On the beneﬁt of inter-operator cooperation in C-RAN

Cooperation of co-located mobile network operators can provide potential beneﬁts for the capacity expansion without further densiﬁcation of radio nodes. However, such ben-eﬁts need to be scrutinised for coexisting cloud radio access network (C-RAN) operators because the inter-operator cooperation may compromise the superb interference coordination capability that each C-RAN has. Furthermore, altering C-RAN infrastructure for the cooperation incurs a high investment cost. In this paper, quantitative gain of inter-operator coordination strategies is evaluated to provide the C-RAN operators with a guideline on their cooperation decisions. The coordination strategies encompass dynamic user association and the spectrum sharing which aim at maximising the total user throughput. A heuristic algorithm is proposed that reduces the computational burden of the coordination signiﬁcantly. Numerical results suggest that the inter-operator cooperation is beneﬁcial particularly when the network size of each operator tends to be highly asymmetric. It is also veriﬁed that the users who belong to smaller network attain more coordination gains.


INTRODUCTION
Faced with ever-growing mobile data traffic resulting from the use of a wide variety of newly-emerging smart devices in the future 5G era, cloud radio access network (C-RAN) has been viewed as a promising countermeasure to address the detrimental influence of interference [1], [2]. All the baseband processing functionalities in C-RAN are centrally carried out at the virtualised baseband unit (BBU) pool, while distributed remote radio heads (RRHs) merely activate the radio frequency processing. However, a dramatic increase of traffic demand after an initial C-RAN deployment is still one of the practical challenges to be overcome [3], [4]. It is commonly known that collaboration among multiple mobile network operators (MNOs) in a specific service region is a cost-effective solution to meet the growth of user demands without further expenditure to densify the number of wireless nodes [5], [6]. The main concept of inter-operator cooperation we will leverage hereafter is that all the user equipments (UEs) in shared cellular networks can be served by any base stations (BSs) belonging to cooperating MNOs within the overlapped coverage area, in a similar manner to the conventional roamingbased network sharing [6], in addition to the spectrum sharing between MNOs [7][8][9]. By virtue of inter-operator cooperation, capacity expansion for each MNO can be fulfilled while the network deployment remains unchanged.
This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. © 2020 The Authors. IET Communications published by John Wiley & Sons Ltd on behalf of The Institution of Engineering and Technology

Related work
In recent years there have been several efforts to address the multi-operator cooperation problems. However, most of the work focused on traditional cellular networks or a single BS architecture. The authors of [10] described the benefits of network sharing between different operators and emphasised the role of the regulator in the shared networks. In [11], [12], various infrastructure sharing scenarios for the multiple coexisting MNOs such as passive RAN sharing and roaming-based sharing have been presented. Strategies for inter-operator roaming and RAN sharing in the same geographical area were proposed in [6], [13]. Cost and energy saving in heterogeneous network environments through multi-operator cooperation was investigated in [14], [15]. The authors of [16] analysed an infrastructure sharing problem among multiple seller MNOs and a buyer MNO based on stochastic geometry approach. In [17], [18], the authors considered a multi-operator network architecture that a single BS simultaneously serves different MNOs' subscribers and provided a framework for maximising energy and spectral efficiency with multiple quality of service and fairness constraints. More recently, the authors of [19] considered the potential benefits of spectrum sharing among different colocated MNOs by leveraging on network virtualisation from the energy efficiency perspective. In addition, taking into account the guarantee of information privacy, a coordinated jamming strategy for the uplink of multiple C-RAN operator system was proposed in [20].
To the best of the authors' knowledge; however, there has been scarce study on the effect of cooperation amongst multiple C-RAN operators which possess their exclusive BBU pools, RRH clusters, and the corresponding subscribers. In [21], cooperation of C-RAN operators at the contractual level, for example, trust building and security, was studied and a trusted cooperation platform was designed. But, C-RAN inter-operator cooperation at the level of baseband signal processing has not been studied thus far.

