LIEN_LAYOUT.qxp_Layout 2/25/14 3:37 PM Page 70

ACCEPTED FROM OPEN CALL

COGNITIVE RESOURCE MANAGEMENT FOR FUTURE CELLULAR NETWORKS

SHAO-YU LIEN, NAT I O N A L FORMOSA UNIVERSITY KWA N G -CHENG CHEN, NAT I O N A L TAIWAN UNIVERSITY YING-CHANG LIANG, INSTITUTE FOR INFOCOMM RESEARCH (I2R) AND UNIVERSITY OF ELEC- TRONIC SCIENCE AND TECHNOLOGY OF CHINA YONGHUA LIN, IBM RESEARCH DIVISION

ABSTRACT ety (METIS) of 100 Gbps/km2 with high mobili- ty. To support high mobility, the evolving packet The heterogeneous network (HetNet) archi- core of LTE-A is modified to simplify the han- tecture, device-to-device (D2D) communications, dover procedure. The orthogonal frequency divi- and coexistence with existing systems sion multiple access (OFDMA) is also adopted have been regarded as new communication to combat frequency selective channels. Howev- paradigms introduced in LTE-A/LTE-B cellular er, achieving a high network efficiency is never networks. To facilitate these paradigms, consider- an easy task, which has been shown infeasible by able research has shown promise of the cognitive link level solutions only. It thus drives the devel- radio (CR) technology, particularly the cognitive opment of a revolutionary network architecture radio resource management (CRRM) on the top known as the heterogeneous network (HetNet) of resource allocation to control Layer-1 and [1]. In the HetNet, in addition to conventional Layer-2 radio operations, thus eliminating the Macrocells formed by evolving NodeBs (eNBs), concerns of potential system impacts and opera- there are heterogeneous small cells including tion unreliability to bridge the gap between cellu- picocells formed by eNBs with smaller transmis- lar and CR technologies. To support diverse sion power (than that of Macrocell eNB), femto- communication paradigms with different chal- cells formed by home eNB (HeNBs) and relay lenges, a variety of CRRM schemes have been nodes (RNs), underlay the coverage of Macro- recently proposed, which however significantly cells to shorten the distance between a transmit- perplexes the system implementation. To provide ter and a receiver, as shown in Fig. 1. However, a general reconfigurable framework, in this arti- even though such a revolutionary architecture is cle, we reveal a software-defined design of the adopted, two challenges remain for the develop- CRRM. Through proper configurations, this soft- ment of future evolutions. ware-defined design is able to adapt to diverse Inter-network interference: Cellular networks communication paradigms in LTE-A/LTE-B, and require dedicated wide bandwidth to support provides transmission reliability in terms of quali- high data rates for advanced releases. However, ty-of-service guarantees via the optimum control the current is very crowded, and of the design. Supporting diverse CRRM only leaves very limited space for future evolu- schemes, this design substantially simplifies the tions, which results in a compact arrangement of system realization to bring the development of frequency bands between /5G releases and the CRRM to the next stage of practice for the other wireless systems. For instance, the fifth generation (5G) . 2400–2483.5 MHz ISM band is utilized by both WiFi and , while the operating band of LTE-A Band-40 ranges from 2300 to 2400 MHz INTRODUCTION as an immediate neighbor. Due to the imperfect Different from previous 3GPP releases of Rel. transceiver components, the compact arrange- 99 for WCDMA, Rel. 5 for HSDPA, Rel. 6 for ment of adjacent frequency bands introduces HSUPA, Rel. 7 for HSPA+, Rel. 8 and Rel. 9 severe inter-network interference, not only for LTE, the recent releases of Rel. 10 and Rel. among wireless stations but also within a device 11 for LTE-A (the fourth generation, 4G) and with multiple (i.e., the well known in- Rel. 12 for LTE-B are evolving to the fifth gen- device coexistence, IDC, interference), which eration (5G) cellular network, and thus need to immensely perplexes system and device designs. satisfy very challenging requirements envisioned From investigations by 3GPP [2], using state-of- by the mobile and wireless communications the-art RF filters does not provide sufficient enablers for the twenty-twenty information soci- rejection to adjacent channel interference.

70 1536-1284/14/$25.00 © 2014 IEEE IEEE Wireless Communications • February 2014 LIEN_LAYOUT.qxp_Layout 2/25/14 3:37 PM Page 71

By utilizing compressed sensing, the eNB of a Game theory can be applied to picocell can acquire RBs occupation in a subframe mitigate interference among small on all grids (locations) with limited reports from UE cells

eNB eNB HeNB eNB HeNB HeNB eNB HeNB HeNBeNB HeNB eNB eNB HeNB eNB HeNB HeNB HeNB HeNB HeNB eNB eNB HeNB eNB eNB eNB eNB eNB eNB eNB eNB eNB

Scenario 1: Mitigating interference between a Scenario 2: Mitigating interference among small cell and Macrocells (and WiFi) small cells

Coverage of a group of WiFi networks Macrocell Coverage of a group of WiFi networks

Macrocell Macrocell

eNB RN eNB HeNB RN HeNB HeNB eNB Picocell HeNB eNB

HeNB Macro-UE HeNB eNB HeNB RN Macro-UE RN HeNB Pico-UEeNB eNB eNB Pico-UE Picocell Macrocell eNB Macrocell Picocell Macrocell

Multi-system coexistence, HetNet, and D2D communications

D2D communications D2D communications The traditional CoMP and HetNet underlay Macrocells underlay small cells multihop CoMP can be applied to mitigate interference from WiFi

eNB eNB eNB RN eNB eNB eNB RN eNB UE eNB eNB Macro-UE UE Macro-UE RN RN eNB Pico-UEeNB eNB eNB eNB Pico-UE eNB eNB

Scenario 3: Mitigating interference for D2D Scenario 4: Mitigating interference from communications WiFi by facilitations of CoMP

Figure 1. Multi-system coexistence, HetNet, and D2D Communications in state-of-the-art E-UTRA of 4G/5G, and a general illustration of the software-defined design.

