDN controller is also illustratedin Fig. 2. Note that, the network is dynamic, i.e., that the user density as well as the traffic acrossthe network is different and could change over the time. This means that the resource managementshould be dynamic in the sense that based on the network conditions, the resource managementshould change the RAN architecture and the transmission technology; it should perform therequired changes as fast as possible which requires to be of low complexity. It should be flexibleenough to incorporate changes resulting from new technologies or new management policieswith less amount of hardware change. These requirements lead to a programmable architecturewhere the network operations are software defined. The proposed
Smart Soft-RAN is able toadopt three types of resource management based on the network status: • SD-CRM: In the SD-CRM, BBU pool is responsible for baseband processing while RRHis responsible for RF functions. Some other BSs are also responsible for providing controlcoverage. Power and subcarrier allocation, user associations, adopted transmission, accesstechnologies, virtualization, and slicing are performed centrally in a central controller. SD-CRM is more suitable for the case where the users density and traffic volume is high(ultra dense scenarios). This is because, when the number of users increases, the processingload of RRHs increases, too. Hence, more bandwidth and power should be considered for control signals. On the other hand, satisfying the QoS at end-users is more critical for theoperators, because of the confined radio resources, which lead them to use more efficientresource management algorithms. Generally, centralized resource management approacheshave more gains, compared to semi-centralized or distributed ones. To this end, to save morebandwidth and power resources for transmission data in ultra-dense scenarios, we use SD-CRM algorithms. Note that SD-CRM algorithms inherently increase the signalling overheadand the computational complexity in the network, specifically in ultra-dense scenarios. But,in the proposed Smart Soft-RAN architecture, by utilizing the software defined approachand powerful processors at BBU pool, we have a fast and efficient resource management. • SD-SCRM: Here, we move away from centralized architecture by letting the BS performsome base band processing. For example, BSs choose the set of connected users and performpower and subcarrier allocation. However, proper transmission technology as well as theaccess technology could be determined in a centralized manner. SD-SCRM algorithms aremore suitable for high and moderate density scenarios, where the density of users andtheir corresponding traffic status are sufficiently decreases, compared to the ultra-densescenario. In these situations, with decreasing the scale of the network, the difference betweencentralized and semi-centralized resource management algorithms is decreased. Since RRHsare responsible for resource management and certain other functionalities, the processingload of RRHs increases. RRHs have much smaller processing capacities than the BBUpool. Hence, the computational complexity is a more important factor when choosing anefficient algorithm compared to the performance, generally, which leads us to use SD-SCRMalgorithms which have lower computational complexity, compared to SD-CRM approaches. • SD-LRM: In this type of resource management, BSs with traditional architecture are re-sponsible for both the control and traffic signals and all the base band processing and RFfunctions are performed in the BS. In this case, BSs perform the task of resource allocationand management locally based on the available local information in a distributed manner.We note that each of the above mentioned resource management schemes could be adoptedfor a part of the network in a dynamic manner. In contrast to the traditional RANs with afixed subcarrier spacing, our proposed
Smart Soft-RAN supports new radio (NR) technologywith dynamic subcarrier spacing and TTI tuning in which the spectrum and time slot dura-tion management, inter-cell interference modeling, synchronization, multiple access schemes, F re qu e n c y Time F re qu e n c y Time 𝑡 𝑡 𝑓 𝑓 𝑓 𝑓 𝑡 𝑡 𝑓 𝑓 𝑓 𝑓 F re qu e n c y Time F re qu e n c y Time
Interference of Cell B to Cell A
Interference of Cell A to Cell B 𝑡 𝑡 𝑓 𝑓 𝑓 𝑓 𝑡 𝑡 𝑓 𝑓 𝑓 𝑓 Fig. 3: Scheduled subcarrier spacing and TTI at each user with inter-cell interferences. and the resource management format have considerable fundamental changes. Specifically, forthe spectrum and time slot duration management perspective,
Smart Soft-RAN has a flexiblesubcarrier spacing based on the corresponding frequency range and wireless bandwidth. In thiscase, the network schedular is able to choose the appropriate subcarrier bandwidth based on thetarget service type and users’ velocity. The inter-cell interferences should be modeled carefullysuch that tackles the overlapping of different time slot durations and subcarrier bandwidthsin multi-cell scenarios. To be more specific, we show an exemplary system with 2 cells eachof which consists of 4 users, where each user is scheduled based on specific subcarrier bandand TTI duration in Fig. 3. As shown in Fig. 3, users in the same frequency range and timeinterfere with each other. For example, user in cell A interferes with users 1 and 2 in cell B in TTI [0 t ] over frequency bands [ f f ] and [ f f ] , respectively. This situation whichis named as partial overlapping needs an exact interference modeling, due to heterogeneity ofsubcarrier bandwidths and TTIs. Moreover, user 4 in cell B only interferes with user 4 in cellA in all the assigned frequency bands and TTIs, since both of them are scheduled in the samefrequency band and TTI. This situation is also named as full overlapping. On the other hand, thenetwork synchronization needs more signalling, since new subcarrier bandwidth allocation andsubcarrier type selection variables are added in the system. Besides, some fundamental changesare needed to be imposed to apply multiple access schemes, such as pattern division multipleaccess (PDMA) and SCMA technologies. In PDMA and SCMA, both the codebook designingand detection algorithms should be configured based on the various subcarrier bandwidths andTTIs, specifically in the partial overlapping scheme. The fundamental changes, which should beapplied in resource allocation perspective due to performing the dynamic framing technology,are presented in Table I. TABLE I: Challenges of
Smart Soft-RAN with dynamic framing
Items CharacteristicsSpectrum and time slot durationmanagement • Smart Soft-RAN : Flexible Subcarrier Spacing (15 KHz, 30 KHz, 60 KHz, 120 KHz and 240 KHz) • Traditional RANs: Fixed Subcarrier Spacing (15 KHz)Inter-cell interference modeling • Smart Soft-RAN : Dynamic mapping (partial overlapping) full inter-slot interference • Traditional RANs: Fixed mapping (full overlapping)Synchronization In
Smart Soft-RAN , due to the new subcarrier bandwidth allocation and subcarrier type selection variables there are more signaling.Multiple access schemes In
Smart Soft-RAN , PDMA and SCMA techniques from the aspects of codebook design and detection algorithm has more challenges.Resource management format
Smart Soft-RAN has more assignment variables (e.g. subcarrier parameter indicator) and new constraints (e.g. format selection) with respectto traditional RANs.
III. C
ASE S TUDY : V
IRTUALIZED C O MP-NOMA B
ASED H ET N ETS
In this section, we present a case study of the proposed
Smart SofRAN architecture with theaim of designing joint RRM and user association algorithms in a CoMP-NOMA based HetNet.
