Cooperation incentives for multi-operator C-RAN energy efficient sharing

Facing the increasing energy demands associated with the perspective of fifth generation (5G) wireless networks, the Mobile Network Operators (MNOs) are motivated to gradually convert their traditional Radio Access Network (RAN) infrastructure to more flexible and power efficient centralized architectures, i.e., Cloud-RAN (C-RAN). Apart from their promising benefits in terms of management and network optimization, these new architectures further enable the sharing of spectrum and network elements, such as the Remote Radio Heads (RRHs) and the Baseband Units (BBUs), among multiple operators. In this paper, we introduce a novel scheme based on coalitional game theory to identify the potential room for cooperation among different MNOs that provide service to the same area. The proposed scheme sets the rules for profitable collaboration and identifies the core formation conditions (i.e., pricing) for various scenarios with different market and spectrum shares among three operators. Our results show that i) cooperation among subcoalitions of MNOs is always beneficial, yielding both higher revenues and enhanced Quality of Service (QoS) for the end users, and ii) the cooperation of all operators (grand coalition) is profitable for given user pricing in different scenarios.


I. INTRODUCTION
Global mobile data traffic will grow up by a factor of eight between 2015 and 2020 [1], due to the increasing number of smart devices (smartphones, tablets, PC, M2M devices) and to the traffic volume they typically generate according to the widespread diffusion of bandwidth-greedy applications.It is estimated that by 2020 the 98% of mobile data traffic will originate from smart devices and only mobile video traffic by itself will represent the 75% of total mobile data traffic.Also, the average downstream cellular connection speed (considering all devices) will grow from 2.0 Mbps in 2015 to nearly 6.5 Mbps by 2020.
In order to follow these trends, Heterogeneous Networks (HetNets) have been widely deployed in order to improve the spatial utilization of the frequency resource.In 2015 the 51% of total mobile data traffic was offloaded onto Wi-Fi and small-cell (SC) networks.In conclusion Next-generation cellular networks (such as the fifth generation 5G network) should support very dense high speed connections.
On the other hand, Base stations (BS) are the most expensive component of traditional distributed RAN [2], where the Operating Expenditure (OPEX) represents the 60% of the Total Cost of Ownership.Also, according to [3], the Baseband (BB) processing is the most power consuming element of SCs, due to the high processing complexity required compared to the low power transmitted.Thus, SCs densification is translated into an increase in costs and CO2 emissions and, in order for 5G networks to be sustainable, it is fundamental to optimize the energy-efficiency.
Many works addressed power consumption minimization for traditional RAN, mainly leveraging BS switching-off concepts [4].The main drawback of switching-off in traditional RAN are the possible coverage holes, because of the disjoint BSs' service areas.
Cloud-RAN is a centralized RAN architecture proposed for an efficient usage of RAN resources [5], which has been widely tested and adopted already by hardware manufacturers and MNOs.Each BS is substituted with a RRH responsible for analog RF functions, a BBU which performs digital BB processing and a fronthaul (FH) link connecting RRH and BBU.BBUs of different RRHs are grouped in a pool (BBUpool) which allows joint and dynamic coordination of the BB processing in a real-time adaptable manner to current network state.Manifold are its applications, among the others, user equipment (UE) association, resource allocation, interference management (e.g.CoMP) and power minimization.
In order to minimize the dominant BB power consumption, the substitution of dedicated hardware BBUs with general purpose processors (GPP) servers [6] has been recently proposed.BB functionalities are implemented according to a virtualized multi-standard approach, as instances of a virtual machine (VM) running on the GPP-BBU.The one-to-one association between BBUs and RRHs is replaced by a more flexible oneto-many association [7]- [9], where BB resources of one BBU can be used for more RRHs at the same time.Hence, in a low loaded region RRHs can be served by a smaller set of GPP-BBUs without performance losses.The network becomes scalable and the overall power consumption can be minimized.In conclusion C-RAN dense HetNet represents a candidate architecture for providing high rates in a cost efficient way.
Many papers have addressed the NP-hard problem of efficiently mapping BBUs and RRHs while guaranteeing some objectives (e.g.power consumption minimization).Among the possible strategies, [8] models the problem as a bin-packing problem, [9] by using graph coloring, [10] as a Knapsack In the literature, it is common opinion that SC offloading by itself will not be sufficient for enabling the forecasted traffic and that mmWave bandwidth extension will be necessary.On the other hand, many works provide strategies for improving the spectrum utilization efficiency by means of opportunistic (Cognitive Radio) or cooperative spectrum sharing.
Many papers have addressed the problem of cooperation among operators in case of traditional RAN.[12] studied the benefits of jointly deploying a new shared network, [13] investigated by means of non-cooperative game theory the feasible infrastructure sharing advantages thanks to base stations switching-off.On the other hand, works concerned with C-RAN are mainly focused on the BBU optimization in single operator case [8], [10], [11].[9] considers the two operators case but with separate spectrum licenses.
In this work, we propose a novel scheme for studying the conditions for beneficial C-RAN sharing among coexisting MNOs.The objective is evaluating how QoS and profits can be improved thanks to a better spectrum utilization efficiency.The problem is modeled as a coalitional game and investigated for the three MNOs case, when different combinations of market and spectrum share are associated with each MNO.Our results demonstrate that cooperation is always advantageous and, depending on the user pricing adopted, the grand coalition can be preferred to smaller coalitions.
The rest of the paper is structured as follows: Section II presents the system model.In Section III the sharing problem formulation and the proposed fair solution are introduced.In Section IV, the conditions for the stability of the sub and grand coalition are evaluated together with the gains provided by the proposed solution.Finally, Section V concludes the paper.