Intra-network interference: Due to the under- small cell base station to a Macrocell user locat- lying deployment of small cells without proper ed within the coverage of the small cell. cell planning, severe interference may occur not To mitigate inter-network interference under only between a small cell and a Macrocell, but the limitation without the knowledge on the also among small cells under highly dense presence of other wireless systems, Band-40 may deployments. There are different classes of be forbidden to waste precious 100 MHz band- interference in the HetNet, among those, the width. To mitigate intra-network interference major challenge lies in the interference from a without a centralized coordinator nor fixed spec-

IEEE Wireless Communications • February 2014 71 LIEN_LAYOUT.qxp_Layout 2/25/14 3:37 PM Page 72

Most importantly, it is trum partition (due to a low spectrum efficien- three years, not only for human-to-human (H2H) cy), one proposed solution is to handover the communications in the HetNet [7–10] but also demonstrated that the victim Macrocell user to the nearby small cell. for machine-to-machine (M2M) communications QoS guarantees be However, the Macrocell user may consume back- with dynamic spectrum accesses in cyber-physical haul resources of the small cell. In addition, a systems [11, 12]. However, all these schemes suf- achieved solely by small cell can not be in the close access mode fer from a new dimension of challenge that the and shall always be in the open access mode to inherent opportunistic channel access results in a optimizing the be available for Macrocell users. As a result, severe channel availability variation to signifi- operation/parameters in although interference could be mitigated, bene- cantly harm the QoS guarantees. To combat fits and security of small cells are sacrificed. such an issue, each CRRM scheme needs corre- the software-defined Consequently, these solutions in 3GPP are away sponding optimum control mechanisms. Howev- design. Such a technical from being effective. er, such a diverse design perplexes the system To solve interference issues in 4G/5G, it is implementation in 4G/5G. merit leads to a suggested that the distributive nature for infor- It is the time to envision facilitation of the mation collections and parameter optimization CRRM for practical implementation of future straightforward empowers the cognitive radio (CR) technology cellular networks. Consequently, the purpose of implementation of the [3] as a novel design paradigm [4]. However, the this article is to organize and restructure the favorable part of the CR technology to 4G/5G is most recent developments of the CRRM into a CRRM toward the not the original idea to develop intelligent software-defined design. Inspired by the wisdom 5G cellular network. devices (with powerful RF front end, powerful of software-defined communication processors signal processing capability, and powerful com- [13, 14] where functions/hardwares in individual putation capability for sophisticated analysis, layers are reused, configured, controlled, and learning, and decision making) and networks optimized by means of software, we illustrate a formed by these intelligent devices (i.e., the cog- layerless design that functions in Layer 1 to nitive radio network, CRN), due to the following Layer 3 are controlled in the central of the critical concerns. resource management. By such a concept, all 1. Reliability issue: The major task for a cel- communication scenarios supported by LTE- lular network is to provide reliable services to A/LTE-B can be taken through proper configu- users while maintaining the network stability. rations. Most importantly, it is demonstrated Specifically, quality-of-service (QoS) should be that the QoS guarantees be achieved solely by guaranteed for timing constrained services. In optimizing the operation/parameters in the soft- addition, continuous traffic congestion, unrecov- ware-defined design. Such a technical merit erable link failure, and uncontrollable communi- leads to a straightforward implementation of the cation behavior should be avoided in a cellular CRRM toward the 5G cellular network. network. These are extremely critical challenges for a CRN with a distributive design. For cellu- lar networks, when a tradeoff between perfor- COGNITIVE RADIO RESOURCE mance and reliability is encountered, reliability MANAGEMENT AND THE SOFTWARE- should have a higher priority. This is the reason that 4G engages in sophisticated network infras- DEFINED DESIGN tructures. Due to potential interference to/from other wire- 2. Potential system impacts and complexity: less systems (e.g., WiFi) and intra-network inter- Due to the conflictions on the design philosophy ference from neighboring cells of 3GPP, each and system architecture between 4G and the resource block (RB) in a subframe shall not be conventional CRN, considerable system impacts occupied by more than one cells (nor more than are introduced as applying the original design one systems). To achieve this goal, a centralized paradigm of CR to 4G. This concern substantial- RB allocation scheme is infeasible, due to ly obstructs the development of this direction, at • An unacceptable high computational complexity least in this decade. • The lacking of an appropriate interface for Since [5], a novel design paradigm based on information exchange of RB occupations the CR technology referred as the cognitive among cells and systems radio resource management (CRRM), using As a consequence, the key idea of the CRRM radio resource as the core of cross-layer design, lies in that each cell obtains RB occupation situ- opens a practical direction harmonious to the ations of neighboring cells and wireless systems system operations of 4G. With the technical through Layer-1 measurements, then only utiliz- merit of the CR technology, the CRRM controls ing unoccupied RBs based on measurement the physical layer radio operations of communi- results. To full reuse functions in existing releas- cation environment cognition and channel access es, the CRRM is restricted to only leverage via the upper layers resource management. Com- Layer-1 procedures and capabilities approved by patible to the state-of-the-art system architecture 4G. In the following, the Layer-1 measurement of 4G, the CRRM smoothly introduces the CR capabilities supported by LTE-A are outlined. technology to 4G. In LTE-A and LTE-B, a vari- Received interference power: In 3GPP TS ety of communication scenarios will be support- 36.214, an E-UTRAN (eNBs of Macrocell and ed beyond the HetNet, such as coexisting with picocell, and HeNBs) is capable of measuring systems operating on the ISM band (e.g., WiFi interference power on each RB in a subframe networks), device-to-device (D2D) communica- basis. The original purpose of this quantity is for tions [6], and heterogeneous coordinated multi- identifying interference from an UE located at point (Het-CoMP) communications. Targeting the cell edge of the neighboring Macrocell. If an at these scenarios, in the literatures, a variety of RB is identified as occupied by other UEs (inter- CRRM schemes have been proposed in recent ference on this RB exceeds a predetermined