A. Network Model
We consider a scenario with multiple InPs and multiple MVNOs with the users of each MVNOspreading over the total coverage area of the network. We assume that each InP network consistsof a set of base stations in which the reuse factor is more than one, and InPs do not interfere witheach other. We assume that all the transmitters are equipped with multiple antennas, i.e., M T antennas, while the receivers are simply single antenna systems. We denote the set of InPs by i ∈ I = { , · · · , I } , the set of MVNOs by v ∈ V = { , · · · , V } , and the set of BSs of InP i by b i ∈ B i = { , · · · , B i } . The set of all users in the network is denoted by K = { , · · · , K } whichis the union of the sets of users of all the MVNOs, i.e., K = ∪ v ∈V K v . By utilizing the PD-NOMAtechnique, we assume that the total bandwidth of each InP network, which is non overlappingwith other InP networks, i.e., BW i , is divided into N i subcarriers of equal bandwidth each ofwhose bandwidth is less than the coherence bandwidth of the network channel. We denote thechannel gain from transmitter b i to receiver k over the subcarrier n i by h b i ,n i ,k ∈ C M T × where C is the complex field, and the beam vector assigned by transmitter b i to receiver k over subcarrier n i by w b i ,n i ,k ∈ C M T × . We define an indicator variable ρ b i ,n i ,k ∈ { , } with ρ b i ,n i ,k = 1 if user k is scheduled to receive information from transmitter b i over subcarrier n i , and ρ b i ,n i ,k = 0 if itis not scheduled to receive from transmitter b i over subcarrier n i . Assume that the informationsymbol s b i ,n i ,k is decided to be transmitted to user k from BS b i over subcarrier n i . We suppose BS 1 BS 2 BS 3
User 1 User 2 User 3 𝒉 𝑖 , 𝑛 𝑖 , 𝒉 𝑖 , 𝑛 𝑖 , 𝒉 𝑖 , 𝑛 𝑖 , 𝒉 𝑖 , 𝑛 𝑖 , 𝒉 𝑖 , 𝑛 𝑖 , 𝒉 𝑖 , 𝑛 𝑖 , BS 3
User 1 User 1
User 3
BS 1
BS 2
User 2
Data Transfer Link to User 1
Data Transfer Link to Users 2 and 3Interference Link to Users 1
Fig. 4: Left: A typical CoMP-NOMA based system consisting of BSs and users. The information signaltransmission links from BSs to users are represented by black arrows. Right: A typical CoMP-NOMA basedinterference system when user selects BS . that each user is assigned to at most one InP, and in each cell, each subcarrier can be assignedto at most L T users, which are, respectively, given by the following constraints: ρ b i ,n i ,k + ρ b j ,n j ,k ≤ , ∀ k ∈ K , i, j ∈ I , i (cid:54) = j, b i ∈ B i , b j ∈ B j , n i ∈ N i , n j ∈ N j . (1) (cid:88) k ∈K ρ b i ,n i ,k ≤ L T , ∀ i ∈ I , b i ∈ B i , n i ∈ N i . (2) B. Signal Model and Achievable Data Rates
In the following, we describe the principles of the considered CoMP-NOMA model in HetNets.We first illustrate the model using a simple example which is shown in Fig. 4-Left. As shown,we assume that there are BSs in InP i which has N i subcarriers and users, i.e., user 1, user2, and user 3, are connected to these three BSs in InP i in such a way that is shown in Fig.4-Left. Note that user 1 is connected to BSs 1 and 2 which means that these BSs constructthe set of BSs which perform CoMP for user 1. User 2 is connected to BSs 1, 2, and 3. Inaddition, user 3 is connected to BSs 2 and 3. BS 1 only sends the signal of user 1, BS 2 sendsthe signals of user 1, 2, and 3. And BS 3 sends the signals of user 2 and 3. For the NOMAtransmission, the SIC ordering should be determined. Note that BSs could send the informationof different users over the same subcarrier and different users could be connected to differentBSs for CoMP. In a real network with many users and BSs, it is difficult to find the NOMAset due to complicated coupling of users and BSs. Note that by the NOMA set, we mean thatusers in the set are ordered based on a specific SIC ordering to decode the other users’ signals,and hence the signal of other users, even if transmitted from the same BSs, is considered as noise. A simple way is to consider all the users as the NOMA set which is prohibitive in areal scenario. To overcome this difficulty, we define the SINRs of a user from the view point ofeach connected BSs and define the set of NOMA users for this case. For example, as shown inFig. 4-Right, we assume that user 1 is connected to BS 1. In our model, if we consider user 1,connected to BS 1, all other users’ signals should be considered as interference. The receivedSINR for user from the viewpoint of BS on subcarrier n i is given by γ i ,n i , = ρ i ,n i , | h H i ,n i , w i ,n i , | + ρ i ,n i , | h H i ,n i , w i ,n i , | I i ,n i , + N , (3)where N is the noise power and I i ,n i , = ρ i ,n i , | h H i ,n i , w i ,n i , | + ρ i ,n i , | h H i ,n i , w i ,n i , | + ρ i ,n i , | h H i ,n i , w i ,n i , | + ρ i ,n i , | h H i ,n i , w i ,n i , | . is the received interference at user fromthe viewpoint of BS on subcarrier n i . Now, consider user 2 which is connected to BSs 2 and 3.If we consider BS 2, since all users are connected to this BS, the NOMA set constraints all thethree users. In this case, user 2 could be able to decode and cancel the other users’ signals basedon the SIC ordering. For example, the SIC ordering of the form → → means that user 2is able to decode and cancel user 1 signal and user 3 is able to decode and cancel users 1 and 2signals. We may have several options for the SIC ordering. One option is to order the users basedonly on the channel gains of users in BS 2. In this case, if (cid:107) h i ,n i , (cid:107) ≥ (cid:107) h i ,n i , (cid:107) ≥ (cid:107) h i ,n i , (cid:107) ,then → → . Another option is the average channel gain of users from BSs to which theusers are connected. In this case, if ¯ h n i , ≥ ¯ h n i , ≥ ¯ h n i , where ¯ h n i , = ( (cid:107) h i ,n i , (cid:107) + (cid:107) h i ,n i , (cid:107) ) ,then → → . Another option is to consider the average gain of all channels between a userand BSs. In this case, we have ¯ h n i , = ( (cid:107) h i ,n i , (cid:107) + (cid:107) h i ,n i , (cid:107) + (cid:107) h i ,n i , (cid:107) ) although BS 3is not transmitting anything to user 1. Now consider the first option and the corresponding SICordering. The SINR of user 2 from the viewpoint of BS 2 is given by γ i ,n i , = ρ i ,n i , | h H i ,n i , w i ,n i , | + ρ i ,n i , | h H i ,n i , w i ,n i , | I i ,n i , + N , (4)where I i ,n i , = ρ i ,n i , | h H i ,n i , w i ,n i , | + ρ i ,n i , | h H i ,n i , w i ,n i , | . Note that there is nointerference from user 3 as we assumed that user 2 is able to decode and cancel its interference.Fig. 5 shows the considered scenario. For user 2, we can write the SINR from the viewpoint ofBS 3. Note that only users 2 and 3 are connected to BS 3, and hence the NOMA set consists ofusers 2 and 3. In this case, the signal of user 1 is considered as interference. Assuming option BS 1
BS 2
BS 3User 1
User 2
User 3
Data Transfer Link to User 2
Data Transfer Link to Users 1 and 3
Interference Link to User 2Canceled Interference Link at User 2
BS 1
BS 2
BS 3User 1
User 2
User 3
Data Transfer Link to User 2
Data Transfer Link to Users 1 and 3
Intercell Interference Link to User 2NOMA Interference Link to User 2
BS 3
BS 2
