Architectures, Key Techniques, and Future Trends of Heterogeneous Cellular Networks
- 154 Downloads
HetNets are seen as new network paradigm evolutions to the fifth-generation wireless systems, which can cost-efficiently improve system coverage, capacity, latency, and so on.
With the explosive increase in mobile data traffic, operators have to continuously improve network performance. As a promising solution, the HetNet is a new technology that can cost-efficiently improve system coverage and capacity (Peng et al., 2015a). Communication nodes with high transmit power, such as MBSs, are deployed for large coverage and high capacity, while LPNs, such as SCAPs, help to supply service in the coverage of some hotspots. By deploying additional LPNs within the local-area range and bringing LPNs closer to end-UEs, HetNets can potentially improve spatial resource reuse and extend the coverage, thus allowing future cellular systems to achieve higher data rates while retaining the uninterrupted connectivity and seamless mobility of cellular networks.
The LPN is identified as one of the key components to increase the capacity of cellular networks in dense areas with high traffic demands. When traffic is clustered in hotspots, such LPNs can be combined with HPN to form a HetNet. HetNets have advantages of serving hotspot customers with high bit rates through deploying dense LPNs, providing ubiquitous coverage, and delivering the overall control signallings to all UEs through the powerful HPNs. Actuarially, too dense LPNs will incur severe interference, which restricts performance gains and commercial developments of HetNets. Therefore, it is critical to control interference through advanced signal processing techniques to fully unleash the potential gains of HetNets. The CoMP transmission and reception is presented as one of the most promising techniques in 4G systems. Unfortunately, CoMP has some disadvantages in real networks because its performance gain depends heavily on the backhaul constraints and even degrades with increasing density of LPNs. Further, it was reported that the average SE performance gains from the uplink CoMP in downtown Dresden field trials were only about 20% with non-ideal backhaul and distributed cooperation processing located on the base station.
To overcome the SE performance degradations and decrease the energy consumption in dense HetNets, a new paradigm for improving both SE and EE through suppressing inter-tier interference and enhancing the cooperative processing capabilities is needed in the practical evolution of HetNets. Meanwhile, cloud computing technology has emerged as a promising solution for providing high energy efficiency together with gigabit data rates across software-defined wireless communication networks, in which the virtualization of communication hardware and software elements place stress on communication networks and protocols.
To achieve these goals, the C-RAN has been proposed as a combination of emerging technologies from both the wireless and the information technology industries by incorporating cloud computing into RANs (Peng et al., 2016a). C-RANs have come with their own challenges in the constrained fronthaul and centralized BBU pool. A prerequisite requirement for the centralized processing in the BBU pool is an interconnection fronthaul with high bandwidth and low latency. Unfortunately, the practical fronthaul in C-RANs is often capacity and time-delay constrained, which has a significant decrease on SE and EE gains.
Consequently, H-CRANs are proposed in Peng et al. (2014) as cost-effective potential solutions to alleviating inter-tier interference and improving cooperative processing gains in HetNets through combination with cloud computing. The motivation of H-CRANs is to enhance the capabilities of HPNs with massive MIMO and simplify LPNs through connecting to a “signal processing cloud” with high-speed optical fibers. As such, the baseband datapath processing and the radio resource control for LPNs are moved to the cloud server so as to take advantage of cloud computing capabilities.
Unfortunately, H-CRANs are still challenging in practice. First, since the location-based social applications become more and more popular, the traffic data over the fronthaul between RRHs and the centralized BBU pool surges a lot of redundant information, which worsens the fronthaul constraints. Besides, H-CRANs do not take full advantage of processing and storage capabilities in edge devices, such as RRHs and “smart” UEs, which is a promising approach to successfully alleviating the burden of the fronthaul and BBU pool. Moreover, operators need to deploy a huge number of fixed RRHs and HPNs in H-CRANs to meet the requirements of peak capacity, which makes a serious waste when the volume of delivery traffic is not sufficiently large.
To solve such challenges in H-CRANs, revolutionary approaches involving new RAN architectures and advanced technologies need to be explored. Fog computing is a term for an alternative to cloud computing that puts a substantial amount of storage, communication, control, configuration, measurement, and management at the edge of a network, rather than establishing channels for the centralized cloud storage and utilization (Bonomi et al., 2012). Inspired by the advanced characteristics of fog computing and to alleviate the existing challenges of H-CRANs and take full advantages of local caching, the F-RAN architecture has been proposed in Peng et al. (2016b).