72 IEEE Wireless Communications • February 2014 LIEN_LAYOUT.qxp_Layout 2/25/14 3:37 PM Page 73

threshold), an eNB will avoid allocating such ference and the amount of available radio In 3GPP TS 36.214, occupied RB to its UEs. Through measuring this resources (similar to the sensing-throughput quantity, an eNB or an RN can identify RBs tradeoff in the CRN [3]). By above software- an UE is able to utilize occupied by neighboring cells (or WiFi if the defined design, QoS can be provided by strike reference signals to Band-40 is utilized). such a tradeoff under all kinds of communica- Reference signal received quality (RSRQ): In tion scenarios. However, a dynamic measure- measure the signal to 3GPP TS 36.214, an UE is able to utilize refer- ment period is not the case for Macrocells, as ence signals to measure the signal to interfer- Macrocells shall support a considerable number interference and noise ence and noise power ratio (SINR) averaged of UEs, and a dynamic measurement period may power ratio (SINR) over considered bandwidth. If the value of this increase the burden in . quantity is too low, interference is severe on averaged over considered bandwidth. considered bandwidth. Number of neighboring small cells: In 3GPP CONFIGURATIONS OF THE TS 25.967, a HeNB (and thus an eNB) is able to SOFTWARE-DEFINED DESIGN If the value of this identify cell identifications (cell IDs) of neigh- quantity is too low, boring cells. Thus, a HeNB (and thus an eNB) is CONFIGURATIONS FOR INTRA-NETWORK able to identify the number of neighboring cells. interference is severe on Since a Macrocell typically handles a larger INTERFERENCE MITIGATION number of UEs than that of a small cell, a small To support all kinds of communication scenar- considered bandwidth. cell should adapt to operations of Macrocells. ios, our software-defined design consequently Furthermore, if Band-40 is utilized (2300 to 2400 integrate a variety of CRRM schemes to miti- MHz), WiFi is one of major sources introducing gate interference under the HetNet, D2D com- considerable inter-network interference to LTE- munications, multi-system coexistence as well as A. Since WiFi is typically extensively deployed Het-CoMP. (as a group of WiFi networks) to provide seam- less services to users, they may affect the entire Scenario 1. Mitigating Interference between A Small Cell Macrocells, small cells, and D2D links. As a con- and Macrocells (and WiFi Networks) — As shown in sequence, UE, small cells, and Macrocells should Fig. 1, all small cells underlay coverage areas of adapt to operations of WiFi networks, due to the Macrocells (and WiFi networks) may invoke/suf- fact that WiFi is a widely deployed legacy sys- fer interference to/from Macrocells (and WiFi tem. Aligning with these operating principles, a networks). For such interference mitigation, in variety of recent CRRM schemes for different Step 2 of above software-defined design, an scenarios [7–12] are actually on a common func- HeNB in a femtocell can identify occupied RBs tional structure. Consequently, a software- in a measurement subframe by measuring the defined design can be constructed with the received interference power on all RBs in a following operations. measurement subframe. If this quantity exceeds 1.Subframe boundaries of small cells align with a pre-determined threshold on an RB, this RB is those of Macrocells. A similar agreement is identified as occupied by Macrocells (or WiFi reached in TR 36.921. networks). By only utilizing unoccupied RBs, 2.For all E-UTRAN (eNBs, HeNBs or RNs), interference can be consequently mitigated. there are logically two kinds of subframes: However, this operation is not generally suffi- measurement subframe and data subframe. In cient for picocells. Since the coverage of a pico- each measurement subframe, each E-UTRAN cell is typically larger than that of a femtocell, identifies those RBs are occupied. Then, in only the measurement by the picocell eNB may subsequent data subframes, the E-UTRAN not completely reflect the interference situation only allocates unoccupied RBs (measured in of all UEs distributed over the coverage of the the latest measurement subframe) to its UEs. picocell. Thus, all UEs shall measure the corre- As a consequence, an E-UTRAN can not per- sponding location-based RSRQ in a measure- form data transmission and reception in a ment subframe and report the measurement measurement subframe. The measurement results to the picocell eNB. However, if there is period (in the unit of the number of sub- a large number of UEs, the uplink channel for frames) is dynamic for small cells, while it is measurement report may suffer from severe con- fixed for Macrocells and RNs. gestions. To overcome this huge challenge, the 3.By identifying occupied RBs in each measure- compressed sensing technology can be applied ment subframe, an E-UTRAN can further [15]. Compressed sensing origins as a signal pro- infer the traffic load and RB allocation corre- cessing technology, which is able to sample an lation in the neighboring cells (or WiFi net- audio/image signal with a sampling rate far works) for the subsequent optimum control. lower than the Nyquist rate. The signal can be From Step 2, it is known that a measurement recovered with an acceptable error rate if certain subframe is an overhead for the communication constraints can be satisfied. Such technical merit environment cognition. If the measurement peri- makes compressed sensing a powerful technolo- od is very short, most subframes are leveraged gy for picocells. As shown in Fig. 1, the coverage for measurement, and thus the available radio area of a picocell is divided into N isomorphic resources are significantly reduced. On the other grids indexed by n = 1, …, N. By applying com- hand, if the measurement period is very long, pressed sensing, within each measurement sub- although the overhead is significantly reduced, frame, each UE performs channel measurement dynamics of interference can not be accurately with a probability q  1. As a result, if there are captured. As a result, the available radio V UEs in a picocell, only qV  V UEs are respon- resource without interference may not increase. sible for performing the measurement and Therefore, there exists a tradeoff between inter- report. Upon receiving feedback from UEs,