Motivation of this work
Joint baseband signal processing between different C-RAN operators should be accompanied by expensive investments to merge the BBU pools which are geographically separated and to rearrange the fiber fronthaul links accordingly. In other words, combining the infrastructure of different MNOs at the level of BBU pools, called hereafter tight coordination and illustrated in Figure 1(a), is likely to be costly. This is only worthwhile if attainable performance gain from the coordination is relatively high. In this context, we are motivated to consider an alternative loose coordination shown in Figure 1(b) for which one BBU pool is linked to the other BBU pool over X2 interface passing through the intermediate common server to coordinate the user association and bandwidth (BW) allocation between C-RAN operators. Although BBU pools in the loosely coordinated multi-operator C-RAN are assumed not to fully share channel state information (CSI), in turn leading to less performance benefits, the required extra cost for implementing the loose coordination will be lower than that of the tight coordination. In fact, a practically meaningful potential benefit from the operator's point of view is not obvious without an appropriate cost-benefit framework. Particularly it is the main goal of this paper to identify and analyse trends in which operators can achieve how much coordination gains under which situation. Therefore, we can determine the most favorable C-RAN operator cooperation strategy only when given the specific coordination cost model. Accordingly, it is crucial for MNOs to consider the loose coordination as well at the expense of a loss of coordination gain, along with the tight one. In short, we categorise the inter-operator cooperation as tight and loose coordination in terms of the extent of infrastructure alteration and corresponding interference coordination capabilities for multioperator C-RAN.

Contributions
In this paper, we consider the inter-operator cooperation between C-RAN MNOs with sharing agreements to accommodate the growing traffic demand in each MNO, which raises the following key question; under which condition and to which operator, how much performance gain is attainable by cooperation between C- RAN operators? Note that, from the business perspective, MNOs intend to carry out a cooperation with a competitor only if the expected economic gains from inter-operator coordination are large enough with regard to their extra investment expense [22]. This is because the cost and the corresponding performance gain may differ depending on the coordination method. Most operators would be unwilling to implement the cooperation with others if they can only achieve marginal performance gain.
The main contributions of this paper can be summarised as follows: • We evaluate the attainable performance gain of interoperator cooperation depending on the extent of coordination. The total expenditure for cooperation is operator-and environment-specific, and thus is beyond the scope of the paper. We emphasise that the analysis of the trade-off between the cost and performance benefits is straightforward once the MNO establishes the relevant cost model and figures out the coordination costs (see, e.g. [5]). • To make a reasonable decision on whether to facilitate inter-operator cooperation, and what coordination strategy a MNO should choose, it is necessary to quantify the gains of feasible coordination options. For this purpose, we firstly elaborate on two loose coordination strategies based on the assumption that two C-RAN operators coordinate resource allocation through a common coordination server. Specifically, we focus on a design of a joint UE association among C-RAN operators and dynamic BW division for the case of loose coordination to optimise the overall system performance. In this regard, we formulate the corresponding optimisation problems. In the conventional cellular structure, the downlink dynamic cell association and resource allocation problems have been studied extensively, for example, user-BS association in heterogeneous networks [23][24][25], BW allocation for downlink OFDM [26][27][28][29][30], and joint optimisation of association and BW allocation [31], [32]. However, the existing approaches in conventional cellular networks [23][24][25][26][27][28][29][30][31][32] cannot be directly applied to this work because the high-complexity precoding design for each individual C-RAN operator under per-RRH power constraints should be accompanied with every feasible set of UE association and BW allocation. Thus, it is difficult to find the optimal solution in C-RAN environments by utilising the methods in the previous work. We propose a heuristic algorithm which has low complexity to solve them. • Applying the proposed algorithm, the quantitative performance advantage of loose coordination between C-RAN operators is evaluated. Next, we consider the tight coordination strategy that features full cooperation with global precoding matrix via the integrated BBU pool, as an upper bound of inter-operator cooperation, and compare it with loose coordination strategies as well as the no cooperation case.
The remainder of the paper is structured as follows. Section 2 provides the system model of coexisting C-RANs, and two different coordination scenarios. In Section 3, we discuss several coordination strategies for C-RAN operators. Section 4 proposes a heuristic algorithm to assess the performance of loose coordination. In Section 5, we present simulation results followed by concluding remarks in Section 6.