Abbreviations in this chapter
Third generation partnership project
Low power node
Convolutional neural network
Macro base station
Coordinated scheduling/coordinated beamforming
Collaborative radio signal processing
Quality of experience
Collaborative radio resource management
Quality of service
Cloud radio access network
Radio access network
Deep reinforcement learning
Recurrent neural network
Small cell access point
Fog access point
Fog computing-based radio access network
Fog user equipment
Heterogeneous cloud radio access network
Wireless local area network
High power node
The Architecture of Traditional HetNets
The Architecture of H-CRANs
However, different from C-RANs, the BBU pool in H-CRANs is interfaced to HPNs for mitigating the cross-tier interference between RRHs and HPNs through the centralized cloud computing-based cooperative processing techniques. Further, the data and control interfaces between the BBU pool and HPNs are added and denoted by S1 and X2, respectively, whose definitions are inherited from the standardization definitions of 3GPP. Since the voice service can be provided efficiently through the packet switch mode in 4G systems, the proposed H-CRAN can support both voice and data services simultaneously, and the voice service is preferred to be administrated by HPNs, while the high-data packet traffic is mainly served by RRHs.
Compared with the traditional C-RAN architecture, the proposed H-CRAN alleviates the fronthaul requirements with the participation of HPNs. Owing to the incorporation of HPNs, the control signallings and data symbols are decoupled in H-CRANs. All control signallings and system broadcasting information are delivered by HPNs to UEs, which simplifies the capacity and time-delay constraints of the fronthaul links between RRHs and the BBU pool, and can make RRHs active or sleep efficiently to save the energy consumption. Further, some burst traffic or instant messaging service with a small amount of data can be efficiently supported by HPNs. The adaptive signalling/control mechanism between connection-oriented and connectionless is supported in H-CRANs, which can achieve significant overhead savings in the radio connection/release by moving away from a pure connection-oriented mechanism. For RRHs, different transmission technologies in the PHY layer can be utilized to improve transmission bit rates, such as IEEE 802.11 ac/ad, millimeter wave, and even optical light. For HPNs, the massive MIMO is one potential approach to extend the coverage and enrich the capacity.
The Architecture of F-RANs
Different from H-CRANs, a large number of CRSP and CRRM functions originally located in the cloud computing layer are shifted to F-APs and F-UEs, which alleviates the burden on the fronthaul and BBU pool. Meanwhile, the limited caching in F-APs and F-UEs can make some requests served locally. The benefit of this enhancement is significant, considering that the location-based social applications are becoming popular and much redundant information over the fronthaul links will worsen the fronthaul constraints. Furthermore, local caching, local signal processing, and local radio resource management can better satisfy the low latency requirement of some applications in 5G era.
Cooperative Spatial Processing
To boost the SE and EE performance of HetNets, the inter-tier interference should be properly handled. In the PHY layer, this interference can be mitigated by inter-tier CoMP transmission. The CoMP technique is aimed at enhancing the performance of cell-edge UEs. By the coordination between transmissions of adjacent cells, the CoMP can effectively suppress the adjacent cells’ interferences when applied in UL or DL. Specifically, all access nodes can jointly decode transmitted signals from UEs in UL via JR while allocate transmission power to different UEs in a network MIMO manner with optimal precoding matrices, i.e., JT and CS/CB (Access, 2010). It should be noted that the main difference between JT and CS/CB is that the data of each UE is transmitted by only one access node in CS/CB, while multiple cooperative nodes serve each UE simultaneously in JT. In Lee et al. (2012), it is demonstrated that JT and CS/CB can improve the performance of cell-edge UEs by 100% and 30%, respectively, compared to multiuser MIMO.
However, the backhaul transmission will incur latency for coordination in practice, and meanwhile the number of quantization bits for the exchanged information is limited, which will both degrade the performance of CoMP. The impacts of non-ideal backhaul are investigated in Xia et al. (2013), and it is concluded that the backhaul latency will affect the coverage and normalized throughput. Moreover, it is shown that the coverage and throughput achieved by CoMP zero-forcing beamforming decrease almost linearly with the average latency growing from zero. Even worse, when the latency exceeds 60% of the fading coherence time, the scheme does not have any benefit. Furthermore, the large number of cooperative cells does not essentially lead to better performance. On the contrary, it is shown that the number of coordinated cells should be kept fairly small in HetNets.
Radio Resource Management
To maximize the SE and EE performances of HetNets, the multidimensional radio resources should be optimized, including power allocation, subchannel allocation, and user association. However, because of the involvement of integer variables like subchannel allocation indicator and the existence of inter-tier and intra-tier interference, multidimensional radio resource allocation in HetNets is always non-convex, making the optimization problem hard to solve. To deal with the non-convexity, there are generally two kinds of ways. One is using heuristic algorithms which are suitable for finding the near-optimal solutions to NP-hard problems. The other one is to transform these non-convex problems into convex problems by different methods, and then the primal problems can be solved by settling the corresponding convex problems. In Peng et al. (2015b), the author focuses on maximizing EE by jointly allocating the resource block and transmit power for a heterogeneous cloud radio access network. Due to the non-convexity of the primal problem, the author tackles the primal problem by solving its dual problem aiming to minimize the Lagrange function of the primal problem which is always convex by definition. Moreover, since the primal problem satisfies the time-sharing condition, zero duality gap holds. Thus, the non-convex primal problem can be equivalently transformed into a convex optimization problem, i.e., its dual problem.