Fig. 5: A typical CoMP-NOMA based interference system. Left: when user selects BS , Right: when user selects BS . one as previous with (cid:107) h i ,n i , (cid:107) ≥ (cid:107) h i ,n i , (cid:107) which implies the SIC ordering → , the SINRof user 2 from the viewpoint of BS 3 is given by γ i ,n i , = ρ i ,n i , | h H i ,n i , w i ,n i , | + ρ i ,n i , | h H i ,n i , w i ,n i , | I i ,n i , + N , (5)where I i ,n i , = ρ i ,n i , | h H i ,n i , w i ,n i , | + ρ i ,n i , | h H i ,n i , w i ,n i , | + ρ i ,n i , (cid:107) h H i ,n i , w i ,n i , (cid:107) + ρ i ,n i , | h H i ,n i , w i ,n i , | . Note that the interference from user 3 is NOMA interference whilethe interference from user 1 is inter-cell interference. Hence, we can say that the interferenceexperienced by a user can be divided into two terms; The term I NOMA comes from the PD-NOMA technique which is the interference from users with higher order in SIC ordering. Theterm I Inter is the interference from all other users. This scenario is also illustrated in Fig. 5. In thissystem, the user’s connections to BSs, i.e., the CoMP set for users, are assumed to be fixed andknown. However, in general, it is an optimization problem and the resulting resource allocationhas to determine this connectivity. To mathematically state the proposed CoMP-NOMA modelin general case, we need to determine the SINR of users. For the general case, we write theSINR of user k from BS b i over the subcarrier n i as follows: γ b i ,n i ,k = (cid:80) b (cid:48) i ∈B i ρ b (cid:48) i ,n i ,k | h Hb (cid:48) i ,n i ,k w b (cid:48) i ,n i ,k | I NOMA b i ,n i ,k + I Inter b i ,n i ,k + N , (6)where I NOMA b i ,n i ,k = (cid:80) k (cid:48)∈K ,k (cid:48) >k (cid:80) b (cid:48) i ∈B i ρ b i ,n i ,k ρ b i ,n i ,k (cid:48) ρ b (cid:48) i ,n i ,k (cid:48) | h Hb (cid:48) i ,n i ,k w b (cid:48) i ,n i ,k (cid:48) | , and I Inter b i ,n i ,k = (cid:88) k (cid:48)∈K ,k (cid:48)(cid:54) = k (cid:88) b (cid:48) i ∈B i ρ b i ,n i ,k (1 − ρ b i ,n i ,k (cid:48) ) ρ b (cid:48) i ,n i ,k (cid:48) | h Hb (cid:48) i ,n i ,k w b (cid:48) i ,n i ,k (cid:48) | , (7) where k (cid:48) > k means that user k (cid:48) is of higher order than user k in SIC ordering based on thefollowing condition which is discussed above as the third option as k (cid:48) ( n i ) > k ( n i ) ⇐⇒ ¯ h n i ,k (cid:48) ≥ ¯ h n i ,k , where ¯ h n i ,k = B i (cid:80) b (cid:48) i ∈B i (cid:107) h b (cid:48) i ,n i ,k (cid:107) . As we know, we have to consider a specific SINR foreach user on each subcarrier which leads us to select only one viewpoint for each user on eachsubcarrier. To overcome the mentioned challenge, we introduce a new binary non-orthogonal set(NOS) selection optimization variable denoted by x b i ,n i ,k ∈ { , } , where if user k at subcarrier n i is on the viewpoint of BS b i , x b i ,n i ,k = 1 , and otherwise, x b i ,n i ,k = 0 . The data rate of user k at subcarrier n i on the viewpoint of BS b i is thus given by r b i ,n i ,k = log (1 + γ b i ,n i ,k ) . C. Problem Formulation
Here, we aim to design a joint the subcarrier allocation, NOS selection and beamformingstrategy to maximize the sum data rate of users. We propose an optimization problem tofind the binary subcarrier assignment and NOS selection variables, and beamforming method,simultaneously, as follows: max W , ρ , X (cid:88) i ∈I (cid:88) b i ∈B i (cid:88) n i ∈N i (cid:88) k ∈K x b i ,n i ,k r b i ,n i ,k (8a)s.t. : (cid:88) n i ∈N i (cid:88) k ∈K ρ b i ,n i ,k (cid:107) w b i ,n i ,k (cid:107) ≤ P b i max , ∀ i, b i ∈ B i , (8b) (cid:88) i ∈I (cid:88) b i ∈B i (cid:88) n i ∈N i (cid:88) k ∈K v ρ b i ,n i ,k (cid:107) w b i ,n i ,k (cid:107) ≤ P v max , ∀ v ∈ V , (8c) (cid:88) i ∈I (cid:88) b i ∈B i (cid:88) n i ∈N i x b i ,n i ,k r b i ,n i ,k ≥ R v min , ∀ v, k ∈ K v , (8d) ρ b i ,n i ,k | h Hb i ,n i ,k w b i ,n i ,k | ≥ ρ b i ,n i ,k ρ b i ,n i ,k (cid:48) | h Hb i ,n i ,k w b i ,n i ,k (cid:48) | , ∀ i ∈ I , b i ∈ B i , n i ∈ N i ,k, k (cid:48) ∈ K , k (cid:48) ( b i ) > k ( b i ) , (8e) ρ b i ,n i ,k γ b i ,n i ,k ( k (cid:48) ) ≥ ρ b i ,n i ,k ρ b i ,n i ,k (cid:48) γ b i ,n i ,k (cid:48) ( k (cid:48) ) , ∀ i ∈ I , b i ∈ B i , k, k (cid:48) ∈ K , k ( b i ) > k (cid:48) ( b i ) ,n i ∈ N i , (8f) x b i ,n i ,k ≤ ρ b i ,n i ,k , ∀ k ∈ K , n i ∈ N i , b i ∈ B i , i ∈ I , (8g) ρ b i ,n i ,k , x b i ,n i ,k ∈ (cid:110) , (cid:111) , ∀ i ∈ I , b i ∈ B i , n i ∈ N i , ∀ k ∈ K , (8h) (cid:88) b i ∈B i x b i ,n i ,k ≤ , ∀ i ∈ I , n i ∈ N i , ∀ k ∈ K , (8i)(1) , (2) , where W = [ w b i ,n i ,k ] , ∀ i, b i , n i , k , ρ = [ ρ b i ,n i ,k ] , ∀ i, b i , n i , k , X = [ x b i ,n i ,k ] , ∀ i, b i , n i , k , (8b) showsthe total available transmit power at each BS, (8c) indicates the total available transmit power foreach MVNO, (8d) represents the minimum rate requirement for each MVNO, (8e) demonstratesthe PD-NOMA constraint, (8f) shows the SIC ordering constraint, and (8i) represents the NOSselection limitation in each InP, at each user and for each subcarrier.IV. S OLUTION M ETHODS B ASED ON N ETWORK D ENSITY
The proposed optimization problem is a nonlinear program incorporating both integer and con-tinuous variables. Moreover, due to the non-concavity of the objective function and constraints, itis not convex, and hence, the available convex optimization methods cannot be used directly. Tosolve the proposed optimization problem, two methods namely, I) centralized resource allocationand II) semi-centralized resource allocation algorithms are exploited. In the centralized approach,all the decisions to find appropriate resource allocation are organized at the SDN controller. Inthe semi-centralized method, each of BSs takes part in finding resource allocation method.
1) Centralized Resource Allocation Algorithm:
Here, we propose a centralized RRM al-gorithm to solve (8). In centralized RRM, channel state information (CSI) of all users aretransformed to BBU, and after solving the optimization problem, final results are transmitted toBSs. In the proposed algorithm, we first relax the combinatorial constraints in (8h) by relaxingvariables ρ b i ,n i ,k and x b i ,n i ,k to have a real value between and . With this relaxation ρ b i ,n i ,k indicates a time sharing factor which is interpreted as the portion of time that subcarrier n i isassigned to a user k for a specific transmission frame, and x b i ,n i ,k shows a time sharing whichis interpreted as the portion of time that user k at subcarrier n i is on the viewpoint of BS b i [46], [47]. We solve the relaxed form of (8) using the monotonic programming approach withpoly block algorithm [48]. Note that all of the processes of driving appropriate beamforming,subcarrier allocation and NOS set selection are done at a central controller in the BBU pool. Theorem 1. (8) can be transformed into the canonical form of a monotonic optimization problem.Proof.
Constraint (8d) can be replaced by a single constraint as follows: min v ∈ V [ Q + v − Q − v ] ≥ R v min , ∀ v, k ∈ K v , in which Q + v = (cid:88) i ∈I (cid:88) b i ∈B i (cid:88) n i ∈N i x b i ,n i ,k log I NOMA b i ,n i ,k + I Inter b i ,n i ,k + N + (cid:88) b (cid:48) i ∈B i ρ b (cid:48) i ,n i ,k | h Hb (cid:48) i ,n i ,k w b (cid:48) i ,n i ,k | , (9) Q − v = (cid:88) i ∈I (cid:88) b i ∈B i (cid:88) n i ∈N i x b i ,n i ,k log (cid:0) I NOMA b i ,n i ,k + I Inter b i ,n i ,k + N (cid:1) . (10)Hence, we have min v ∈ V (cid:34) Q + v + (cid:80) v (cid:48) ∈V / { v } Q − v (cid:48) (cid:35) − (cid:80) v (cid:48) ∈V Q − v (cid:48) ≥ R min , ∀ v, k ∈ K v . By defining ˆ Q + =min v ∈ V (cid:34) Q + v + (cid:80) v (cid:48) ∈V / { v } Q − v (cid:48) (cid:35) , and ˆ Q − = (cid:80) v (cid:48) ∈V Q − v (cid:48) − R min , (8d) is transformed into ˆ Q + + ˆ Q − ≥ . Subsequently, by introducing the auxiliary variables S , (8d) is transformed into the followinginequalities: ˆ Q + + S > ˆ Q − (cid:0) W mask , ρ mask , X mask (cid:1) , (11) ˆ Q − + S < ˆ Q − (cid:0) W mask , ρ mask , X mask (cid:1) , (12) ≤ S ≤ ˆ Q − (cid:0) W mask , ρ mask , X mask (cid:1) − ˆ Q − (0 , , , (13)where W mask , ρ mask , and X mask