mm‐Wave micro‐wave integrated Sub‐RAN for CRAN performance enhancement

Correspondence Brijesh Kumbhani, Department of Electrical Engineering, Indian Institute of Technology Ropar, Rupnagar, Punjab, 140001, India. Email: brijesh@iitrpr.ac.in Abstract Coordinated Multi-Point (CoMP) transmission in Cloud Radio Access Network (CRAN) requires a large amount of data transmission and processing within a coherence time window. Hence, CoMP transmission puts a lot of burden on the central processor and backhaul unit. Also, establishing CoMP for high-mobility users is challenging due to small coherence window and large beamforming overhead over mm-Wave transmission. This paper proposes a two-layer CRAN architecture with intelligent mm-Wave and micro-Wave allocation. A dual connectivity framework has been introduced to increase the coverage of high-mobility users. Further, it is shown that the proposed scenario reduces the load on the central processor and central back-haul. To avoid unnecessary handoffs, a mobility management algorithm is also proposed, which can provide seamless connectivity to the users irrespective of their velocity. Further, through simulation results, it is shown that the proposed network outperforms the existing CRAN framework.


INTRODUCTION
There has been a tremendous growth in cellular radio network [1]. This is made possible by seamless adoption of communication framework with increasing demand as well as continual efforts of the network operators [2,3]. In addition, academia and telecommunication industries have suggested several innovative network up-gradations to fulfil this ever increasing demand. Though, each generation of communication system came up with a series of network/architecture level up-gradations, Base Station (BS) densification has always been a key contributor in all the communication generations [4]. Furthermore, the advancement in the electronic circuit techniques made it possible to involve millimetre-Wave (mm-Wave) frequency band in cellular communications which made several fold increment in the resource bank in terms of bandwidth [5].
Traditional communication infrastructure is pillared on frequency reuse networks in which co-channel interference is avoided by smart frequency allocation [6]. This is because the BS density is sufficiently low (i.e. about 4-5 BSs/km 2 ) in the traditional cellular networks [7]. Further, in the upcoming generations, a continuous increment has been observed in BS density.
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. © 2021 The Authors. IET Communications published by John Wiley & Sons Ltd on behalf of The Institution of Engineering and Technology However, with time, people became more data hungry which further raised the demand to increase BS density, but a large BS density may lead to severe interference from the neighbouring BSs. Therefore, next generation communication infrastructure demands a coordinated framework among the BSs to avoid interference from near-by BSs. In addition, increase in the number of BSs, further increases the capital and operational expenditure of the communication network. The nextgeneration infrastructure has to be cost effective so that overall system remains economically sustainable. Cloud Radio Access Network (CRAN) strives to fulfil both the given conditions and is expected to become the next generation communication architecture due to several advantages over the existing network [8].
In CRAN, processing units of all the BSs are pooled to a central processor, also known as cloud processor, leaving behind radio units, known as Access Points (AP). Each AP is connected to the central processor via back-haul links (i.e. fiber optic cables). Thus, a centralized network is created. For transmission and reception in CRAN, each user is served by a number of preferred APs known as intra-cluster APs. Hence, this network is often referred to as user centric network. To utilize full potential of the available resources, CRAN is operated on frequency reuse-1 1 scenario which leads to severe co-channel interference from near-by APs. Therefore, Coordinated Multi-Point (CoMP) transmission technique has been introduced to prepare a coordinated road-map among the APs [9]. CoMP technique mainly uses the variants of either Coordinated Beamforming (CB) or Joint Transmission (JT). CB enables each clustered AP to form a beam in the direction of the user. Thus, it reduces the interference from the other APs. On the other hand, data is transmitted from all the clustered APs in JT. Basically, both the given schemes utilize the Channel State Information (CSI) to create beam or to establish JT. Hence, to achieve CoMP, large amounts of data (in the form of training signals and CSI) need to be exchanged among APs and the processor. Therefore, CoMP transmission is subjected to the large amount of meta-data transmission.
CRAN is highly capable to provide required Quality of Service (QoS) to each user and seems quite effective from user's view-point, but large amount of meta-data transmission makes CRAN highly inefficient from the processor's viewpoint. Though centralizing the computational unit has indeed provided several advantages (e.g. cost reduction, central governance, CoMP etc), the validity of CRAN becoming an efficient network is still under investigation. In spite of getting sufficient research spotlight, CRAN has several key challenges such as: • Large computational complexity: Centralizing the computational unit has shifted the burden of all the BSs to the central processor. Thus, the central processor is responsible for cluster formation, Training Resource (TR) allocation, CSI acquisition, and pre-coder design 2 , within the coherence window. Hence, the processor has to perform large amount of data processing within a fraction of second. In addition, computation and/or time complexity of the system also increases with increase in network size and sometimes, may alleviate the gain provided by the CoMP transmission. • Mobility management: CRAN relies on CoMP transmission which requires large amount of data transmission and processing within the coherence time window. However, the coherence window is comparatively small for a high-mobility user. Therefore, beam overhead is significantly large for highmobility users [10,11]. Thus, mobility management in CRAN demands some innovative solutions which posses low computational and time complexities to maintain fairness among the low-mobility users and the high-mobility users. • Large back-haul traffic: Every second, a lot of data is exchanged between the processor and APs. Specially, sharing of training signals require swift transmission so that the pre-coders may be designed within the coherence time frame which puts a lot of burden on the back-haul network and may exceed the back-haul capacity.