IEEE Wireless Communications • February 2014 73 LIEN_LAYOUT.qxp_Layout 2/25/14 3:37 PM Page 74

ty of the true RBs occupations in Macrocells (and 0.17 WiFi networks) and q. The sparsity of the true ρ=0.8 ρ=0.7 RBs occupations in Macrocells (and WiFi net- 0.16 ρ=0.6 ρ works) is uncontrollable for a picocell. However, =0.5 through the evaluation on the error rate of the 0.15 spectrum map construction in Fig. 2, it shows that there is only a 7.5 percent error (7.5 RBs are 0.14 occupied by Macrocells but they are estimated as unoccupied, or 7.5 RBs are not occupied by 0.13 Macrocells but they are estimated as occupied, among 100 RBs in a subframe) by properly setting 0.12 q  [0.6,0.8]. Please note that it suggests that only 60 percent among V UEs are adopted for channel 0.11 measurement. This result sufficiently shows practi- cability and effectiveness of the software-defined 0.1 design on supporting compressed sensing.

Spectrum map construction error rate 0.09 Scenario 2. Mitigating Interference among Small Cell — By the software-defined design, all small cells 0.08 can autonomously mitigate interference to/from Macrocells and WiFi networks. However, the subsequent challenge is interference among 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 small cells deployed closely with each other q (referred as collocated small cells), as shown in Fig. 1. For collocated small cells, an identical Figure 2. Error rate of the spectrum map construction under different q. In this “pool” of unoccupied RBs are identified. To simulation, there are 7 hexagonal-grid Macrocells with wrap-around and three avoid interference among small cells, each small picocells within the coverage of Macrocells. All Macro- and picocells share a cell shall utilize distinct RBs from each other. 20MHz bandwidth with M = 100 RBs per subframes. r = M¢/M denotes the However, in LTE-A, there is no interface pro- traffic load of Macrocells, where M¢ is the number of RBs utilized by Macro- vided for femtocells to exchange information. cells in each subframe. Other parameters and assumptions for simulation fol- For picocells, although the X2 is available for lows the definition in 3GPP TR 36.814. data exchanges via wired link, the large latency may not fully reflect the resource variation in Layer 1. As a result, lacking an appropriate compressed sensing takes place in the eNB by interface for a centralized coordination among multiplying a R × N sampling matrix A (with small cells in LTE-A, each collocated small cell each element taking “1” with probability should randomize the utilization of unoccupied RBs to minimize interference among small cells. V However, there is another factor that can signifi- q n , V cantly affect the performance. Since demands in small cells are typically different from one to and taking “0” with probability another, if a small cell decides to utilize as many unoccupied RBs as it needs, interference may V 1,− q n V become severe when all collocated small cells V n have heavy demands. On the other hand, if a small cell decides to utilize unoccupied RBs in a is the number of UEs within the nth grid) on the very conservative manner, RBs are underutilized T true RBs occupation  =[y 1y 2 … y N] of if all collocated small cells have light demands. Macrocells (and WiFi networks), where yn indi- Since the traffic demand in a small cell is cates RB occupations among M RBs in a mea- unavailable to other small cells, an effective surement subframe on the nth grid. That is, solution to determine the optimum number of unoccupied RBs for utilization lies in game theo- y = A + e. (1) ry [16]. In the following, for interference mitiga- tion among femtocells or among picocells, the We particularly call the RB occupations at all following enhancement can support the soft- the grids as a “spectrum map.” The spectrum ware-defined design for collocated small cells. map can be constructed by searching the mini- 1. By measuring the number of neighboring small mum l1 norm of  cells (i.e., the number of players in the game) with the Layer-1 capability defined in TS A * = arg min 1 s.t.   – y 2  e,(2) 25.967, the number of unoccupied RBs in the measurement subframe and the number of if total RBs in a subframe, the utility function can be formulated as the number of available ⎛ NM ⎞ ROK= log , RBs without interference. ⎝⎜ K ⎠⎟ 2.The optimum number of unoccupied RBs for utilization is determined by the equilibrium A where K is the sparsity of , e = |e|2 and  is solution. a random basis. 3.After determining the optimum number of The accuracy on the spectrum map construc- unoccupied RBs for utilization, these RBs are tion by compressed sensing depends on the sparsi- utilized by a small cell in a randomized manner.