are the maximum thresholds of W , ρ , and X , respectively.For constraint (8e), we let T − i,b i ,n i ,k,k (cid:48) = − ρ b i ,n i ,k | h Hb i ,n i ,k w b i ,n i ,k | , T + i,b i ,n i ,k,k (cid:48) = ρ b i ,n i ,k ρ b i ,n i ,k (cid:48) | h Hb i ,n i ,k w b i ,n i ,k (cid:48) | . Then by introducing the auxiliary variables S = [ S i,b i ,n i ,k,k (cid:48) ] , (8e) is trans-formed into the following inequalities: T + i,b i ,n i ,k,k (cid:48) + S i,b i ,n i ,k,k (cid:48) ≤ T + i,b i ,n i ,k,k (cid:48) (cid:0) W mask , ρ mask (cid:1) , (14) T − i,b i ,n i ,k,k (cid:48) + S i,b i ,n i ,k,k (cid:48) ≥ T + i,b i ,n i ,k,k (cid:48) (cid:0) W mask , ρ mask (cid:1) , (15) ≤ S i,b i ,n i ,k,k (cid:48) ≤ T + i,b i ,n i ,k,k (cid:48) (cid:0) W mask , ρ mask (cid:1) − T + i,b i ,n i ,k,k (cid:48) (0 , . (16)For constraint (8f), after simplifying it, we let ˜ T + i,b i ,n i ,k,k (cid:48) = ρ b i ,n i ,k ρ b i ,n i ,k (cid:48) (cid:88) b (cid:48) i ∈B i ρ b (cid:48) i ,n i ,k (cid:48) | h Hb (cid:48) i ,n i ,k (cid:48) w b (cid:48) i ,n i ,k (cid:48) | (cid:0) I NOMA b i ,n i ,k + I Inter b i ,n i ,k + N (cid:1) , (17) ˜ T − i,b i ,n i ,k,k (cid:48) = ρ b i ,n i ,k (cid:88) b (cid:48) i ∈B i ρ b (cid:48) i ,n i ,k | h Hb (cid:48) i ,n i ,k w b (cid:48) i ,n i ,k (cid:48) | (cid:0) I NOMA b i ,n i ,k (cid:48) + I Inter b i ,n i ,k (cid:48) + N (cid:1) . (18)Hence, (8f) is transformed into ˜ T + i,b i ,n i ,k,k (cid:48) − ˜ T − i,b i ,n i ,k,k (cid:48) ≤ , ∀ i ∈ I , b i ∈ B i , k, k (cid:48) ∈ K , k ( b i ) >k (cid:48) ( b i ) . With the same way applied to constraint (8e), by introducing the auxiliary variables S ,constraint (8f) is transformed into the following inequalities: ˜ T + i,b i ,n i ,k,k (cid:48) + S i,b i ,n i ,k,k (cid:48) ≤ ˜ T + i,b i ,n i ,k,k (cid:48) (cid:0) W mask , ρ mask , X mask (cid:1) , (19) ˜ T − i,b i ,n i ,k,k (cid:48) + S i,b i ,n i ,k,k (cid:48) ≥ ˜ T + i,b i ,n i ,k,k (cid:48) (cid:0) W mask , ρ mask , X mask (cid:1) , (20) ≤ S i,b i ,n i ,k,k (cid:48) ≤ ˜ T + i,b i ,n i ,k,k (cid:48) (cid:0) W mask , ρ mask , X mask (cid:1) − ˜ T + i,b i ,n i ,k,k (cid:48) (0 , , . (21)In order to transform the objective function to monotonic form at first, we define ˜ Q + and ˜ Q − as ˜ Q + = (cid:88) i ∈I (cid:88) b i ∈B i (cid:88) n i ∈N i (cid:88) k ∈K x b i ,n i ,k log (cid:18) I NOMA b i ,n i ,k + I Inter b i ,n i ,k + N + (cid:88) b (cid:48) i ∈B i ρ b (cid:48) i ,n i ,k | h Hb (cid:48) i ,n i ,k w b (cid:48) i ,n i ,k | (cid:19) , (22) ˜ Q − = (cid:88) i ∈I (cid:88) b i ∈B i (cid:88) n i ∈N i (cid:88) k ∈K x b i ,n i ,k log (cid:0) I NOMA b i ,n i ,k + I Inter b i ,n i ,k + N (cid:1) , (23)then, by introducing the auxiliary variables S , the objective function (8a) is transformed intothe following form: max W , ρ , X , S , S , S , S ˜ Q + + S (24) ˜ Q − + S ≤ ˜ Q − (cid:0) W mask , ρ mask , X mask (cid:1) , (25) ≤ S ≤ ˜ Q − (cid:0) W mask , ρ mask , X mask (cid:1) − ˜ Q − (0 , , . (26)Finally, the equivalent form of (8) in the canonical form of monotonic optimization is expressedas follows [49], [50]: max W , ρ , X ,S ,S , S , S ˜ Q + + S (27a)s.t. : (1), (2), (8b), (8c), (8g), (8i), (11)-(16), (19)-(21) , (25) , (26) . The feasible set of problem (27), can be described as the intersection of the following twosets:
N S = (cid:8) ( S , S , S , S , W , ρ , X ) : W ≤ W mask , ρ ≤ ρ mask , X ≤ X mask , (12) , (14) , (19) , (25) (cid:9) , (28) CN S = { ( S , S , S , S , W , ρ , X ) : W ≥ , ρ > , X > , (11) , (15) , (20) } , (29)where N S is normal and
CN S is co-normal.To solve the monotonic optimization problem the poly block algorithm is utilized [48].
2) Semi-Centralized Resource Allocation Algorithm:
In this method, to solve the proposedoptimization problem, an iterative algorithm is exploited in which each iteration decouples themain optimization problem into two sub-problems, namely, I) beamforming, and II) joint NOSselection and subcarrier allocation. To solve the beamforming problem in each iteration, theSCA approach with the dual method is used. Meanwhile, to solve the joint NOS selectionand subcarrier allocation problem, at first, the time sharing method is exploited, then the SCAapproach with the dual method is utilized. In the following, first, the beamforming, NOSselection, and subcarrier allocation are described, then the semi-centralized method is presented. • Beamforming : To obtain beamforming vectors, we solve the following problem: max W (cid:88) i ∈I (cid:88) b i ∈B i (cid:88) n i ∈N i (cid:88) k ∈K x b i ,n i ,k r b i ,n i ,k , s.t. (8 b ) − (8 f ) . (30)By introducing a new variable as t = [ t b i ,n i ,k ] , ∀ i, b i , n i , k , the optimization problem (30)can be transformed into the equivalent form as: max W , t (cid:88) i ∈I (cid:88) b i ∈B i (cid:88) n i ∈N i (cid:88) k ∈K x b i ,n i ,k log t b i ,n i ,k (31a)s.t. : (8b)-(8f) ,γ b i ,n i ,k ≥ t b i ,n i ,k − , ∀ i ∈ I , b i ∈ B i , k ∈ K , n i ∈ N i . (31b)We can show that (31) can be transformed into the following form max W , t ,(cid:36) (cid:88) i ∈I (cid:88) b i ∈B i (cid:88) n i ∈N i (cid:88) k ∈K x b i ,n i ,k log t b i ,n i ,k (32a)s.t. : (8b)-(8f) , (cid:88) b (cid:48) i ∈B i ρ b (cid:48) i ,n i ,k | h Hb (cid:48) i ,n i ,k w b (cid:48) i ,n i ,k | ≥ (cid:36) b i ,n i ,k t b i ,n i ,k − (cid:36) b i ,n i ,k , ∀ i ∈ I , b i ∈ B i , k ∈ K ,n i ∈ N i , (32b) I NOMA b i ,n i ,k + I Inter b i ,n i ,k + N ≤ (cid:36) b i ,n i ,k , ∀ i ∈ I , b i ∈ B i , k ∈ K , n i ∈ N i . (32c)Du to the constraint (32b) and and constraint (8f) the presented optimization problem in(32) is non-convex. By using the first order Taylor approximation we approximate it by aconvex one. In iteration q the non-convex term in the left side of the constraint (32b),is written by | h Hb i ,n i ,k w b i ,n i ,k | = (cid:107) θ b i ,n i ,k (cid:107) = ( θ Rlb i ,n i ,k ) + ( θ Imb i ,n i ,k ) , where θ Rlb i ,n i ,k = Real ( h Hb i ,n i ,k w b i ,n i ,k ) , θ Imb i ,n i ,k = Image ( h Hb i ,n i ,k w b i ,n i ,k ) . Then, we use a linear approxima-tion of (cid:107) θ b i ,n i ,k (cid:107) as follows: (cid:107) θ b i ,n i ,k (cid:107) (cid:39) (cid:107) θ q − b i ,n i ,k (cid:107) + 2( θ q − b i ,n i ,k ) T ( θ b i ,n i ,k − θ q − b i ,n i ,k ) . Now, todeal with the bilinear product on the right side of (32b), we rewrite (cid:36) b i ,n i ,k t b i ,n i ,k as follows (cid:36) b i ,n i ,k t b i ,n i ,k = 14 (cid:20) ( (cid:36) b i ,n i ,k + t b i ,n i ,k ) − ( (cid:36) b i ,n i ,k − t b i ,n i ,k ) (cid:21) . (33)The first term of (33) is convex, consequentially, the first order Taylor approximation isapplied to the second term as follows (cid:36) b i ,n i ,k t b i ,n i ,k (cid:39)
14 ( (cid:36) b i ,n i ,k + t b i ,n i ,k ) − (cid:20) ( (cid:36) q − b i ,n i ,k − t q − b i ,n i ,k ) + 2( (cid:36) q − b i ,n i ,k (34) − t cb i ,n i ,k )( (cid:36) b i ,n i ,k − (cid:36) q − b i ,n i ,k − t b i ,n i ,k + t q − b i ,n i ,k ) , Finally, constraint (32b) is written by (cid:88) b (cid:48) i ∈B i ρ b (cid:48) i ,n i ,k (cid:18) (cid:107) θ q − b i ,n i ,k (cid:107) + 2( θ q − b i ,n i ,k ) T ( θ b i ,n i ,k − θ q − b i ,n i ,k ) (cid:19) ≥