Related work
The challenges in the implementation of CRAN (discussed above) got significant research spotlight. Some recent works like [12][13][14] used multi-cloud network for load reduction of the processor where local cloud is shared by a set of BSs. Basically, it involves local BSs to perform the controlling tasks for the central load reduction. However, this framework suffers from the fact that it involves complicated resource allocation algorithm [15]. In contrast, Mei et al. proposed a new type of Mobile Edge Computing (MEC) architecture using dynamic computing offloading [16]. Further, it is shown that introducing MEC reduces computational complexity of the central processor. In a similar context, MEC offloading mechanisms are also investigated in the heterogeneous networks to achieve low-latency computing [17,18]. Several studies used the notion of caching for back-haul load reduction [19,20]. Chen et al., proposed a caching framework for offloading the back-haul of CRAN networks [19]. A novel scheme has been proposed which enables the processor to predict the content request distribution of each User Equipment (UE) with limited information on the network state. Further, simulation results have shown that the given scheme provides significant performance gains as compared to that of the conventional approaches. In [20], the authors proposed a cooperative caching paradigm for CRAN to enhance the hit ratio of caching and reducing back-haul load using joint caching at both the processor side and the AP side. A few works have studied mobility management in the centralized network; for example, a users' mobility prediction algorithm based on Markov model based method has been proposed in [21,22]. Specifically, the authors in [22] used previous mobility histories utilizing average distance travelled by the user to predict the next serving AP. Also, a scheme to reduce average number of hand-offs has been proposed in [11]. The given scheme used the notion of virtualized BS where adjacent transmission points are used to serve the UE when it traverses the cell edge. The authors solved the hand-off reduction optimization problem using an integer linear program.
Some recent works also studied the possibility of opportunistic mm-Wave and micro-Wave integration. In [23], Deng et al. studied two-hop downlink relaying in an integrated mm-Wave/sub-6 GHz network. The authors proposed a hierarchical architecture for scalable network management to provide multi-connectivity, beam selection, and interference management. mm-Wave/sub-6 GHz multi-connectivity with relaying shows considerable promise for reaching consistent user experience with high end-to-end throughput. Besides that, [24,25] proposed an integrated vehicular communication framework over mm-Wave and sub-6 GHz bands for vehicle to vehicle (V2V) and Vehicle to Infrastructure (V2I) connectivity, respectively. Moreover, both the works used IEEE 802.11p to obtain relative position information of the adjacent transmission points for beam overhead reduction.
Though quite impressive solutions have been given so far, still there is a huge margin for performance enhancement in CRAN. Also, most of the existing works target on the single objective problems. To the best of our knowledge, none of the existing work has proposed multi-objective solution to overcome various problems of CRAN. This paper proposes a twolayer CRAN architecture with intelligent resource utilization. The proposed scenario not only reduces the load on the central processor and central back-haul but also suggests a mobility management scheme for the centralized network, which can provide a seamless connectivity to the users irrespective of their velocity. Specifically, the contribution of this work can be summarized as follows: • We introduce a hybridized network pillared on high-density micro APs and low-density macro BSs. Large-capacity mm-Wave spectrum is assigned to the APs while more reliable sub-6 GHz band is utilized at the BSs. • To realize sub-RAN, each macro BS is equipped with high computational unit. In this way, a zonal Sub-RAN is created at the local level in which intra-zonal APs are connected to its macro BS via backhaul link. • The proposed network strives to provide high-speed links with required QoS to slow/stationary users and reliable communication to fast moving users. To achieve this and for efficient resource utilization, dual cluster formation scheme is proposed for fast-moving users and CoMP transmission is used for slow-moving users. • A velocity-based dynamic BS association is proposed. Further, to eliminate ping-pong effect, hysteresis-based switching scheme is introduced. • Further, it is shown that the proposed framework gives manyfold advantages; namely, it improves users' connectivity irrespective of their mobility, decreases the burden and back-haul load on the central processor, and experiences small number of inter-zonal handoffs as compared to the number of handoffs in a typical CRAN network.