74 IEEE Wireless Communications • February 2014 LIEN_LAYOUT.qxp_Layout 2/25/14 3:37 PM Page 75

Scenario 3. Mitigating Interference for a D2D Link — such transmission diversity, the timing constraint In future releases of D2D is a new type of communications that will of a service is violated only if none of the links be supported by 3GPP Rel. 12 to reduce bur- can successfully deliver the packet before expira- 3GPP, eNBs are no dens of eNBs by localized communications. tion of timing constraints, thus the QoS is longer the only There can be two possible cases for D2D com- enhanced. Although, by CoMP transmissions, munications: the eNB allocating RBs for D2D additional radio resources in the spatial domain in CoMP links, and UEs autonomously identifying unoc- are required, the communication suspension due cupied RBs for D2D communications. The first to interference from WiFi can be alleviated. In transmissions. case is suitable for D2D links underlay small future releases of 3GPP, eNBs are no longer the In addition, RNs are cells, as shown in Fig. 1. Since a small cell typi- only transmitter in CoMP transmissions. In addi- cally handles a smaller number of UEs, schedul- tion, RNs are involved as a major part of CoMP involved as a major part ing for D2D communications does not introduce transmissions. Thus, the traditional CoMP of CoMP transmissions. significant computational burdens. Furthermore, evolves to the Het-CoMP, as shown in Fig. 1. To a small cell typically has a smaller coverage, and support this new feature, the software-defined Thus, the traditional thus the impact to other UEs from a small cell design can be configured by the following way. can be precisely characterized. The second case 1.All eNBs and RNs involved in CoMP trans- CoMP evolves to the is suitable for D2D links underlay Macrocells missions identify interference from WiFi Het-CoMP. due to similar reasons. For D2D links underlay based on the operation of the software-defined small cells, an HeNB or an eNB of a picocell can design. The measurement period of these identify unoccupied RBs from Macrocells (and eNBs and RNs are set to a fixed value. WiFi networks) and neighboring small cells 2.When a packet with timing constraint is through the configurations for the Scenario 1 required to be transmitted to an UE, eNBs and Scenario 2. For D2D links underlay Macro- involved in CoMP simultaneously transmit cells, an UE can also utilize the configuration identical packets (the duplication of the pack- for the scenario 1 to identify unoccupied RBs et) to the UE (or to an RN if there is one) via from Macrocells (and/or WiFi groups) and different links. neighboring D2D links, respectively. Under this –RNs transmit the packet to the UE or the case, the measurement period of an D2D link is next hop RN as they receive the packet from also adjustable. an eNB or the previous hop RN. –The eNBs and RNs only utilize unoccupied CONFIGURATION FOR INTER-NETWORK RBs in each data subframe, and it may take multiple subframes to transmit a packet to the INTERFERENCE MITIGATION UE. In addition, since eNBs (and RNs) suffer Scenario 4. Mitigating Interference between Macrocells from different levels of interference from and WiFi — Similar to the above three scenarios, WiFi, it also takes distinct numbers of sub- to mitigate interference to/from WiFi networks, frames to deliver a packet to an UE from dif- Macrocells also adopt measurement subframes ferent links (with or without RNs). Since all for interference estimations. However, different these eNB deliver an identical packet, the tim- from the above three scenarios for small cells ing constraint is violated only if none of the and D2D communications, Macrocells face a links can deliver the packet by the timing con- huge challenge of handling a large number of straint expiration. UEs. As a result, the measurement period of 3.For packets without timing constraints is Macrocells and RNs may have to be fixed, as a required to be transmitted, eNBs are involved dynamic measurement period may perplex the in CoMP transmissions distinct packets to the system operation for stability to support and UE (or to an RN if there is one) via different coordinate a large number of UEs. In addition, links to enhance the throughput. communications in a small cell enjoy a stronger –RNs transmit the packet to the UE or the signal strength than that in a Macrocell. There- next hop RN as they receive the packet from fore, data rates in a small cell typically are invul- an eNB or from the previous hop RN. nerable to interference from WiFi groups. –The eNBs and RNs only utilize unoccupied However, data rates of longer distance commu- RBs in each data subframe. nications in a Macrocell are typically vulnerable to WiFi groups. As a result, most RBs in data subframes may be unavailable for Macrocells THE OPTIMUM CONTROL OF THE due to severe interference under concurrent SOFTWARE-DEFINED DESIGN transmissions of WiFi networks and a Macrocell. For IEEE 802.11, the network allocation vector THE OPTIMUM CONTROL FOR INTRA-NETWORK (NAV) for each station can be set to the maxi- mum of around 33 ms for continuous packet INTERFERENCE MITIGATION transmissions. Under this circumstance, commu- To provide QoS guarantees while maximizing nications in a Macrocells may be suspended for resource utilization for interference mitigation 33 subframes, which may disable a Macrocell intra LTE-A/B, the measurement period as well from providing QoS guarantees for services with as the allocation of unoccupied RBs shall be timing constraints. For the traditional link optimized. For this purpose, a novel control the- between an eNB and an UE, such interference is ory joint queueing theory and information theo- never alleviated. To combat this critical issue, a ry, i.e., the effective bandwidth theory and the powerful multi-antenna technology of CoMP effective capacity theory, can be exploited. The transmissions [17] can be applied. For traditional effective bandwidth theory specifies the minimum CoMP transmissions, multiple eNBs transmit an service rate to serve a given arrival process sub- identical packet to an UE via different links. By ject to a given QoS requirement. On the other

IEEE Wireless Communications • February 2014 75 LIEN_LAYOUT.qxp_Layout 2/25/14 3:37 PM Page 76

Since the measurement Calculate the effective bandwidth for period is fixed for the real-time service Characterize packet transmission delays under CoMP transmissions of the Macrocells and RNs, to interference scenario 4 for all admitted provide QoS guarantees Set the initial measurement period to a services when a new (VoIP/Video) call predetermined value arrives. while maximizing resources utilization, the Allocate l=1 unoccupied RB to the UE for utilization Allocate / unoccupied RBs on each key is to determine the CoMP path within t subframes where l can be fixed, while t is determined by t t t optimal packet transmis- min{ }, 0< < max, such that timing Calculate the effective capacity of the constraints of all admitted services and system according to the measured RBs the new service are not violated. sion time for different occupation situation services in Het-CoMP transmissions, as identi- Estimate the QoS performance on current operation point (l and No measurement period) on the equilibrium cal packet transmissions If the optimum t can be obtained that the effective capacity equals to the from multiple eNB may effective bandwidth waste spatial domain radio resources. Yes Yes Update t for all services and admit the The QoS performance satisfies new service the constraint

No The new service cannot be served

Decrease the measurement period by 1

Determine l by min{l}, s.t. 0

Done

Figure 3. The optimal control procedures for the software-defined design. For a HeNB, an eNB of a picocell, or an UE performing D2D communications, the optimum control calculates the effective capacity denoted l,T (q) by EC s under different values of allocated unoccupied RB l and measurement period Ts, and calcu- lates the effective bandwidth of voice/video traffic denoted EB(q), where q is a parameter known as the l,T QoS exponent. l and Ts are optimized by achieving EC s(q) = EB(q) = d such that the delay violation −qdd probability Pr{Delay > dmax} = e max satisfies the requirement. For a Macrocell eNB, the optimum control calculates packet transmission delays under different number of paths supported by Het-CoMP transmissions. If the optimum solution on the number of allocated RB and the number of Het-CoMP paths can be obtained such that QoS requirements are satisfied, the service can be admitted.