14 ( (cid:36) b i ,n i ,k + t b i ,n i ,k ) − (cid:36) b i ,n i ,k − (cid:20) ( (cid:36) q − b i ,n i ,k − t q − b i ,n i ,k ) + 2( (cid:36) cb i ,n i ,k − t q − b i ,n i ,k )( (cid:36) b i ,n i ,k − (cid:36) q − b i ,n i ,k − t b i ,n i ,k + t q − b i ,n i ,k ) (cid:21) , ∀ i ∈ I , b i ∈ B i , k ∈ K , n i ∈ N i , (35)which presents a convex constraint. In addition, (8d) is transformed into the following convexform as (cid:80) n i ∈N i ( x b i ,n i ,k log t b i ,n i ,k ) ≥ R v min , ∀ v, k ∈ K v . For (8f) we have | h Hb i ,n i ,k (cid:48) w b i ,n i ,k | = (cid:107) θ k (cid:48) b i ,n i ,k (cid:107) = ( θ Rl,k (cid:48) b i ,n i ,k ) + ( θ Im,k (cid:48) b i ,n i ,k ) , where θ Rl,k (cid:48) b i ,n i ,k = Real ( h Hb i ,n i ,k (cid:48) w b i ,n i ,k ) , and θ Im,k (cid:48) b i ,n i ,k = Image ( h Hb i ,n i ,k (cid:48) w b i ,n i ,k ) . Then, we use a linear approximation of (cid:107) θ b i ,n i ,k (cid:107) as follows: (cid:107) θ k (cid:48) b i ,n i ,k (cid:107) (cid:39) (cid:107) θ q − ,k (cid:48) b i ,n i ,k (cid:107) + 2( θ q − ,k (cid:48) b i ,n i ,k ) T ( θ k (cid:48) b i ,n i ,k − θ q − ,k (cid:48) b i ,n i ,k ) . (36) (cid:107) θ k b i ,n i ,k (cid:107) (cid:107) θ k (cid:48) b i ,n i ,k (cid:107) = 14 (cid:20) ( (cid:107) θ k b i ,n i ,k (cid:107) + (cid:107) θ k (cid:48) b i ,n i ,k (cid:107) ) − ( (cid:107) θ k b i ,n i ,k (cid:107) − (cid:107) θ k (cid:48) b i ,n i ,k (cid:107) ) (cid:21) . (37)By applying (36), (37) is written by (cid:107) θ k b i ,n i ,k (cid:107) (cid:107) θ k (cid:48) b i ,n i ,k (cid:107) = 14 (cid:20) (( (cid:107) θ q − ,k b i ,n i ,k (cid:107) + 2( θ q − ,k b i ,n i ,k ) T ( θ k b i ,n i ,k − θ q − ,k b i ,n i ,k )) (38) + ( (cid:107) θ q − ,k (cid:48) b i ,n i ,k (cid:107) + 2( θ q − ,k (cid:48) b i ,n i ,k ) T ( θ k (cid:48) b i ,n i ,k − θ q − ,k (cid:48) b i ,n i ,k ))) − (( (cid:107) θ q − ,k b i ,n i ,k (cid:107) + 2( θ q − ,k b i ,n i ,k ) T ( θ k b i ,n i ,k − θ q − ,k b i ,n i ,k )) − ( (cid:107) θ q − ,k (cid:48) b i ,n i ,k (cid:107) + 2( θ q − ,k (cid:48) b i ,n i ,k ) T ( θ k (cid:48) b i ,n i ,k − θ q − ,k (cid:48) b i ,n i ,k ))) (cid:21) . With the same steps applied in (34), (38) is approximated by a convex function.Consequentially, by applying the described liner approximation a convex form of beamform-ing problem is achieved. Then, we apply the dual method to solve the convex approximatedproblem. The Lagrangian function can thus be written by L ( W , t , (cid:36) , (cid:37) , α , β , λ , δ ) = (cid:88) i ∈I (cid:88) b i ∈B i (cid:88) n i ∈N i (cid:88) k ∈K x b i ,n i,k log ( t b i ,n i,k ) (39) − (cid:88) b i ∈B i α b i ( (cid:88) n i ∈N i (cid:88) k ∈K ρ b i ,n i ,k (cid:107) w b i ,n i ,k (cid:107) − P b i max ) − (cid:88) v ∈V β v ( (cid:88) i ∈I (cid:88) b i ∈B i (cid:88) n i ∈N i (cid:88) k ∈K v ρ b i ,n i ,k (cid:107) w b i ,n i ,k (cid:107) − P v max )+ (cid:88) i ∈I (cid:88) b i ∈B i (cid:88) n i ∈N i (cid:88) k ∈K v δ i,b i ,n i ,k (cid:18) (cid:88) b (cid:48) i ∈B i ρ b (cid:48) i ,n i ,k ( (cid:107) θ b i ,n i ,k (cid:107) + 2( θ q − b i ,n i ,k ) T ( θ b i ,n i ,k − θ q − b i ,n i ,k ) − ( 14 ( (cid:36) b i ,n i ,k + t b i ,n i ,k ) − (cid:36) b i ,n i ,k −
14 [( (cid:36) q − b i ,n i ,k − t q − b i ,n i ,k ) + 2( (cid:36) q − b i ,n i ,k − t q − b i ,n i ,k )( (cid:36) b i ,n i ,k − (cid:36) q − b i ,n i ,k − t b i ,n i ,k + t q − b i ,n i ,k )]) + (cid:36) b i ,n i ,k − R v min (cid:19) − (cid:88) i ∈I (cid:88) b i ∈B i (cid:88) n i ∈N i (cid:88) k ∈K λ b i ,n i ,k ( I NOMA b i ,n i ,k + I Inter b i ,n i ,k + N − (cid:36) b i ,n i ,k )+ (cid:88) i ∈I (cid:88) b i ∈B i (cid:88) n i ∈N i (cid:88) k ∈K (cid:88) k (cid:48)(cid:48) ∈K (cid:37) b i ,n i ,k,k (cid:48)(cid:48) (cid:20) ρ b i ,n i ,k ρ b i ,n i ,k (cid:48) (cid:88) b (cid:48) i ∈B i ρ b (cid:48) i ,n i ,k (cid:48) | h Hb (cid:48) i ,n i ,k (cid:48) w b (cid:48) i ,n i ,k (cid:48) | (cid:0) I NOMA b i ,n i ,k + I Inter b i ,n i ,k + N (cid:1) − ρ b i ,n i ,k (cid:88) b (cid:48) i ∈B i ρ b (cid:48) i ,n i ,k | h Hb (cid:48) i ,n i ,k w b (cid:48) i ,n i ,k (cid:48) | (cid:0) I NOMA b i ,n i ,k (cid:48) + I Inter b i ,n i ,k (cid:48) + N (cid:1) (cid:21) , where α = [ α b i ] , β = [ β v ] , λ = [ λ b i ,n i ,k ] , (cid:37) = [ (cid:37) b i ,n i ,k,k (cid:48) ] , and δ = [ δ i,b i ,n i ,k ] are the nonnega-tive Lagrange multipliers. The dual function is given by max W , t , (cid:36) L ( W , (cid:37) , t , (cid:36) , α , β , λ , δ ) . Finally, calculating the stationary point of Lagrangian function, we have w b i ,n i ,k ( i ) = (cid:80) b (cid:48) i ∈B i δ i,b (cid:48) i ,n i ,k + Ω (cid:37) b i ,n i ,k,k (cid:48) β v h b i ,n i ,k ( i ) + Ψ( i ) , (40)where Ω is a constant value coming from the approximated form of constraint (8f) Ψ( i ) = 2 α b i h b i ,n i ,k ( i ) + (cid:88) b (cid:48) i ∈B i /b i (cid:88) k (cid:48) ∈K /k (1 − ρ b (cid:48) i ,n i ,k (cid:48) )2 λ b (cid:48) i ,n i ,k (cid:48) h b (cid:48) i ,n i ,k (cid:48) ( i )+ (41) Total number of users O v e r a ll s i gn a li ng ov e r h ea d s B=4, SD-CRMB=8, SD-CRMB=12, SD-CRMB=4, SD-SCRMB=8, SD-SCRMB=12, SD-SCRM
Fig. 6: Overall signaling overhead in terms of Kilo Bytes (KB) for the proposed SD-CRM and SD-SCRM algorithmsversus the number of users for different number of BSs. (cid:88) b (cid:48) i ∈B i /b i (cid:88) k (cid:48) ∈K ,k (cid:48) ( b (cid:48) i ) To solve the joint NOS selection andsubcarrier allocation problem, first we relax ρ b i ,n i ,k and x b i ,n i ,k to have ≤ ρ b i ,n i ,k ≤ and ≤ x b i ,n i ,k ≤ . Then, similar steps as for the beamforming are applied.In the proposed semi-centralized resource allocation method, the solution steps are presentedin four groups, namely, I) initialization, II) beamforming, III) joint NOS selection and subcar-rier allocation IV) convergence conditions check, which are described in Alg. 1. By defining R s tot ( W , ρ , X ) = (cid:80) i ∈I (cid:80) b i ∈B i (cid:80) n i ∈N i (cid:80) k ∈K x b i ,n i ,k r b i ,n i ,k as the sum-rate in iteration s of Alg.1, the convergence condition at the SDN controller is | R s tot ( W , ρ , X ) − R s − tot ( W , ρ , X ) | ≤ (cid:15) ,where (cid:15) is a positive small value.V. S IGNALING O VERHEAD AND O PERATIONAL COMPLEXITY A. Signaling Overhead In this subsection, we study the effect of the number of users and BSs on the overall signallingoverhead for the SD-CRM and SD-SCRM algorithms. As discussed in Section II, the SDNcontroller selects the SD-CRM algorithm for ultra-dense scenarios while the SD-SCRM algorithmis selected for both the moderate and high density scenarios. Fig. 6 shows the effect of the numberof users and BSs on the overall signalling overhead for the SD-CRM and SD-SCRM algorithms.As shown, the signalling overhead of the proposed SD-CRM algorithm is higher than that of theproposed SD-SCRM approach. Moreover, it can be seen that when the number of users and/or Algorithm 1 Semi-centralized resource allocation algorithm Initialize ρ (0) , X (0) , Υ (small value) and β at BBU, α i , λ b i ,n i ,k , ∀ n i ∈ N i , k ∈ K , (cid:37) b i ,n i ,k,k (cid:48) , ∀ n i ∈ N i , k, k (cid:48) ∈ K and δ b i ,n i ,k , ∀ n i ∈ N i , k ∈ K at BS b i , q in = 0 (inner iterationnumber) and q out = 0 (outer iteration number) at the SDN controller. while the convergence conditions are satisfied by the SDN controller do Beamforming :With ρ = ρ ( q out ) and X = X ( q out ) the following steps should be done at BS b i : Broadcast (cid:80) k (cid:48) ∈K /k (1 − ρ b i ,n i ,k (cid:48) )2 λ b i ,n i ,k (cid:48) h b i ,n i ,k (cid:48) ( i ) + (cid:80) k (cid:48) ∈K ,k (cid:48) ( b i ) Feedback variable Number of bits β v , δ b i ,n i ,k h b i ,n i ,k M T × (cid:80) k (cid:48) ∈K /k (1 − ρ b i ,n i ,k (cid:48) )2 λ b i ,n i ,k (cid:48) h b i ,n i ,k (cid:48) ( i ) + (cid:80) k (cid:48) ∈K ,k (cid:48) ( b i ) B. Operational Complexity: A Metric for RRM Type Selection Here, the processing ability of each BS is the maximum amount of operations that eachBS should perform during the solution of an optimization problem. In other words, BSs in aspecific time should perform a specific amount of processing. Optimizing a