PROPOSED ARCHITECTURE AND SYSTEM MODEL
We propose a hybrid CRAN consisting of Micro Stations (MSs) and Pico APs (PAs) distributed using Poisson Point Process (PPP), with densities and p , respectively, as shown in Figure 1. All the PAs are connected to the nearest MS via fiber optic back-haul. MSs are embedded with a computing unit and are responsible for CoMP management among the PAs located within their range. In this way, an individual MS with its subordinate PAs create a Zonal-RAN where all the clusters located within the zones are governed by the respective MS. Further, all the MSs are connected to the cloud processor/central processor which governs all the inter-zonal clusters. Specifically, the proposed network (given in Figure 1) makes a two-layer RAN. PAs are operated using mm-Wave and strive to serve stationary/slow mobility users because transmission over mm-Wave requires CoMP-based directional transmission using beam formation. On the other hand, MSs are assumed to operate using sub-6 GHz band and used to enhance the communication reliability. Since, sub-6 GHz band is limited, MSs are specifically used for high-mobility users. Thus, MSs are intended to serve intra-zonal high-mobility users at the zonal processor and thus, reduce the burden of central processor. Let,  and  p denote the set of MSs and PAs, respectively, while  denotes the set of users; then the signal received by i th low-mobility user (Y l i ) can be denoted as- First term on r.h.s. denotes the desired signal while second and third terms denote interference from intra-cluster APs and inter-cluster APs, respectively. In (1), C i is the serving cluster of u i which contains all the coordinated PAs, x i denotes the transmitted signal to i th user and n i denotes the additive Gaussian noise. w ik and g ik , respectively, denote the pre-coding vector and the channel gain between i th user and k th PA, where channel gain is defined as g ik =h m ik .
Here, A m 0 denotes the mm-Wave link path loss at the reference point of unit distance, r ik and h m ik , respectively, denote the distance and mmWave channel coefficient between i th user and k th PA, while m denotes the mm-Wave link path-loss exponent. It is considered that all the symbol are transmitted at constant power P t . Perfect CSI estimation is assumed and zero forcing pre-coding is considered; hence, interference from intra-cluster APs can be neglected. Defining Gaussian noise power by P n , signal to interference plus noise ratio (SINR) for u i can be given as- It has been discussed that establishing CoMP for highmobility users is quite difficult due to large Doppler shift (i.e. frequently changing scenario) and large beam overhead. Thus, we propose uncoordinated framework for high-mobility users in which all such users are to be served by a single PA (via Non-CoMP transmission). In addition, a sub-6 GHz connection is also provided from the most preferred MS. In this way, a dual connected framework is prepared for the high-mobility users. When k th PA and r th MS serve a high-mobility user, the signal received by i th user, u h i , can be denoted as- In the above expression, first two terms denote the desired and interfering signals from PAs, respectively. The subsequent two terms denote the desired and interfering signals from MSs, respectively. P ′ n denotes the joint noise power. h denotes the channel coefficient at sub-6 GHz and denotes the sub-6 GHz link path-loss exponent. Low-and high-mobility users are discriminated by superscripts l and h, respectively, while a k ∈  p and a r ∈  , respectively, denote PA and MS which provide the most dominant powers.
A UE is said to be in coverage if it satisfies the QoS requirements of the intended user. Accordingly, coverage probability for low-mobility users and high-mobility user can be given as- where, min denotes the SINR equivalent of QoS requirement of the user. Ψ l /Ψ h represents the instantaneous SINR experienced by the low-/high-mobility user and can be calculated as-