hand, the effective capacity theory specifies the tive bandwidth equals to the effective capacity, maximum constant arrival rate that can be sup- and the corresponding QoS performance satisfies ported by the system subject to a given QoS the requirement, as detailed in Fig. 3. requirement. When the effective bandwidth equals to the effective bandwidth, an equilibrium THE OPTIMUM CONTROL FOR of a statistical delay guarantee can be reached (that is, the probability that packet delivery delay INTER-NETWORK INTERFERENCE MITIGATION exceeds the required value is upper bounded by a Since the measurement period is fixed for Macro- certain value). To apply this in cellular networks, cells and RNs (and thus the overhead is fixed), to a HeNB, an eNB of a picocell, or an UE per- provide QoS guarantees while maximizing forming D2D communications calculates the resources utilization, the key is to determine the effective bandwidth of timing-constrained services optimal packet transmission time for different [18], and calculates the effective capacity of the services in Het-CoMP transmissions, as identical configurations for interference Scenario 1, Sce- packet transmissions from multiple eNB may nario 2 or Scenario 3 according to the traffic load waste spatial domain radio resources. To achieve and RB allocation correlation of neighboring cells this goal, packet transmission delays under Het- under different measurement periods and num- CoMP transmissions shall be characterized by bers of utilized RBs. The measurement period eNBs for different (voice/video) services based and the number of utilized RBs can consequently on different activities of WiFi networks. After be optimized under the condition that the effec- analytically characterizing packet transmission

76 IEEE Wireless Communications • February 2014 LIEN_LAYOUT.qxp_Layout 2/25/14 3:37 PM Page 77

Violation prob. Violation prob. of Delay-bound Violation prob. of Violation prob. of Delay- software-defined software-defined Traffic violation Randomization Randomization bound design (low design (high prob. (low correlation) (high correlation) correlation)a correlation)b

Star Wars 40ms  0.02 0.0199 0.0066 0.0464 0.0331

Die hard 40ms  0.02 0.0172 0.0058 0.0398 0.0313

Jurassic Park 40ms  0.02 0.0181 0.0055 0.0415 0.0299

VoIP 20ms  0.02 0.0012 0.0003 0.0013 0.0008

a There is a low correlation of RBs allocation in Macrocells (and WiFi networks) among subframes (that is, if a particular RB is occupied in a subframe, this RB is occupied by Macrocells (and WiFi networks) in subsequent subframe with a low probability of 0.3).

b There is a high correlation of RBs allocation in Macrocells (and WiFi networks) among subframes (that is, if a particular RB is occupied by Macrocells (and WiFi networks) in a subframe, this RB is occupied in subsequent subframe with a high probability of 0.8).

Table 1. QoS requirements and simulation results on the delay bound violation probability for the VoIP and high quality MPEG4 video transmissions in a small cell for intra-network interference mitigation.

delays for different services, eNBs can make the stream is considered. The arrival process of the optimum RB allocation as well as admission con- VoIP is the well-known ON-OFF fluid model. trol such that timing constraints of all admitted The holding times in “ON” and “OFF” states services can be satisfied. Such the optimum con- are exponentially distributed with the means 6.1s trol for Macrocells is also illustrated in Fig. 3. and 8.5s, respectively. The data rate of the “ON” state is 32Kb/s. The delay bound is 20ms and the PERFORMANCE EVALUATIONS delay bound violation probability is 0.02. For video traffic, a high quality MPEG4 movie trace To show technical merits of the software-defined is considered. The delay bound of the video traf- design of the CRRM, reliability in terms of QoS fic is 40ms and the delay bound violation proba- guarantees is evaluated under different interfer- bility is 0.02. The results in Table I show the ence scenarios. These simulations are conducted effectiveness on the support of the statistical based on system parameters and assumptions delay guarantees by the software-defined design, defined by 3GPP TR 36.814 for LTE-A with while the video traffic can not be supported by 20MHz bandwidth. Each RB is composed of 12 the randomized scheme. subcarriers over 7 OFDM symbols. There are 7 hexagonal-grid Macrocells with wrap-around, MITIGATING INTERFERENCE FROM WIFI NETWORKS and 5 clusters of small cells deployed over the For inter-network interference mitigation, Table coverage areas of Macrocells. Each cluster is 2 shows simulation results of the timing con- composed of 1 to 5 small cells. Transmission straint violation probabilities of 5 VoIP and 5 power of a Macrocell, a picocell, a femtocell, an video streams. The results are demonstrated in RN, and a WiFi network is set to 46 dBm, 30 the form of (timing constraint violation probabil- dBm, 20 dBm, 30 dBm, and 3dBm, respectively. ity, average activity of WiFi networks on each By the software-defined design, a station in Het-CoMP link) under different number of Het- LTE-A/LTE-B can autonomously acquire RB CoMP links that shall be involved (denoted by occupation situations of neighboring cells to pro- S). We can observe from Table 2 that, for S = 1, ceed to an optimized action according to QoS of all VoIP and video can be guaranteed acquired RB occupation situations. Without when the average activity of WiFi networks on autonomous acquisition of RB occupation situa- each Het-CoMP link does not exceed 0.8. For S tions, a station has no information about RBs = 3 and S = 5, QoS can be guaranteed when occupation in a subframe (and the potential the average activity of WiFi networks on each trend of such RB occupation) by neighboring Het-CoMP link does not exceed 0.5 and 0.4, cells. Under this constraint, the optimal scheme respectively. for interference mitigation lies in the randomized Although Table 2 shows that QoS of all VoIP scheme. That is, each eNB, HeNB or UE in and video can be guaranteed, it does not reveal D2D communications randomizes the RB uti- the efficiency on the utilization of the spatial lization in a subframe to avoid interference on domain radio resources in the Het-CoMP, as successive RBs. In the following simulations, the transmitting identical packets via multiple eNBs performance on the support of real-time voice potentially wastes radio resources. In Fig. 4, the and video services is particularly evaluated. numbers of required Het-CoMP (or CoMP) paths of the software-defined design and the MITIGATING INTRA-NETWORK INTERFERENCE state-of-the-art CR transmission scheme. We can Table 1 shows the simulation results of a small observe from Fig. 4 that, to provide QoS guaran- cell on the support of real-time voice and video tees for 5 VoIP and 5 video streams, the existing transmissions. For the voice traffic, a VoIP CR scheme requires a larger number of Het-