system via solvingan optimization problem can improve the system performance if the considered optimizationproblem can be solved during a specific time. With increasing the network density, the amountof processing required to solve an optimization problem in each BS increases. Therefore, if itexceeds the maximum processing ability of BSs, the optimization problem can not be solveddynamically. By increasing the network density, if the required processing is more than theprocessing ability of BSs, the SD-CRM approach should be exploited. The BBU comprisespowerful processors which can solve optimization problems in the required time. Consequently,in networks with high density if we choose SD-DRM, the system can not work efficiently asit is not possible to achieve the solution of the resource allocation problem before transmissiontime is started. Moreover, when the network density is low, BSs can satisfy the processingrequirement. In this case, if we choose the SD-CRM approach as shown in Subsection V-A,more signaling overhead is imposed on the network. As an example, Fig. 7 shows the numberof active users in different hours for two days in a week. Accordingly, based on the networkdensities in different hours of a day, RAN architecture and its corresponding RRM type shouldbe dynamically changed. In this evaluation, operational complexity refers to the amount ofoperations which should be performed at a BS to solve the proposed optimization problems.As shown in Fig. 7-Left, for some hours, the density of users and its corresponding trafficrate increase. This needs more computational resources which causes more pressure at BSs in Hour of day T o t a l nu m b e r o f ac ti v e u s e r s ThursdaySunday Hour of day O p e r a ti on a l c o m p l e x it y p e r B S ( b it s ) × ThursdaySunday SD-SCRM SD-CRM Fig. 7: Left: Total number of active users at different hours in various days for a week. Right: Per-BS operationalcomplexity at different hours of a day in the proposed SD-SCRM algorithm for iterations of Alg. 1. the SD-SCRM algorithm. Fig. 7-Right, shows the operational complexity of each BS, which isdefined as operational complexity per BS, versus of the hours of a day and indicates the selectedRRM type in each time which is function of user density. When the density of users is high, theoperational complexity of the proposed SD-SCRM algorithm at each BS is high which couldbe out of the BS computational ability. Hence, due to the existing of high power processorat the cloud, it is better to forward the RRM processing tasks of BSs to the cloud and adoptthe SD-CRM algorithm to guarantee the efficiency of our Smart Soft-RAN from the resourcemanagement perspective (see Fig. (10a)). To this end, we define a complexity threshold whereif the operational complexity at a time exceeds it, the SD-CRM algorithm is selected and whenthe complexity of SD-SCRM is lower than this threshold, SD-SCRM is chosen. Figures (8a) and(8b) depict the achievable and outage region of the traditional and the proposed RRM algorithms,respectively. In the figures, there are two different thresholds for SD-CRM and SD-SCRM. Thethreshold of SD-SCRM is defined as the maximum tolerable operational complexity of each BSand the threshold for SD-CRM is defined as the maximum tolerable operational complexity ofcentral processor. From these figures, we can see that the SD-SCRM approach is feasible for asystem with low operational complexity. This follows from the outage event, if the operationalcomplexity of a BS is higher than its determined maximum value. Moreover, due to the highsignaling overhead of the SD-CRM approach, shown in Fig. 6, it is not appropriate for a systemwith low density to exploit the SD-CRM approach. In Fig. (8a), the area shown by the greencolor is the achievable region of the SD-SCRM approach, the area shown by the blue color is theachievable region of the SD-CRM approach, the area shown by the vertical lines is the outageregion of the SD-SCRM scheme, and the area shown by the red color illustrates the outageregion of the SD-SCRM approach. In Fig. (8b), the area shown by the vertical lines depicts the Number of Active Users O p e r a ti on a l C o m p l e x it y ( b it s ) Overall Operational Complexity for SD-CRMPer-BS Operational Complexity for SD-SCRMSD-CRM Operational Complexity ThresholdSD-SCRM Operational Complexity ThresholdOutage Region for the Traditional Scheme: SD-SCRMNon-Outage Region for Traditional Scheme: SD-SCRMOutage Region for the Traditional Scheme: SD-CRMNon-Outage Region for Traditional Scheme: SD-CRM Outage Region for the TraditionalScheme: SD-CRMNon-Outage Region forTraditional Scheme: SD-SCRM Non-Outage Region forTraditional Scheme: SD-CRM Outage Region for theTraditional Scheme: SD-SCRM Number of Active Users O p e r a ti on a l C o m p l e x it y ( b it s ) Overall Operational Complexity for SD-CRMPer-BS Operational Complexity for SD-SCRMSD-CRM Operational Complexity ThresholdSD-SCRM Operational Complexity ThresholdNon-Outage Region for the Proposed SchemeOutage Region for the Proposed SchemesNon-Outage Region for theProposed Scheme Outage Region for theProposed Schemes Fig. 8: Left: Non-outage region off the traditional approaches. Right: Non-outage region of the proposed approach. achievable region of the proposed algorithm, and the area shown by the red color illustrates theoutage region of it. From the figures, we can conclude that by exploiting the proposed approach,total number of users which can be supported by the system is increased by approximately in contrast to the conventional approach SD-SCRM. Moreover, we can see that the signalingoverhead of the system for number of users belong to the interval [1 , , as can be seen inFig. 6, is decreased by approximately in contrast to the conventional approach SD-CRM.VI. N UMERICAL R ESULTS This section provides numerical results to evaluate the performance of the proposed solutionsfor the downlink of CoMP-NOMA system. We assume two different InPs, each of which consistsof a single MBS and FBSs randomly located in the main area with radius m. Moreover,each InP has a frequency bandwidth of MHz. The frequency bandwidth of each subcarrier isalso assumed to be . KHz. Hence, the total number of subcarriers for each InP is . It is alsoassumed that there are users spread in the coverage area of the network with antennas each.The channel power gains are exponentially distributed with mean . . The pathloss exponentof the considered pathloss model is set to 2. The power spectral density (PSD) of the receivedAWGN noise is also set to − dBm/Hz. Besides, the considered network is sliced into twovirtual networks, where there are users in each virtual network. In addition, we assume aminimum required data rate for users assigned to each virtual network. Accordingly, we set R min = 2 bps/Hz and R min = 3 bps/Hz, for the first and second virtual networks, respectively.In addition, we assume a maximum transmit power for each MBS, FBS and slice as watts, . watts and watts, respectively. We suppose that maximum each subcarrier can be assignedto at maximum users in each cell, i.e., L T = 2 . We compare the performance of the proposed Maximum transmit power of each MBS (Watts) T o t a l s p ec t r a l e ff i c i e n c y ( bp s / H z ) SD-SCRM,P b i max =0.7 (Watts)SD-CRM,P b i max =0.7 (Watts)SD-SCRM,P b i max =0.5 (Watts)SD-CRM,P b i max =0.5 (Watts) P vmax ( Watts ) T o t a l s p ec t r a l e ff i c i e n c y ( bp s / H z ) SD-SCRM,R =2 (bps/Hz)SD-CRM,R =2 (bps/Hz)SD-SCRM,R =3 (bps/Hz)SD-CRM,R =3 (bps/Hz) Fig. 9: Left: Total spectral