II. SYSTEM MODEL
In this section, we introduce the system model taking Fig. 1 as a reference and we provide the state-of-the-art power model for its main elements.
We consider a set M of mobile MNOs that deployed their own 4G HetNet in a given area A. For each MNO m the HetNet consists of a typical RAN macro cell (MC) layer and a C-RAN SC layer.The two layers are separate in the frequency domain and each MNO owns an exclusive spectrum license for a band of B m MHz reserved for SCs.Hence interference is considered only between equipment belonging to the same layer of the same MNO.We consider a unitary frequency reuse factor for SCs.Each operator m has deployed H m RRHs uniformly distributed, which, according to Fig. 1, are connected to U m GPP BBUs through a FH link.BBUs are grouped in a centralized physical site named BBU-hotel.We assume that the BBU-hotel is co-located with the MC eNodeB and connected to the core network (CN) through the eNodeB.We assume that MNOs share eNodeB and BBU-hotel site.
In the same area, are present N UE UE uniformly distributed with activity factor f a , which represents the probability of being active at a given time.Each MNO m has an exclusive market share μ m over the total number of UE.We assume best SNR (Signal-to-Noise Ration) association, hence a specific UE will associate to the eNodeB or RRH with the highest received power above sensitivity SN R min .At the end of the association process, on average a portion of N UE will associate to the eNodeB while a given percentage O SC will be offloaded to the SC layer.Also we assume proportional-fairness as scheduling strategy, thus each of the UE associated to a particular eNodeB/RRH gets an equal amount of resources.
For each operator m, given B m and μ m , the number of the deployed RRH is constrained by a minimum guaranteed DL data rate R min for the SC layer.As well a minimum SCoffloading factor O SC min is set, as we assume higher data rates for the SC layer.