SYSTEM MODEL
We consider a downlink of coexisting two C-RAN named C 1 and C 2 , respectively, managed by different MNOs as the default scenario. For this, the C 1 consists of a BBU pool 1 and M distributed RRHs, while the C 2 also has exclusive BBU pool 2 and N RRHs. Each MNO uses the assigned BW, and independently employs the zero-forcing (ZF) beamforming as a precoder without cooperation between them. For simplicity, each UE is assumed to be always jointly served by all RRHs belong to the same C-RAN operator. Thus, inter-user interference within the same C-RAN can be completely eliminated, which leads to the noise-limited scenario for each MNO. Meanwhile, in the shared service domain of both MNOs, K UEs are randomly distributed and each UE can be determined to be a subscriber of either C 1 or C 2 . The subscriber decision rule will be discussed in more detail in Section 3.1. It is assumed that all RRHs and UEs are equipped with a single antenna for simplicity and thus at most M + N UEs will be randomly chosen from all K UEs at each scheduling epoch where we let the set of indices of RRHs in C-RAN C 1 , RRHs in C-RAN C 2 , and total UEs be denoted Throughout this paper, we will cover two major coordination scenarios between two initial C-RAN deployments serving the corresponding subscribers: (i) tight coordination and (ii) loose coordination. It is noteworthy that tight and loose, respectively, mean high and low level of an effort to carry out coordination. In Figure 1(a), a tight coordination requires two different MNOs' BBU pools to be combined to allow full cooperation with perfectly global precoding matrix. A loose coordination illustrated in Figure 1(b) retains the original BBU pools of two MNOs. Instead, respective C-RAN needs the inter-BBU connections, and it is assumed that there exists a central entity that is in charge of collecting and sharing fragmentary information of the opposite C-RAN to jointly design both UE association and BW allocation.

INTER-OPERATOR COORDINATION STRATEGIES
We describe in detail several feasible coordination strategies for multi-operator C-RAN environments according to the level of coordination which depends on the class of shared information by means of network architecture modification. Then, corresponding performance metrics are presented for each strategies to assess the achievable performance.

No cooperation: Baseline strategy
First, we consider the completely separate C-RANs where each C-RAN operator individually provides service to their original subscribers without any cooperation. Moreover, it is assumed that the respective C-RAN operators utilise the exclusive spectrum as their BW W 1 and W 2 , which ensures no inter-operator interference. Also, there exists no common server between both operators so that they do not have an ability to implement the inter-operator cooperation. Here, to enable the equivalent comparison, we assume that the original number of subscribers is proportional to the number of RRHs within each C-RAN. This is due to the fact that C-RAN operator with fewer original subscribers cannot afford to invest enough in their infrastructure, that is, increase the number of RRHs within C-RAN, from a cost-benefit perspective. Based on this assumption, sets of UE indices assigned to either C 1 or C 2 at each scheduling epoch is predetermined at random, and these are, respectively, denoted as  1 = {U 1,1 , U 1,2 , … , U 1,| 1 | } and  2 = {U 2,1 , U 2,2 , … , U 2,| 2 | }. Note that if the number of assigned UEs is larger than the number of RRHs in that C-RAN, that is, | 1 | > M and/or | 2 | > N , surplus UEs are unable to be served at that epoch, which causes the instantaneous per-user rate to be 0. Instead, those will be served at the next scheduling epoch.