IEEE Wireless Communications • February 2014 77 LIEN_LAYOUT.qxp_Layout 2/25/14 3:37 PM Page 78

Stream VoIP1 VoIP2 VoIP3 VoIP4 VoIP5 Video1 Video2 Video3 Video4 Video5

(0.001, (0.001, (0.001, (0.002, (0.002, (0, 0.9), (0, 0.9), 0.9), 0.9), 0.9), 0.9), 0.9), (0, 0.9), (0, 0.9), (0, 0.9), S = 1 (0.006, (0.007, (0.008, (0.008, (0.008, (0.008, (0.008, (0, 0.8) (0, 0.8) (0, 0.8) 0.8) 0.8) 0.8) 0.8) 0.8) 0.8) 0.8)

(0, 0.9), (0, 0.9), (0, 0.9), (0, 0.9), (0, 0.9), (0, 0.9), (0, 0.9), (0, 0.9), (0, 0.9), (0, 0.9), (0, 0.8), (0, 0.8), (0, 0.8), (0, 0.8), (0, 0.8), (0, 0.8), (0, 0.8), (0, 0.8), (0, 0.8), (0, 0.8), (0, 0.7), (0, 0.7), (0, 0.7), (0, 0.7), (0, 0.7), (0, 0.7), (0, 0.7), (0, 0.7), (0, 0.7), (0, 0.7), (0.002, S = 3 (0.002, (0.002, (0.002, (0.002, (0.002, (0.001, (0, 0.6), (0, 0.6), (0, 0.6), 0.6), 0.6), 0.6), 0.6), 0.6), 0.6), 0.6), (0.007, (0.007, (0.007, (0.005, (0.007, (0.007, (0.007, (0.007, (0.007, (0.005, 0.5) 0.5) 0.5) 0.006, 0.5) 0.5) 0.5) 0.5) 0.5) 0.5) 0.5)

(0, 0.9), (0, 0.9), (0, 0.9), (0, 0.9), (0, 0.9), (0, 0.9), (0, 0.9), (0, 0.8), (0, 0.8), (0, 0.8), (0, 0.8), (0, 0.8), (0, 0.9), (0, 0.9), (0, 0.9), (0, 0.8), (0, 0.8), (0, 0.7), (0, 0.7), (0, 0.7), (0, 0.7), (0, 0.7), (0, 0.8), (0, 0.8), (0, 0.8), (0, 0.7), (0, 0.7), (0, 0.6), (0, 0.6), (0, 0.6), (0, 0.6), (0, 0.6), (0, 0.7), (0, 0.7), (0, 0.7), (0, 0.6), (0, 0.6), S = 5 (0.001, (0.001, (0.001, (0.001, (0.001, (0, 0.6), (0, 0.6), (0, 0.6), (0.001, (0.001, 0.5), 0.5), 0.5), 0.5), 0.5), (0, 0.5), (0, 0.5), (0, 0.5), 0.5), 0.5), (0.003, (0.004, (0.004, (0.004, (0.004, (0, 0.4) (0, 0.4) (0, 0.4) (0.003, (0.003, 0.4) 0.4) 0.4) 0.4) 0.4) 0.4) 0.4)

Table 2. Simulation results of QoS provisioning for the scenario of inter-network interference mitigation.

CoMP (or CoMP) paths as compared with that ware-defined network processor architecture for of the software-defined design. These results suf- LTE family of systems. At this standpoint, we ficiently demonstrate the effectiveness of the delineate the new research opportunities toward optimum operation of the software-defined the next evolution of cellular network designs. design. ACKNOWLEDGEMENT CONCLUSION This research is sponsored by MediaTek Inc. with project number 102H32048 under the The CRRM reforms the original concept of the MediaTek-NTU Advanced Research Center. CR to a realistic layerless technology to solve challenging interference in cellular networks. REFERENCES Considering all practical scenarios, we present a [1] A. Damnjanovic et al., “A Survey on 3GPP Heteroge- software-defined design with the optimum con- neous Networks,” IEEE Wireless Commun., vol. 18, no. 3, pp. 10–21, June 2011. trol relieving individual sub-system function. [2] CMCC, “RP-100671: Signalling and Procedure for In- Such design thus enables the development of the Device Coexistence Interference Avoidance,” 3GPP TSG CRRM from a promising concept to a successful RAN Meeting 48, June 2010. practice, and leads to a novel realization of soft- [3] Y.-C. Liang et al., “Sensing-Throughput Tradeoff for Cognitive Radio Networks,” IEEE Trans. Wireless Com- mun., vol. 7, no. 4, Apr. 2008, pp. 1326–37. [4] A. Attar, V. Krishnamurthy, and O. N. Gharehshiran, 40 “Interference Management Using Cognitive Base-Sta- tions for UMTS LTE,” IEEE Commun. Mag., vol. 49, no. 8, Aug. 2011, pp. 152–59. 35 [5] S.-Y. Lien et al., “Cognitive Radio Resource Manage- ment for QoS Guarantees in Autonomous Femtocell Networks,” Proc. IEEE ICC, 2010. 30 [6] 3GPP RAN WS, “RWS-120003: LTE Release 12 and Beyond,” 3GPP RAN WS on Rel-12 and Onwards, June 2012. 25 [7] S.-Y. Lien, Y.-Y. Lin, and K.-C. Chen, “Cognitive and Game-Theoretical Radio Resource Management for Autonomous Femtocells with QoS Guarantees,” IEEE 20 Trans. Wireless Commun., vol. 10, no. 7, July 2011, pp. 2196–206. Existing CR transmissions [8] S.-Y. Lien and K.-C. Chen, “Statistical Traffic Control for 15 Cognitive Radio Empowered LTE-Advanced with Net- work MIMO,” Proc. IEEE INFOCOM Wksps., 2011.