A. Power Model
According to the power model provided by the EARTH project [3], [14], RAN power consumption can be divided into a term related to RRHs and the other to the BBU-pool: In both terms the power consumed is provided as a function of the Physical Resource Blocks (PRB) used P RB us m,n in RRH n of MNO m out of the available ones P RB m in the bandwidth B m .P RB us m,n represents the total load of a generic RRH n and can be expressed as the sum of the PRBs needed for PHY layer overhead (control and signaling) and UE transmission P RB us m,n = P RB ov m + P RB UE m,n .We assume constant and equal overhead for the RRHs and we define it as a percentage θ ov of the total number of PRBs: P RB ov m = θ ov P RB m .1) RRH power model: For the power consumption of a generic RRH n, we have adopted the EARTH model for the C-RAN architecture: where P TX is the RF output power over one PRB assuming that no power adaptation is performed.η PA is the power amplifier efficiency, N a is the number of antennas, P RF is the power consumption of the RF transceiver for one PRB and σ DC , σ m are the loss coefficients due to DC-DC power supply and mains supply.For each RRH we assume P RF linearly scalable with the number of carriers.Hence, the linear model of (1) can be rewritten as: Where ] and P h,ov m is the RRH power consumption component due to PHY layer overhead.P h,ov m can be calculated by substituting P RB us m,n = P RB ov m in (1).Since we assume equal overhead in the RRHs, we don't use the subscript n.In conclusion the total power consumed by all the RRHs in the network is 2) BBU power model: As introduced in Section I we assume that BBUs are deployed by using identical x86 GPP servers with equal processing capacity X cap expressed in Giga Operations Per Second (GOPS).Each BBU server is able to instantiate multiple RRHs functionalities in the form of VMs, which are soft resources that can be migrated among BBUs and shared among RRHs.We consider uniform workload share among the servers and we model with a constant K TX the necessary computation for one PRB, when a specific transmission configuration is used [14].
Given that in the worst-case of saturated RRHs (f a = 1) the number of deployed BBUs U m has to be sufficient for supporting the BB operations of the RRHs in the area, we define U m = (K TX H m P RB m ) /X cap .In average load case, only some of the available BBUs need to be active U act m for supporting the total network load, while the remaining U id m = U m − U act m are considered idle and can go into sleep mode for energy consumption optimization.As already mentioned, one possible way of calculating the optimum BBU-RRH mapping is by solving a Knapsack problem [10].In this context we consider the ideal minimum number of active BBUs defined, similarly to U m , as U act m = K TX H m n=1 P RB us m,n /X cap .We model the BBU power consumption as a function of the total network load [14] and we add a power component P u id which accounts for idle-state BBUs (cooling, power supply, etc. [7], [8]).The power consumption of the whole BBU-pool can be expressed as: (3) where P u st is the component due to those functions independent from the network load (e.g.FFT and IFFT [5]), Δ u p is the power consumed per PRB when a specific transmission configuration and server kind are used and P u,ov m = Δ u p P RB ov m is the consumption due to overhead processing of one RRH.
3) Total power consumed: With some arithmetics over (2) and (3) and by defining P ov m = P h,ov m + P u,ov m and Δ p = Δ h p + Δ u p , P m can be rewritten as: (4)

III. COOPERATIVE GAME
In this section, we define the cooperative game approach for the assessment of MNOs incentives for running a shared SC C-RAN.Given the set of MNOs M, we assume that each operator may decide to keep running its network independently or to cooperate by creating a coalition ω where the SC layers are pooled together.In other words, the MNOs agree on sharing RRHs, BBUs, FH and SCs' spectrum and each member of the coalition has the same rights of using infrastructure and resources.By cooperating operators provide their UE with higher rates and an extended coverage, which are translated in increased revenues.On the other hand, by sharing the costs of a larger network we expect that forming a coalition will be profitable only under given conditions and depending on the particular market and spectrum share of the cooperating MNOs.To this end, we formulate the problem as a coalitional game where MNOs are the players and we investigate which are these conditions and which is their physical meaning.

A. Architecture Adaptation for Cooperation
In order to make effective the spectrum pooling, some architectural adaptations are needed.First of all RRHs are modified for supporting multiple carriers and LTE-A compliant UE will be able to access the extended band according to Carrier Aggregation (CA) technique.In addition by assuming 3GPP MOCN (multi-operator core network) compliant RRHs, multitenant traffic can be to sustained in a transparent way.Since we assumed in Section II that MNOs are already sharing the BBUhotel, BBUs can be pooled by deploying a common switch and by implementing a central shared coordinator (Fig. 1).The latter is responsible for RRHs' state monitoring and for performing the joint BBU optimization, while respecting objectives and traffic profiles of different MNOs.This architecture is consistent with the architectures defined in [15], [16] for extending the Software Defined Networks (SDN) concepts to RAN.
As we already stated in Section I we only investigate the costs due to OPEX and thus we consider the expenses for the architecture adaptation as exceptional costs out of this context.