efficiency of users versus maximum allowable transmit power of each MBS for differentmaximum transmit power of each FBS. Right: Total spectral efficiency of users versus the maximum powerconsumption at each MVNO for different values of R min . SD-CRM and SD-SCRM algorithms in terms of sum data rate of users for different values of P b i max in Fig. 9-Left. As shown, the sum data rate increases, when the maximum transmit powerof each MBS and/or FBSs are increased. This is because, increasing P b i max , extends the feasibleregion of (8b) which means that more transmit powers can be allocated to users. Hence, thesum data rate of users increases. In Fig. 9-Right, we investigate the effect of the consideredvirtualization scheme in our model by evaluating the performance of the proposed algorithmsfor different values of P v max and R v min . In doing so, we increase the maximum allowable transmitpower of each MVNO from Watts to Watts for different values of R min as bps/Hz and bps/Hz and obtain the sum data rate of users in each scenario. As shown, the performanceof the proposed algorithms are improved by increasing P v max which allows each MVNO tohave more power consumption. Whereas, more R min shrinks the feasible region of (8d) for allusers assigned to virtual network which causes more reduction in sum data rate of users.Same conclusion can be derived by increasing R min for all users of virtual network . Fig.(10a) evaluates the performance of the utilized CoMP technique in our algorithms and alsoinvestigates the performance gain achieved by optimizing the NOS selection in the CoMP-basedschemes. We specifically compare our proposed sub-optimal joint radio resource allocation andNOS selection algorithms with a heuristic NOS selection algorithm, where each user is assignedto the NOMA set of a BS based on the following condition: each user selects the nearest FBSif the distance of the user and the nearest FBS be less than m, otherwise MBS is selectedfor that user. To be more specific, we compare the performance of the proposed algorithmsin terms of the sum data rate versus the total number of users for different CoMP-based and Total number of users in each MVNO T o t a l s p ec t r a l e ff i c i e n c y ( bp s / H z ) SD-CRM,CoMP,Proposed NOS selectionSD-CRM,CoMP,Heuristic NOS selectionSD-SCRM,CoMP,Proposed NOS selectionSD-SCRM,CoMP,Heuristic NOS selectionSD-CRM,No CoMPSD-SCRM,No CoMP Minimum rate of users (bps/Hz) O u t a g e p r ob a b ilit y SD-SCRM,No CoMPSD-CRM,No CoMPSD-SCRM,CoMPSD-CRM,CoMP Fig. 10: Left(a): Total spectral efficiency of users versus total number of users for different CoMP-based andnonCoMP-based schemes. Right: Outage Probability versus minimum required rate of users. nonCoMP-based schemes which are indicated by ‘CoMP’ and ‘No CoMP’, respectively . FromFig. (10a), it can be seen that utilizing CoMP technique in our model has more gains comparedto the nonCoMP-based scheme. It is seen that the sum data rate of users increases by increasingthe total number of users. Moreover, optimizing the NOS selection in the CoMP scheme by ourproposed SD-CRM and SD-SCRM algorithms increases the sum-rate of users more than theheuristic approach. This is because, in NOS selection, the NOMA set of each BS is optimized.Fig. (10b) measures the capability of our proposed SD-CRM and SD-SCRM algorithms to meetthe users QoS requirements based on the outage probability performance metric. The outageprobability is obtained by the number of situations that the minimum required data rate of userscan not be satisfied in the network over the total number of situations. As seen, the CoMP-based scheme offers better outage performance than the No CoMP scheme, since it increasesthe users data rates more than the No CoMP scheme. Besides, the SD-CRM approach has moreperformance gain than the SD-SCRM approach. From Fig. (10b), it can derived that the outageprobability grows exponentially when the minimum required rates of users increases, since thefeasible region of constraint (8d) becomes tighten. Moreover, for small values of R v min , the outageprobability tends to zero, since the only challenge is satisfying the QoS constraint (8d).VII. C ONCLUSION In this paper, a new smart soft-RAN architecture for the 5G of cellular wireless networks waspresented in which five important tasks, namely dynamic radio resource management, dynamicBS type selection, dynamic functionality splitting, dynamic technology selection based on thenetwork density and traffic of users, and dynamic framing are applied. In this paper, due to the In the nonCoMP scheme, each user is assigned to at most one BS in the network. space limitation, we only focused on the first task and proposed three resource managementapproaches, namely, centralized, semi-centralized and distributed RRM. Based on the proposed Smart Soft-RAN architecture, as a case study, considering NOMA and CoMP techniques, in adownlink of HetNet MISO based system, a novel radio resource allocation problem with a novelCoMP-NOMA model was proposed and to solve it, based on the network density, two solutionalgorithms, namely, centralized and semi centralized methods were applied. Finally, the signalingoverhead and throughput analysis of each solution method were presented.R EFERENCES [1] Z. Niu, S. Zhou, S. Zhou, X. Zhong, and J. Wang, “Energy efficiency and resource optimized hyper-cellular mobilecommunication system architecture and its technical challenges,” SCIENTIA SINICA Informationis , vol. 42, pp. 1191–1203, Jan. 2012.[2] H. Ishii, Y. Kishiyama, and H. Takahashi, “A novel architecture for LTE-B :C-plane/U-plane split and phantom cellconcept,” in Proc. 2012 IEEE Globecom Workshops , Dec 2012, pp. 624–630.[3] Y. Lin, L. Shao, Z. Zhu, Q. Wang, , and R. Sabhikhi, “Wireless network cloud: Architecture and system requirements,” IBM Journal of Research and Development , vol. 54, no. 1, pp. 1–12, Feb. 2010.[4] C. M. R. Institute, “The road towards green RAN,” C-RAN White Paper , Jun. 2014, [Online]. Available:http://labs.chinamobile.com/cran/.[5] A. Checko, H. L. Christiansen, Y. Yan, L. Scolari, G. Kardaras, M. S. Berger, and L. Dittmann, “Cloud RAN formobile networks 2014 - a technology overview,” IEEE Communications Surveys & Tutorials , vol. 17, no. 1, pp. 405–426, Firstquarter 2015.[6] T. Salman, “Cloud RAN: Basics, advances and challenges,” Wireless and Mobile Networking , Apr. 2016.[7] T. X. Tran, A. Hajisami, P. Pandey, and D. Pompili, “Collaborative mobile edge computing in 5G networks: New paradigms,scenarios, and challenges,” IEEE Communications Magazine , vol. 55, no. 4, pp. 54–61, Apr. 2017.[8] M. Peng, Y. Li, J. Jiang, J. Li, and C. Wang, “Heterogeneous cloud radio access networks: A new perspective for enhancingspectral and energy efficiencies,” IEEE Wireless Communications , vol. 21, no. 6, pp. 126–135, Dec. 2014.[9] M. Peng, K. Zhang, J. Jiang, J. Wang, and W. Wang, “Energy-efficient resource assignment and power allocation inheterogeneous cloud radio access networks,” IEEE Transactions on Vehicular Technology , vol. 64, no. 11, pp. 5275–5287,Nov. 2015.[10] M. Peng, S. Yan, K. Zhang, and C. Wang, “Fog-computing-based radio access networks: Issues and challenges,” IEEENetwork , vol. 30, no. 4, pp. 46–53, July 2016.[11] A. Gudipati, D. Perry, L. E. Li, and S. Katti, “SoftRAN: Software defined radio access network,” in Proc. the Second ACMSIGCOMM Workshop on Hot Topics in Software Defined Networking . New York, NY, USA: ACM, 2013, pp. 25–30.[12] M. Yang, Y. Li, D. Jin, L. Su, S. Ma, and L. Zeng, “OpenRAN: A software-defined RAN architecture via virtualization,”in Proc. the ACM SIGCOMM 2013 conference on SIGCOMM , vol. 43, no. 4, Aug. 2013, pp. 549–550.