PROPOSED TRANSMISSION SCHEME
The proposed work aims to provide a seamless communication framework to each user without burdening the central processor. As discussed in the previous section, the proposed network posses two types of transmission schemes, one for low-mobility users and the other for high-mobility users. This section gives an insight into the initialization and management of the proposed transmission schemes which involves two step process, namely, cluster formation and mobility management.

Cluster formation
In the proposed scenario, the clusters are formed on the basis of mobility of the user. Moreover, a UE is eligible for CoMP transmission if its instantaneous velocity is less than threshold velocity v th . In the Low-Mobility Mode (LMM), user u l i selects several preferred PAs to form C i based on the received signal strength (RSS). Then, PAs in C i , serve u l i , either using CB or JT. PAs for cluster C i can be selected by using some appropriate scheme, for example, [26], which is briefly summarised in Appendix. In the High-Mobility Mode (HMM), high-mobility user u h i is jointly served by the most dominant MS and AP selected by using RSS information. A flow graph for cluster formation is given in Figure 2. Once, the initial connection of the intended UE is established, its velocity is continuously monitored for mode up-gradation which is governed by mobility management mode. Thus, at some instance, following transmission scenarios can be observed-

Mobility management
Mobility management takes care that the intended user operate in the appropriate mode and must be capable of avoiding unnecessary switching between different transmission modes. Flow graph for the proposed mobility management scheme is given in the Figure 3, where v i denotes instantaneous velocity of the intended user u i . The notion of hysteresis is used to eliminate ping-pong effect. Lower and upper limits of the hysteresis are calculated by using velocity margin Δ as, v l = v th − Δ and v h = v th + Δ, respectively. Further, to avoid frequent switching, an intentional delay is introduced to ensure that the user remains in the appropriate mode. Accordingly, when the velocity of u h i (i.e. v i ) goes below v l , it is not switched to LMM, immediately. Rather, velocity is further monitored in the contention window of size T s . If still the velocity remains below v l , the user is switched to CoMP LMM and vice versa is true for low-mobility users. Introducing delay (i.e. T s ) gives two-fold advantages. First, it helps to consider momentary fluctuations in the velocity. Second, once the contention period is over, the subroutine is called to update the transmission points for next contention window.

RESULTS AND DISCUSSION
In this section, performance analysis of the proposed network has been discussed using the simulation results performed on MatLab. A hybrid network (with MSs and PAs) within the area of coordinates [−500, 500] × [−500, 500] meters 2 has been considered, while all the simulation parameters for both MSs and PAs are listed in Table 1. The users are distributed according to PPP model, with mean = 1000 within the considered area. Moreover, all the plots have been obtained as a result of averaging the performance by simulationg the proposed network for 10 5 times; hence, the plots depict mean performance of the considered frameworks. The proposed framework is compared with the existing CRAN network (i.e. considered as baseline approach) where baseline approach uses single centralized processor for all CRAN operations. Hence, for the baseline approach, all the CRAN operations, that is, clustering, training resource allocation, pre-coder design etc., are done by the central processor. To verify the proficiency of the proposed network, several performance metrics have been introduced and used for performance comparison.

Coverage probability
This subsection measures the degree of fairness achieved by the proposed network in satisfying the QoS to the low-mobility users and the high-mobility users. Figure 4 shows average CP for different users against min . The curve for ICT or OCT depicts the CP experienced by low-mobility users while rest of the curves denote the CP for high-mobility users in two different scenarios, that is, proposed network (IHT or OHT) and existing CRAN network (Non CoMP Transmission). It can be observed that for the proposed scenario, the CP for high-mobility users is significantly close to the coverage of low-mobility users. In other words, the proposed scheme maintains the degree of fairness between high-mobility users and low-mobility users. However, the coverage probability of high-mobility users in the baseline approach lie significantly below that of the proposed approach. Specifically, the enhancement in coverage performance is achieved due to the notion of dual connected framework, which subsequently creates a virtual diversity network.