B. Coalitional Game
For the general cooperative game (M, V ), we represent with Ω the set of all the 2 M \∅ possible coalitions and with V ω the coalition payoff, which can be considered as the maximum utility value that the set of players in coalition ω can jointly obtain.Let v m be the portion of V ω assigned to player m when participating to that coalition, named player's payoff, then a payoff allocation v ∈ R ω is the vector representing a possible distribution of the payoffs among the players in coalition ω.The core C is the set of payoff allocations such that no group of players is willing to leave the grand coalition for one of the sub-coalitions.
Coalitional games are a specific class of cooperative games [17], which address those problems where forming coalitions is preferred by the players.A particular class of coalitional games are the canonical ones where joining the grand coalition M represents the most convenient choice.This means that the payoff that player m receives out of V M is at least as large as the payoff it would receive in any of the disjoint sets of sub-coalitions Ω\M.In this terms, the core C guarantees the stability of the grand coalition as the players don't have incentives for leaving it.
Expressing the payoff allocation in the grand coalition with v ∈ V M and the one for a subcoalition with y ∈ V ω , a possible definition of the core is [17]: The core C doesn't always exist and in those cases the grand coalition is considered unstable.
Our objective is to determine under which conditions the problem of cooperation between MNOs for sharing SC C-RAN resources can be considered as a canonical coalitional game, or in other terms when the grand coalition of MNOs is preferred to the sub-coalitions and when the opposite is true.
We model MNOs payoff in ω as their profit [12], defined as the difference between revenues ρ m and costs C m , when m ∈ ω.We assume that the revenue of each MNO only depends on its own UEs and will not be shared with other MNOs.On the other hand, operators share the total C-RAN costs C ω and c ∈ R ω is the cost sharing vector which tells us the portion of it that each MNO is willing to pay ( m∈ω c m = 1, 0 ≤ c m ≤ 1).The payoff of MNO m according to c is: Thus, the value of the generic coalition ω can be defined as the sum of its members' profit: This particular definition applies to non transferable utility (NTU) coalitional games [17], where each MNO's payoff depends on the joint actions of the other MNOs in that coalition.Indeed being revenues independent for each MNOs, they all need to agree on the cost sharing vector c.The core definition provided in ( 5) is valid for NTU games.When C exists, among all the possible cost shares c ∈ R M , we choose the market share vector μ ∈ R M as unique and fair solution.This means that each operator will pay a portion of the coalition cost proportional to the number of UEs it owns, as it is a rough but logical estimation of its load contribution into the shared C-RAN (see ( 4)).In other words, c m = μ m /μ ω , where μ ω = m∈ω μ m is the market share of coalition ω.
We proceed now to the specific definition of revenues and costs for the system model defined in Section II.

C. Revenue Model
We model the revenue ρ m as a price proportional to the minimum rate R g guaranteed to the UE when MNO m participates to coalition ω.Thus, the operator charges a flat tariff τ r per unit of data rate per month [e/M bps/month] [12].Considering an investment period of T years, the revenue of MNO m over this period can be defined as below: By focusing on OPEX, the cost model reduces to the power consumption P ω of coalition ω multiplied by a constant γ p [e/W/year], which represents the price per unit of power consumed in the investment period T [Y ears].Considering average power consumption P ω over T , the total cost function C ω of coalition ω is: where ρ KW h [e/KW h] is the realistic tariff set by energy providers and γ p [e/W/year] = ρ KWH • 10 −3 • 365 • 24 is the tariff adapted to W atts for the reference period of one year.Finally, P ω is calculated as in (4), after substituting m with ω, and taking into account that operators agree on pooling together their C-RAN elements as explained in Section III-A.Hence, the aggregated BW B ω = m∈ω B m and the total number of PRB in a coalition can be represented as P RB ω = m∈ω P RB m , while the total numbers of RRHs and BBUs in coalition ω become H ω = m∈ω H m and U ω = (H ω P RB ω K TX ) /X cap .