Loose coordination I: Dynamic bandwidth allocation
Here, we introduce the first loose coordination strategy in which each C-RAN operator has the potential to provide service to any subscribers of both operators, while their exclusive BWs W d, 1 and W d,2 are flexibly split on the shared spectrum W = W 1 + W 2 . More specifically, the common server shown in Figure 1(b) not only enables entire RRH clusters of different MNOs to be accessible among total UEs within the same coverage region, but also conducts the optimisation of BW division for two MNOs towards achievable rate enhancement. Because C 1 and C 2 occupy the exclusive subbands in a similar manner to baseline scenario, it is clear that there is no inter-operator interference for this strategy.
The problem of UE association and dynamic BW allocation can be formulated as the maximisation of the total sum-rate under the use of ZF precoder subjected to the per-RRH transmit power constraint as follows: x k , y k ∈ {0, 1}, ∀k ∈ , where R loose-I total is the total sum-rate of both MNOs and is given by where  1 and  2 represent the set of UEs assigned to C-RAN C 1 and C 2 , respectively. In (2), the signal-to-noise ratio (SNR) of UE k is expressed as where x k and y k are the binary decision variables for UE k to be served by C 1 and C 2 , respectively, P max is the maximum transmit power of RRH, W d, 1 and W d,2 are, respectively, the private BW of C 1 and C 2 , 2 is the background noise spectral density, h k = [h 1k , h 2k , … , h (M +N )k ] T denotes the CSI vector from all the RRHs of two C-RAN operators to the UE k, and w k = [w 1k , w 2k , … , w Mk ] T and w k = [ŵ (M +1)k ,ŵ (M +2)k , … ,ŵ (M +N )k ] T are the ZF precoding vectors corresponding to  1 and  2 , respectively. It is noteworthy that W d, 1 2 and W d,2 2 in (3) mean the total noise power over the separate BW of C 1 and C 2 , respectively, because 2 denotes the noise power per unit of BW. The problem (1) is a mixed-integer combinatorial optimisation problem, and thus it is complex and time-consuming to solve in general. For this reason, we will propose the heuristic approach to obtain the near-optimal solution of (1) in the following Section 4.

Loose coordination II: Full spectrum sharing
We discuss the second loose coordination, that is, full spectrum sharing approach which corresponds to the following optimisation problem formulation with the same constraints in (1): where the total sum-rate for this strategy is expressed as . (5) Unlike loose coordination I, inter-operator interference should be considered because full spectrum W is simultaneously utilised by both operators. Accordingly, the signal-to-interference-plusnoise ratio (SINR) in (5) is represented as For this strategy, we concentrate only on the adaptive UE association and precoding design without having to consider the BW division design.

Tight coordination
As the best scenario, that is, an upper bound in terms of achievable rate, an integrated C-RAN including one combined BBU pool is taken into consideration as illustrated in Figure 1(a). In the tight coordination, two separate C-RAN MNOs invest a significant amount of their money into reconstitution of united BBU pool and implement full cooperation with each other with the use of global precoding matrix. Then, the total sum-rate can be written as