Furthermore, to achieve better global performance, it is very important to consider both the conventional physical layer performance metrics and the traffic delay in the upper layer. As a simple framework to deal with delay-aware RRM optimization problems, the Lyapunov optimization approach has been commonly adopted in HetNets. In Urgaonkar and Neely (2012), the Lyapunov optimization approach is utilized to solve the problem of opportunistic cooperation in a two-tier HetNet. The access nodes would like to maximize their own throughput under average power constraints by optimizing access admission, cooperation decision, and power control. The obtained online control algorithm can stabilize the traffic queue without requiring any knowledge of the traffic arrival rates. While in Chen and Lau (2013), a two-stage queue-aware cross-layer resource management algorithm is proposed by minimizing drift-plus-utility. The cross-layer RRM consists of a queue-aware limited CSI feedback filtering optimization stage over a long timescale and an SINR-based optimal user scheduling stage over a short timescale. The utility is defined as the average feedback cost, and a large parameter V reduces the average CSI feedback at the cost of a larger average queue length.
Due to the ever-increasing number of radio resource management functionalities, it is impractical to manually adjust resource management parameters in a multidimensional parameter scenario. Faced with this issue, SON techniques have attracted a lot of attentions from both industry and academia. Generally speaking, the self-organizing capability of HetNets can be divided into three components: self-configuration, self-optimization, and self-healing. By using SON, the network configuration, installation, coverage, capacity, spectrum, and quality can be automatically managed and optimized taking into account target QoS performance, interference level, signal strength, traffic pattern, and so on.
To take full advantages of both centralized and distributed SON architectures, the author in Peng et al. (2013) proposes a hybrid SON architecture, and differences between traditional homogeneous networks and HetNets are discussed in terms of self-configuration like prediction-based radio resource configuration as well as self-optimization including mobility robustness optimization and energy saving. Moreover, it should be noted that time-varying traffic load makes the deployment of completely self-organized HetNets nontrivial. Hence, the features of SON coordination in HetNets should be emphasized. Specifically, the graph-based decision framework proposed in Gelabert et al. (2014) can be adopted to enable efficient interaction and coordination of SON mechanisms, where the interaction and conflict relationship between multiple SON mechanisms are described by the metric event and action graph. Based on this proposed framework, the strategy-based SON coordination for HetNets can be easily implemented for a given event or a combination of events.
HetNets are the key to satisfying the diverse and stringent performance requirements in future wireless networks.
Access Slicing in HetNets
As a cost-efficient solution to support diverse use cases in the 5G era, network slicing enables the provision of networks in an as-a-service fashion. However, current studies mainly concentrate on CN-based slicing, and the impact of the characteristics of RANs is not well considered. To get a more effective network slicing solution, the author in Xiang et al. (2017) proposes an enhanced network slicing approach in F-RANs, termed as access slicing. The proposed access slicing architecture is featured with a centralized orchestration layer, a slice instance layer, and the information awareness capability to guarantee diverse QoS and QoE requirements. Although the above work makes a big step for network slicing, several challenges exist for its implementation. First, as stated in Xiang et al. (2017), the information awareness allows the access slice orchestration layer makes intelligent decisions on slice instance creation and resource management. Nevertheless, more specifications should be investigated like the types of information needed, the frequency for collecting information, and the way of slice instance configuration based on processed information. Second, resource allocation between multiple slices should be investigated to fully utilize the resources of HetNets to meet diverse performance requirements. Such a problem can be very challenging, since it is basically a multi-objective optimization problem. Meanwhile, besides allocating radio resource, some slices also need caching resource and computation resource, which incurs an extra challenge.
HetNets Driven by Deep Learning
As an important technology for enabling artificial intelligence, deep learning has been successfully applied in many areas, including computer vision and speech recognition, and has drawn a lot of attentions of researchers in the wireless communication area. In Cao et al. (2017), a learning framework is proposed in which a CNN and an RNN are used to extract features in spatial domain and time domain from raw information collected by wireless networks, respectively, and this manner avoids identifying features manually. Taking the derived features as state input, DRL is adopted to control wireless networks intelligently, and the superiority of the proposal is verified by applying it to mobility management in WLAN. However, to apply deep learning in HetNets, more studies should be conducted. For example, more theoretical analysis should be done to provide guidelines on the architecture design of neural networks and the selection of hyper-parameters. Furthermore, the infrastructure of HetNets needs to be upgraded to support the implementation of deep learning algorithms, especially at the network edge to facilitate real-time network optimization and control.
- Access EUTR (2010) Further advancements for E-UTRA physical layer aspects. 3GPP technical specification TR 36:V2Google Scholar
- Bonomi F, Milito R, Zhu J, Addepalli S (2012) Fog computing and its role in the internet of things. In: Proceedings of the first edition of the MCC workshop on mobile cloud computing. ACM, Helsinki, Finland pp 13–16Google Scholar