provide guaranteed QoS Software-defined design with [9] J. W. Huang and V. Krishnamurthy, “Cognitive Base Sta- the optimized operation 10 tions in LTE/3GPP femtocells: A Correlated Equilibrium Game-Theoretic Approach,” IEEE Trans. Commun., vol. 59, no. 12, Dec. 2011, pp. 3485–93. The minimum number of paths required to 5 [10] Y.-Y. Li and E. S. Sousa, “Cognitive Uplik Interference Management in 4G Cellular Femtocells,” Proc. IEEE PIMRC, 2010. 0 [11] S.-Y. Lien et al., “Radio Resource Management for QoS 01 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Guarantees in Cyber-Physical Systems,” IEEE Trans. Par- The average activity of WiFi networks on eah Het-CoMP link allel Distrib. Sys., vol. 23, no. 9, Sept. 2012, pp. 1752–61. [12] Q. Wang et al., “Interference Management for Smart Figure 4. The numbers of required Het-CoMP paths to provide QoS guarantees Grid Communication Under Cognitive Wireless Net- for 5 VoIP and 5 Video streams. work,” Proc. IEEE SmartGridComm, 2012.

78 IEEE Wireless Communications • February 2014 LIEN_LAYOUT.qxp_Layout 2/25/14 3:37 PM Page 79

[13] C.-K. Liang and K.-C. Chen, “A Green Software-Defined of Electrical Engineering, Stanford University, from Dec Communication Processor for Dynamic Spectrum 2002 to Dec 2003, and taught graduate courses in Access,” Proc. IEEE PIMRC, 2010. National University of Singapore from 2004–2009. His [14] L. B. Thiagarajan, S. Attallah, and Y.-C. Liang, “Recon- research interest includes cognitive communications and figurable Transceivers for Wireless Broadband Access networking, broadband communications, information Schemes,” IEEE Wireless Commun. Mag., vol. 14, no. 3, theory and statistical signal processing. He currently June 2007, pp. 48–53. serves as Editor-in-Chief of IEEE Journal on Selected Areas [15] D. Donoho, “Compressed Sensing,” IEEE Trans. Info. in Communications–Cognitive Radio Series, and is on the Theory, vol. 52, no. 4, Apr. 2006, pp. 1289–306. editorial board of IEEE Signal Processing Magazine. He [16] D. Fudenberg and J. Tirole, Game Theory, The MIT was an Associate Editor of IEEE Transactions on Wireless Press, 1991. Communications and IEEE Transactions on Vehicular Tech- [17] D. Lee et al., “Coordinated Multipoint Transmission nology. He is a Distinguished Lecturer of the IEEE Com- and Reception in LTE-Advanced: Deployment Scenarios munications Society and the IEEE Vehicular Technology and Operational Challenges,” IEEE Commun. Mag., vol. Society. 50, no. 2, Feb. 2012, pp. 148–55. [18] C.-S. Chang, Performance Guarantees in Communica- YONGHUA LIN [M] ([email protected]) is the Senior Techni- tion Networks, Springer, 2000. cal Staff Member (STSM) of IBM and Senior Research Man- ager leading the Networking and Wireless research group BIOGRAPHIES in IBM Research - China. She leads two global research SHAO-YU LIEN ([email protected]) is an assistant professor directions across China, US and India labs. One is next gen- of the Department of Electronic Engineering, National For- eration radio access network (RAN) architecture (wireless mosa University, Taiwan. He received a number of presti- network cloud), and the other one is wireless system and gious recognitions, including IEEE ICC 2010 Best Paper optimization for IoT (Internet of Things). Yonghua Lin is Award and URSI AP-RASC 2013 Young Scientist Award. His the major driver of IT-based software defined radio (SDR) research interests lie in optimization techniques for net- technology, and the first one proposing the Wireless Net- works and communication systems. Recently, focuses are work Cloud concept in the world. Under her leadership in particularly on cyber-physical systems, 4G/5G communica- wireless research, IBM Research team has successfully driv- tion networks, and network sciences. en the new trend of Wireless Network Cloud for next gen- eration RAN in industry, established the wide collaboration KWANG-CHENG CHEN [F] ([email protected]) is the in industry including mobile operators, telecom equipment Distinguished Professor and Associate Dean for Academic providers and IoT enterprises, and created lots of FIRST Affairs, College of Electrical Engineering and Computer Sci- prototypes in industry to demonstrate the IBM’s leading ence, National Taiwan University. He received 2011 IEEE technologies in wireless system (ICASSP08, GLOBECOM09, ComSoc WTC Recognition Award. His research interests IMIC10 etc.). With her leadership in both China and global include wireless communications, wireless networks, and IBM wireless research activities, IT and wireless conver- network science. gence has become the main topic in IBM Global Technolo- gy Outlook 2010. Yonghua Lin is also active in academic YING-CHANG LIANG [F] ([email protected]) is a Principal Sci- communities. She is the member of ACM. She has more entist in the Institute for Infocomm Research, A*STAR, than 30 patents granted worldwide and publications in top Singapore. He was a visiting scholar with the Department conferences and journals.

IEEE Wireless Communications • February 2014 79