IV. PERFORMANCE EVALUATION
We have implemented a custom simulator in Matlab to evaluate the revenue and costs of the MNOs under different coalitions and to determine the existence of the core.In the following sections, we define the network setup and the results of our experiments.

A. Simulation Set Up
We consider a network as depicted in Fig. 1, where three operators have deployed their networks in an area of 4 Km2 .We assume that there are N UE = 20000 users in this specific area and we consider two different scenarios (Table I): • Scenario A: The operators have equal market share and bandwidth capabilities.• Scenario B: The operators have different market share and bandwidth capabilities proportional to their market share.In both cases, the MNOs have deployed their network in order to satisfy the constraints on O SC min and R min .The number H m of RRHs for each operator is calculated in the worstcase scenario where f a = 1 and the guaranteed rate is R g (on average f a < 1 and the offered data rate R of f is greater than the rate charged R g ).The number of RRHs H m and of BBUs U m are provided in Table I, while the remaining system parameters are summarized in Table II.

B. Performance Results
In Fig. 2, the average offered UE rate R of f versus the offloading factor O SC min is presented for Scenario B, in order to study the benefits of cooperation among MNOs.As we may Fig. 2: Scenario B: Offered data rate vs SC offloading factor see, in stand-alone scenarios, both the offloading factor and data rate are quite low but always above O SC min and R min .By forming coalitions of two, the MNOs may significantly improve their offloading potential and offered rate, while the grand coalition (cooperation among all three operators) provides the highest rate of all scenarios and a very high offloading factor.This can be explained taking into account that by pooling the network elements (i.e., RRHs and BBUs) and aggregating the bandwidth, we manage to fully exploit the aggregated resources and optimize the spectrum usage.Please note that by cooperating and without mmWave bandwidth extension, the offered average rate is always greater than the average 6.5 Mbps estimated for 2020 (see Section I).
In continuation, in Fig. 3, we plot the profit V ω for all possible stable coalitions (applying (7) when the core is nonempty).For Scenario A, MNOs have similar profits when operating individually, since they all have same market share and spectrum.On the other hand, in Scenario B, the MNO with highest market share (i.e., MNO 3 ) also has higher profits.
We can clearly observe that, in both scenarios, any pair of MNOs always forms a stable sub-coalition, meaning that for their members it is always preferable cooperating rather than working individually.This can be explained by the fact that, when forming a coalition, the spatial optimization of the pooled resources and the enhanced spectrum usage enable better QoS, thus increasing MNOs' revenues.On the other hand, the total cost for pooling is distributed according to operator's market share and the cost of the involved MNOs remains approximately unchanged.In conclusion, by participating in a sub-coalition, MNOs' profit increases.However, not all sub-coalitions offer the same profit to their members.This can be observed in Scenario B, where sub-coalitions involving the largest operator (i.e., MNO 3 ) achieve higher aggregate profits.
As far as the grand coalition is concerned, it can be seen that it can always provide significantly higher profits than any subcoalition, for both scenarios.However, it becomes stable (i.e., the core exists) only when a minimum tariff τ r is reached.This is mainly due to the incremental costs associated with spectrum pooling, which represents the price to pay for a better spatial spectrum usage.Indeed the size of the BBU-pool U ω and the power consumed for control and signalling depend on the total BW (see Section II-A).In the case of sub-coalitions, this term can be approximated with that of individual operation, since the bandwidth increase is relatively small.Therefore, its impact on the total cost is negligible.However, when all resources are pooled to form the grand coalition, the aggregated  BW increases significantly, and this term becomes dominant.Hence, tariffs must be set above a given minimum value, in order to ensure that the obtained revenues will payback the increased OPEX, leading to a stable grand coalition.
By comparing the two scenarios, we notice that the minimum tariff required in Scenario A (τ r = 0.23) is much smaller with respect to Scenario B (τ r = 0.62).Indeed, as we show below, for low values of τ r , the core coincides with the market share which operators are forced to adopt as the sole stable payoff distribution.Thus, for Scenario A, the three equal sized MNOs have the same incentive for joining the grand coalition when the tariff compensates for the increased costs.In Scenario B, the market share is unbalanced, and as a result, MNO 3 with the highest market share must assume the greater portion of the costs.For that reason, when the tariff is very low, MNO 3 better prefers forming sub-coalitions, where, as explained before, the operational cost is not noticeably affected.Overall, some useful insights can be gained: • Cooperation is always beneficial for MNOs, even when the tariffs are low.• The profits that can be obtained in each particular coalition depend on the QoS guarantees offered to the end users (i.e., the R g ), represented by the slopes of the profit curves.Hence, the achieved QoS and the selected tariff represent the criteria for MNOs in order to choose partners for cooperation.• MNOs have a higher profit margin when cooperating with equal-sized operators.As seen in the case of Scenario A, the cooperation of MNOs with equal market share and bandwidth can yield higher profits with much lower tariffs, which is an appealing solution for both MNOs and end users.For each scenario, three tariff values are considered, starting from the minimum value that supports the formation of a stable grand coalition (i.e., the existence of the core).The grey areas represent the allocations within the core, whereas the white star represent the point for which the cost allocations coincide with the market share.As mentioned before, we can observe that the market share always belongs to the core.In Scenario A, where all MNOs have equal market shares, the core is symmetrical, whereas in Scenario B, the core moves towards the low-left corner, as the major part of the cost is assigned to MNO 3 (which holds the largest market share).Furthermore, for similar tariffs, the core dimension is much higher in Scenario A, due IEEE ICC 2017 Green Communications Systems and Networks Symposium