[13] J. Liu, T. Zhao, S. Zhou, Y. Cheng, and Z. Niu, “CONCERT: A cloud-based architecture for next-generation cellularsystems,” IEEE Wireless Communications , vol. 21, no. 6, pp. 14–22, Dec. 2014.[14] K. S. M. Y. Arslan and S. Rangarajan, “Software-defined networking in cellular radio access networks: Potential andchallenges,” IEEE Communications Magazine , vol. 53, no. 1, pp. 150–156, Jan. 2015. [15] Z. Zaidi, V. Friderikos, and M. Imran, “Future RAN architecture: SD-RAN through a general-purpose processing platform,” IEEE Vehicular Technology Magazine , vol. 10, no. 1, pp. 52–60, Mar. 2015.[16] S.-C. L. I. F. Akyildiz, P. Wang, “SoftAir: A software defined networking architecture for 5G wireless systems,” ComputerNetworks , vol. 85, pp. 1–18, July 2015.[17] S. Zhou, T. Zhao, Z. Niu, and S. Zhou, “Software-defined hyper-cellular architecture for green and elastic wireless access,” IEEE Communications Magazine , vol. 54, no. 1, pp. 12–19, Jan. 2016.[18] K. Liang, L. Zhao, X. Chu, and H. H. Chen, “An integrated architecture for software defined and virtualized radio accessnetworks with fog computing,” IEEE Network , vol. 31, no. 1, pp. 80–87, Jan. 2017.[19] X. Chen, Z. Han, Z. Chang, G. Xue, H. Zhang, and M. Bennis, “Adapting downlink power in fronthaul-constrainedhierarchical software-defined RANs,” in Proc. IEEE Wireless Communications and Networking Conference (WCNC) , Mar.2017, pp. 1–6.[20] Y. Saito, A. Benjebbour, A. Li, K. Takeda, Y. Kishiyama, and T. Nakamura, “System-level evaluation of downlink non-orthogonal multiple access (NOMA) for non-full buffer traffic model,” in Proc. 2015 IEEE Conference on Standards forCommunications and Networking (CSCN) , Oct. 2015, pp. 94–99.[21] H. Nikopour and H. Baligh, “Sparse code multiple access,” in Proc. 2013 IEEE 24th Annual International Symposium onPersonal, Indoor, and Mobile Radio Communications (PIMRC) , Sep. 2013, pp. 332–336.[22] M. C. Chuang, M. C. Chen, and Y. H. Lin, “SDN-based resource allocation scheme in ultra-dense OFDMA smallcellnetworks,” in Proc. International Conference on Advanced Materials for Science and Engineering (ICAMSE) , Nov. 2016,pp. 524–527.[23] J. Liu, X. Xu, W. Chen, and Y. Hou, “QoS guaranteed resource allocation with content caching in SDN enabled mobilenetworks,” in Proc. 2016 IEEE Communications in China (ICCC Workshops) , Jul. 2016, pp. 1–6.[24] C. Liang, F. R. Yu, H. Yao, and Z. Han, “Virtual resource allocation in information-centric wireless networks withvirtualization,” IEEE Transactions on Vehicular Technology , vol. 65, no. 12, pp. 9902–9914, Dec. 2016.[25] Y. Cai, F. R. Yu, C. Liang, B. Sun, and Q. Yan, “Software-defined device-to-device (D2D) communications in virtualwireless networks with imperfect network state information (NSI),” IEEE Transactions on Vehicular Technology , vol. 65,no. 9, pp. 7349–7360, Sep. 2016.[26] X. Zhang and Q. Zhu, “Information-centric network virtualization for QoS provisioning over software defined wirelessnetworks,” in Proc. 2016 Military Communications Conference(MILCOM) , Dec. 2016, pp. 1–6.[27] J. Choi, “Non-orthogonal multiple access in downlink coordinated two-point systems,” IEEE Communications Letters ,vol. 18, no. 2, pp. 313–316, Feb. 2014.[28] Y. Tian, A. R. Nix, and M. Beach, “On the performance of opportunistic NOMA in downlink CoMP networks,” IEEECommunications Letters , vol. 20, no. 5, pp. 998–1001, May 2016.[29] Z. Liu, G. Kang, L. Lei, N. Zhang, and S. Zhang, “Power allocation for energy efficiency maximization in downlink CoMPsystems with NOMA,” in Proc. IEEE Wireless Communications and Networking Conference (WCNC) , Mar. 2017, pp. 1–6.[30] A. Beylerian and T. Ohtsuki, “Coordinated non-orthogonal multiple access (CO-NOMA),” in Proc. IEEE GlobalCommunications Conference (GLOBECOM) , Dec. 2016, pp. 1–5.[31] M. S. Ali, E. Hossain, and D. I. Kim, “Coordinated multi-point (CoMP) transmission in downlink multi-cell NOMAsystems: Models and spectral efficiency performance,” arXiv:1703.09255v1 [cs.NI] , 27 Mar. 2017.[32] M. F. Hanif, Z. Ding, T. Ratnarajah, and G. K. Karagiannidis, “A minorization-maximization method for optimizing sumrate in the downlink of non-orthogonal multiple access systems,” IEEE Transactions on Signal Processing , vol. 64, no. 1,pp. 76–88, Jan 2016. [33] Q. Sun, S. Han, Z. Xu, S. Wang, I. Chih-Lin, and Z. Pan, “Sum rate optimization for MIMO non-orthogonal multipleaccess systems,” in Proc. IEEE Wireless Communications and Networking Conference (WCNC) , March 2015, pp. 747–752.[34] Q. Zhang, Q. Li, and J. Qin, “Robust beamforming for nonorthogonal multiple-access systems in MISO channels,” IEEETransactions on Vehicular Technology , vol. 65, no. 12, pp. 10 231–10 236, Dec 2016.[35] B. Kimy, S. Lim, H. Kim, S. Suh, J. Kwun, S. Choi, C. Lee, S. Lee, and D. Hong, “Non-orthogonal multiple access ina downlink multiuser beamforming system,” in Proc. MILCOM 2013 - 2013 IEEE Military Communications Conference ,Nov 2013, pp. 1278–1283.[36] Q. Sun, S. Han, C. L. I, and Z. Pan, “On the ergodic capacity of MIMO NOMA systems,” IEEE Wireless CommunicationsLetters , vol. 4, no. 4, pp. 405–408, Aug 2015.[37] Z. Ding and H. V. Poor, “Design of massive-MIMO-NOMA with limited feedback,” IEEE Signal Processing Letters ,vol. 23, no. 5, pp. 629–633, May 2016.[38] M. Moltafet, N. M. Yamchi, M. R. Javan, and P. Azmi, “Comparison study between PD-NOMA and SCMA,” IEEETransactions on Vehicular Technology , vol. 67, no. 2, pp. 1830–1834, Feb 2018.[39] G. Li, S. Jin, F. Zheng, X. Gao, and X. Wang, “Dynamic base station sleeping scheme for potential super density networks,”in Proc. International Conference on Wireless Communications and Signal Processing, Hangzhou, China , Oct. 2013, pp.1–5.[40] E. Oh and B. Krishnamachari, “Energy savings through dynamic base station switching in cellular wireless access networks,”in Proc. IEEE Global Telecommunications Conference (GLOBECOM) , Dec. 2010, pp. 1–5.[41] E. Oh, K. Son, and B. Krishnamachari, “Dynamic base station switching-on/off strategies for green cellular networks,” IEEE Transactions on Wireless Communications , vol. 12, no. 5, pp. 2126–2136, May 2013.[42] S. Parkvall, E. Dahlman, A. Furuskar, and M. Frenne, “NR: The new 5G radio access technology,” IEEE CommunicationsStandards Magazine , vol. 1, no. 4, pp. 24–30, Dec. 2017.[43] F. Schaich, T. Wild, and R. Ahmed, “Subcarrier spacing - how to make use of this degree of freedom,” in Proc. IEEEVehicular Technology Conference (VTC Spring), Nanjing, China , May 2016, pp. 1–6.[44] Qualcomm, “Making 5G NR a reality,” Qualcomm Technologies IEEE Journal on Selected Areas in Communications , vol. 17, no. 10, pp. 1747–1758, Oct. 1999.[47] M. Tao, Y. C. Liang, and F. Zhang, “Resource allocation for delay differentiated traffic in multiuser OFDM systems,” IEEE Transactions on Wireless Communications , vol. 7, no. 6, pp. 2190–2201, June 2008.[48] A. Rubinov, H. Tuy, and H. Mays, “An alogrithm for monotonic global optimization problems,” A Journal of MathematicalProgramming and Operations Research , vol. 49, no. 3, pp. 205–221, 2001.[49] A. Zappone, E. Bjrnson, L. Sanguinetti, and E. Jorswieck, “A framework for globally optimal energy-efficient resourceallocation in wireless networks,” in , Mar. 2016, pp. 3616–3620.[50] ——, “Achieving global optimality for energy efficiency maximization in wireless networks,” in arXiv preprintarXiv:1602.02923 .[51] D. T. Ngo, S. Khakurel, and T. Le-Ngoc, “Joint subchannel assignment and power allocation for OFDMA femtocellnetworks,”