Central load analysis
CRAN is a user centric network. Thus, the clusters in CRAN are inevitably overlapped. Lack of isolation in the clusters increases the complexity of the network. For example, central processor alone is responsible for all the CoMP operations, including training resource assignment. Hence, traditional CRAN exerts huge burden on the central processor. Thus, central load reduction is of foremost importance. The proposed approach shares the burden of central processor with the MSs by embedding each MS with a computational unit. Each MS is able to handle intra-zone operations without burdening the central processor. Performance metric for central load reduction can be defined as- where, i and o denote the number of intra-zone clusters and out of zone clusters, respectively. Since, out of zone clusters are handled by the central processor, o clusters are mainly responsible for burdening the central processor.
Note that min specifies QoS requirement (i.e. SINR in dB) of the users. An increase in the value of min forces to incorporate more number of serving points in each cluster (i.e. upto three), which further increases the chances of creating inter-zone clusters. Thereby, an increase in min results in the excessive burden on CP which is reflected in terms of central load coefficient ( ). Figure 5 shows the plots for central load coefficient (Γ) for slow and fast moving users against min . It can be observed that the value of Γ lies significantly below unity, even for higher min values. Moreover, an increase in the values of Γ for larger min is obvious due to the increased possibility of creation of inter-zone clusters. Overall, the plots depict that the proposed scheme significantly reduces the burden on CP. Back-haul load analysis: Back-haul mainly carries meta-data information (eg., CSI) from PAs to the serving processor. The amount of data carried by the central back-haul network is directly related to the number clusters handled by the central processor. Note that the central back-haul load decreases to Γ times the back-haul load carried in the baseline approach. Central backhaul load reduces mainly due to the relaxation in the responsibility of CP, i.e. CP handles only inter-zone clusters for the proposed scenario. Similar to central load coefficient, backhaul traffic increases with the inter-zone clusters. The results given in Figure 5 also hold true for back-haul data traffic, which indicate that proposed scheme significantly reduces the backhaul traffic.

Hand-off analysis
The proposed network provides provision for mobility management, where users are served on the basis of their mobility. Further, to avoid unnecessary switching between the proposed transmission modes, the notion of hysteresis is used as discussed in Section 3.2. Threshold velocity is calculated as, v th = mean [0 , v max ]. User velocity is assumed to be uniformly distributed, [0 , v max ]. Furthermore, a contention window T s is introduced to verify the change in velocity. The contention window is considered as the time equivalent of the half inter-PA Inter Site Distance (ISD) travelled by the fastest moving user so that the users may connect the next PA, before the connection from previous PA breaks. Figure 6 shows average number of hand-offs per user during a call against QoS requirement min of the users (for v max = 126 km/h). The above curve depicts the number of hand-offs (switching within PAs), while the curve given below denotes number of inter-zone hand-offs. Intra-zonal hand-offs are handled by local processors at MSs, while inter-zonal hand-offs are taken care by CP. Thereby, the gap between both curves shows reduction in central hand-offs. Overall, the curves depict that the proposed approach significantly reduces central hand-offs.

CONCLUSION
This paper proposes a novel mm-Wave micro-Wave integrated CRAN to improve the performance of conventional CRAN. The proposed architecture introduces two-layer dual connected framework for efficient service to users with low as well as high mobility. The proposed architecture was simulated for performance comparison with conventional CRAN. The results show that the proposed framework reduces the load on central processor and central back-haul with required QoS. Further, the results also show that the proposed scheme can provide a seamless connectivity to the users irrespective of their velocity through proposed mobility management. The performance comparison in terms of the number of hand off shows that the proposed technique reduces the average number of hand offs for fast users. So, to conclude, the proposed network architecture is found to outperform the existing CRAN architecture in terms of QoS, load on central processor and the number of handoffs through mm-Wave micro-Wave integrated Sub-RAN architecture with mobility management unit.