FHFig. 1 :
Fig.1: System Model problem, while[11] uses a listing algorithm.In the literature, it is common opinion that SC offloading by itself will not be sufficient for enabling the forecasted traffic and that mmWave bandwidth extension will be necessary.On the other hand, many works provide strategies for improving the spectrum utilization efficiency by means of opportunistic (Cognitive Radio) or cooperative spectrum sharing.Many papers have addressed the problem of cooperation among operators in case of traditional RAN.[12] studied the benefits of jointly deploying a new shared network,[13] investigated by means of non-cooperative game theory the feasible infrastructure sharing advantages thanks to base stations switching-off.On the other hand, works concerned with C-RAN are mainly focused on the BBU optimization in single operator case[8],[10],[11].[9]considers the two operators case but with separate spectrum licenses.In this work, we propose a novel scheme for studying the conditions for beneficial C-RAN sharing among coexisting MNOs.The objective is evaluating how QoS and profits can be improved thanks to a better spectrum utilization efficiency.The problem is modeled as a coalitional game and investigated for the three MNOs case, when different combinations of market and spectrum share are associated with each MNO.Our results demonstrate that cooperation is always advantageous and, depending on the user pricing adopted, the grand coalition can be preferred to smaller coalitions.The rest of the paper is structured as follows: Section II presents the system model.In Section III the sharing problem formulation and the proposed fair solution are introduced.In Section IV, the conditions for the stability of the sub and grand coalition are evaluated together with the gains provided by the proposed solution.Finally, Section V concludes the paper.

Figure 4
Figure4represents the cost allocations c m for each MNO m , with the allocations c 1 and c 2 represented in the x and y axis, respectively, and c 3 derived as c 3 = 1 − c 1 − c 2 .For each scenario, three tariff values are considered, starting from the minimum value that supports the formation of a stable grand coalition (i.e., the existence of the core).The grey areas represent the allocations within the core, whereas the white star represent the point for which the cost allocations coincide with the market share.As mentioned before, we can observe that the market share always belongs to the core.In Scenario A, where all MNOs have equal market shares, the core is symmetrical, whereas in Scenario B, the core moves towards the low-left corner, as the major part of the cost is assigned to MNO 3 (which holds the largest market share).Furthermore, for similar tariffs, the core dimension is much higher in Scenario A, due

TABLE I :
Scenarios

TABLE II :
System Parameters