HEURISTIC ALGORITHM FOR LOOSE COORDINATION
Since the optimisation problem in (1) involves substantially numerous possible combinations of solutions when M , N , K , and the value of increment of the allocation ratio of BW is especially large, the exhaustive search requires an intractable computation time. Hence, we propose an alternative low-complexity heuristic algorithm based on a simple intuition to find a suboptimal solution while reducing the processing time in numerical experiments. By leveraging the conventional greedy method, we first decide which UEs would be associated with either C 1 or C 2 . To be specific, for each UE, we calculate the channel gain with C 1 and C 2 defined as g 1 k and g 2 k , respectively, and then find the larger one. After this, we sort all the corresponding K channel gain in descending order, and associate UE in sequence, according to the sorted order. Here, the arbitrary UE should be re-associated with C 2 if the same number of UEs as the value of M have already been served by C 1 before, and vice versa. A detailed proposed UE association algorithm is summarised in Algorithm 1.
With the determined  1 and  2 , we then design how the entire BW will be dynamically allocated to each operator in Algorithm 2 Algorithm 2: Dynamic bandwidth allocation Algorithm 2. Assuming an equal transmit power from each RRH to all the served UEs in each operator, which are given as , respectively, the approximated total sum-rate can be expressed as (8) w.r.t. r which indicates the allocated BW ratio of operator 1. Due to the fact that the second derivative ofR(r ) is always negative, (8) is strictly concave. Therefore, by differentiating (8) with the iterative change of where the increment is Δ , it is possible to obtain the optimal BW allocation ratio on domain r ∈ (0, 1) based on an iterative numerical method.
Finally, substituting the obtained  1 ,  2 , and from Algorithm 1 and 2 into (2), where W d,1 = W and W d,2 = W (1 − ), we find the optimal ZF precoders w k andŵ j to satisfy (1) under per-RRH power constraints by using a convex optimisation technique aiming at maximisation of sum-rate for all UEs served by both operators.
For the problem in (4) of loose coordination II,  1 and  2 are decided in the same heuristic manner to that of loosecoordinationI . In contrast with (1), we need to further take account of interoperator interference, whereas dynamic BW allocation in Algorithm 2 does not need to be considered in loose coordination II. In this context, the design of ZF precoders w k andŵ j for both C-RANs when employing loose coordination II can be achieved by regarding the worst-case optimisation problem as follows: where P max ∑ n∈ |h * nk | 2 and P max ∑ m∈ |h * m j | 2 imply the predictable worst interferences from the opposite C-RAN operator. Because the exact inter-operator interference from the opposite C-RAN operator remains unknown before precoder design of the opposite operator is performed, we need to solve the optimisation problem of loose coordination II assuming the worst inter-operator interference scenario, that is, maximum transmit power used by the opposite C-RAN.

NUMERICAL RESULTS
In this section, we provide numerical results that present the performance benefits from the inter-operator cooperation for two different C-RAN operators. Specifically, we numerically compare representative coordination strategies according to the coordination level between two C-RAN MNOs given in Section 3. Here, all RRHs and correponding subscribers of each C-RAN operator are randomly distributed within 50 × 50 square meters area, and Rayleigh small scale fading with zero mean and unit variance is assumed in simulations. The detalied simulation parameters are summarised in Table 1. Note that in this simulation our objective is maximising the total user throughputs of two C-RAN operators under the use of ZF precoding, and then identifying which coordination strategy is advantageous from a cost-benefit perspective. For reliable numerical evaluation, we perform Monte Carlo simulation with 10 4 iterations. Figure 2 shows that the curve of proposed heuristic algorithm is very close to that of optimal solution via exhaustive search in terms of the average per-user rate when using loose coordination I, where M = 7 and N = 3. From the observation, the effectiveness of proposed algorithm can be verified, which allows us to employ the proposed heuristic method throughout the numerical evaluations with much less time.
In Figure 3, we quantify the average per-user rate for the scheduled UEs as a function of K under the different strategies. It is observed in Figure 3(a) that the performance of loose coordination I drops more sharply than that of tight coordination as the number of UEs increases. Especially, the performance of loose coordination I approaches to that of no cooperation case for large K . This indicates that the tight coordination is more advantageous to both operators in high traffic demand condition due to performance gain from loose coordination is very marginal. On the other hand, we can see that the result of loose coordination I is closer to that of tight coordination over the   Figure 3(b). From this observation, it is sufficient to implement loose coordination I when the network size, that is, the number of RRHs of two operators is asymmetric. In addition, there is a range of K in which loose coordination II is inferior to no cooperation, for example, K > 4 in Figure 3(a). This implies that the risk of inter-operator interference due to simultaneous use of full spectrum by both operators can exceed the additional benefits from cooperation particularly if the overall network is less asymmetric for both operators. From both Figure 3(a) and 3(b), we notice that the asymmetry level of operators may affect the inter-operator cooperation capability. In this regard, Figure 4 plots the average per-user rate for original UEs of MNO 1 according to changes in the pair (M, N ) for low (K = 3) and high (K = 10) user demand environments. This figure explicitly demonstrates the influence of different network size on the cooperation incentive of operator 1. It is worth noting in both Figure 4(a) and 4(b) that attainable gain with loose coordination I with regard to no cooperation case, from the perspective of the subscribers of C 1 , decreases as the network size of C 1 increases until it reaches the same value as N , that is, the network size of opposite operator. Particularly, Figure 4(b) shows the average per-user rate of operator 1 with loose coordination I almost meets that of no cooperation when the network sizes are symmetric. After the point of symmetry, the performance gap between loose coordination and no cooperation begins to increase again along with the increase in . However, the larger network size leads to relatively less performance improvement for their original subscribers. For example, achievable gains from loose coordination I in terms of the average per-user rate of operator 1 are approximately 3.5 and 1.7 for M = 1 and M = 9, respectively, in Figure 4(a). Then, we can surmise that operator collaboration results in significant performance improvements when both operators' network sizes are remarkably different. This can be explained by two aspects; first, the cooperation between operators can allow the original UEs of each MNO to use BW more flexibly; second, inter-operator cooperation can provide the UEs of small size C-RAN with extra opportunities to be served by the other operator, thereby dealing with the instantaneous increase in traffic demands. Moreover, as expected, the curve of tight coordination is nearly constant regardless of (M, N ) since all the RRHs of both C-RANs are controlled by one combined BBU pool in tight coordination. Figure 5 illustrates the 5-percentile, 50-percentile, and 95percentile user throughput versus the (M, N, K ) tuple. It is clearly observed that the median user throughputs, that is, 50-percentile throughput with tight coordination and loose coordination I greatly improve compared to that achieved without cooperation. As discussed above, the 50-percentile user throughput gain from loose coordination I becomes larger as the asymmetry level grows. Similarly, high SINR users represented by 95-percentile user throughput can also reap more performance benefits from inter-operator cooperation with high asymmetry level. In Figure 5(a), while the tight coordination always outperforms the case of no cooperation for celledge, that is, low SINR users, loose coordination I is inferior to no cooperation for small K , for example, K = 4 in Figure 5(a). This might be an inherent limitation of greedy heuristic algorithm, which requires an improvement in the future study. Also, when the number of total UEs is large, it is seen that 5-percentile user throughput without cooperation is 0 regardless of (M, N ). From this result, we suggest that it would be beneficial to adopt the cooperation between operators especially for a lot of UEs if operators intend to guarantee the

CONCLUDING REMARKS
We investigated the potential gains from inter-operator cooperation between two C-RANs. We developed several coordination strategies in terms of the level of infrastructure modification and corresponding cooperation abilities. Tight coordination allows both operators to facilitate full cooperation by means of the global precoder at great expense. On the other hand, operators share the common coordination server and can realise a joint design of UE association and dynamic BW allocation in the loose coordination at relatively low cost. To explore the usefulness of inter-operator cooperation in C-RAN, we compared the quantitative coordination gain of these strategies with respect to the case of non-cooperation by applying the proposed heuristic algorithm. Particularly, we identified which coordination scheme is preferable under which network conditions. Numerical results suggest that the loose coordination with dynamic BW division gives satisfactory gains for low traffic demands, whereas C-RAN operators may prefer the tight coordination when user demands increase. In addition, it is observed that the asymmetry level of participating MNOs' network size has crucial influence on not only the overall system performance but also the user rate of individual operator. To be specific, highly asymmetric scenario leads to substantial inter-operator cooperation gains, especially for the operator who owns comparatively small number of RRHs. These results provide C-RAN operators with the quantified guidelines to make a decision on whether it is better to invest in cooperation with other operator by considering the expected return.
As a further study, we will present a large-scale analysis of inter-operator cooperation in C-RAN environments. For this, it is worth applying the low-complexity precoding scheme and establishing a relevant analytic model to provide general guidelines and quantified insight for more realistic ultra-dense C-RAN.