Interference-Aware Resource Allocation for Device-to-Device Communication in 5G H-CRAN Networks

Transcription

1 Interference-Aware Resource Allocation for Device-to-Device Communication in 5G H-CRAN Networks Xingwang Mao 1,2, Biling Zhang 1,2, Xuerong Gou 1 1. School of Network Education, Beijing University of Posts and Telecommunications, Beijing, , P. R. China. 2. The State Key Laboratory of Integrated Services Networks, Xidian University, Xian, , P. R. China. xingwangmao@bupt.edu.cn, bilingzhang@bupt.edu.cn, xrgou@163.com Abstract Both heterogeneous cloud radio access networks (H-CRAN) and device-to-device (D2D) communication are promising new paradigms to improve the network throughput and spectral efficiency. However, no D2D resource allocation schemes are proposed for H-CRAN whose architecture is totally different from existing LTE Advanced cellar network and where more than one D2D pairs are allowed to share the resource blocks (RBs) of one cellular user equipment (CUE). To address this problem, we study the uplink resource allocation problem in H-CRAN for D2D communications. We first analyze the interference scenario and formulate the resource allocation problem into a NPhard optimization problem. Then a throughput aware maximum association (TAMA) algorithm and a channel condition aware maximum association (CCAMA) algorithm are introduced to solve the proposed problem. Numerical results demonstrate that the proposed algorithms can improve the system performance in terms of date rate and throughput. Index Terms H-CRAN, D2D Communication, interference avoidance, resource allocation

2 I. INTRODUCTION Nowadays, the research pace of the next generation mobile networks is accelerating worldwide to meet the increasing demand for high data rate applications [1] [4]. In the 5G mobile networks, higher system spectrum efficiency and energy efficiency, greater number of terminal devices, and better user experience in terms of data rate and lower latency are expected. Therefore, researchers are seeking for new paradigms. To achieve these goals, the heterogeneous cloud radio access network (H-CRAN) which incorporates cloud computing into heterogeneous networks (HetNets), is proposed as one of the promising radio access network architecture. In the H-CRAN networks, only partial functionalities in the physical layer are executed in remote radio head (RRHs), while the other important baseband physical processing functionalities and procedures of the upper layers are incorporated in the baseband unit (BBU) pool which has powerful computing capabilities. Flexible deployed RRHs are responsible for providing high data rate, and the macro base stations are mainly used to provide coverage blanket and execute the functions of control plane. Benefiting from this novel radio access network architecture, the cloud computing based cooperation processing and networking gains are fully exploited, the operating expenses are lowered, and the energy consumption of the wireless infrastructure is decreased [5] [7]. Meanwhile, by sharing the the resource of the cellular user equipments (CUEs), device-to-device (D2D) communication is proposed as one of the promising technologies. To improve the system performance in terms of spectrum efficiency, throughput, delay, as well as energy efficiency, in recent literatures, many aspects of D2D communication have been investigated. For instance, in [8] and [9], D2D communication session setup and management mechanisms for the LTE System Architecture Evolution (SAE) were introduced. In [14] and [15], resource allocation schemes for D2D communication underlaying cellular networks were proposed. To manage the interference incurred by the resource reuse, several schemes were proposed for receiving mode selection [16], [18], power control [9], [17], [19] and fractional frequency reuse (FFR) [19], [20]. Resource allocation and interference cancelation are often jointly considered [10] [13]. In [10], the authors provided methods of allocating the reusing channels for D2D pairs so as to maximize the number of permitted D2D connection and minimize the average interference caused by D2D pairs simultaneously. In [11], Zulhasnine et al. formulated the problem of radio resource allocation to the D2D communications as a mixed integer nonlinear programming (MINLP) to maximize the system throughput. To solve the MINLP problem, the authors propose an alternative greedy heuristic algorithm utilizing channel gain information. In [12] and [13], considering network interference, the

3 resource block (RB) allocation problem for D2D pairs was modelled by conflict graph (CG). Although existing schemes are proved to be efficient in cellular networks, most of them can not be applied directly in 5G due to the following reasons. First, existing works largely assume that the number of CUEs is larger than the number of D2D pairs, and any RBs allocated to a CUE can be reused by at most one D2D pair. Such assumptions may not be true in 5G scenario where proximity based services can be seen everywhere and the number of D2D pairs can be several times more than the number of the CUEs. In such a case, it is very likely that more than one D2D pair shares one RB, which may incurs severe interference between the D2D user equipmenst (DUE) and the CUE as well as among D2D pairs. Second, existing schemes are proposed for the LTE-Advanced network whose architecture does not consider the mixed deployment of macro and micro base stations and thus is totally different with that of 5G, especially when H-CRAN is adopted as a promising radio access network architecture [5] [7]. Therefore, how to allocate resource for D2D communications in the 5G H-CRAN system avoiding interference should be carefully considered. Recently, the authors of [6] have studied the performance of a downlink H-CRAN with nonuniformly deployed D2D communication, where the SINR coverage and average ergodic rate at a randomly located typical UE are analyzed and the average data traffic delivery latency as a QoS perceived by the CUEs and DUEs is derived. Motivated by [6], in this paper, we study the uplink resource allocation for D2D communication in a H-CRAN network by carefully analyzing the interference of the CUEs and DUEs and designing a rule to choose the candidate sub-channels. Then, the resource allocation problem is formulated as a optimization problem. To solve the problem which is NP-hard, we propose two low computational complexity algorithms, i.e., the throughput aware maximum association (TAMA) algorithm and channel condition aware maximum association (CCAMA) algorithm. Simulation results are finally shown to verify the proposed algorithms. The rest of this paper is organized as follows. Section II describes the system model of D2D communication in H-CRAN networks. In Section III, we formulate the resource allocation problem. Later, in Section IV, we give a detail description of the proposed resource allocation algorithms. The performance evaluation are shown by extensive simulations in Section V. Finally, we conclude this work in Section VI.

4 II. SYSTEM MODEL A. Scenario Description We consider a typical H-CRAN scenario illustrated in Fig.1. A evolved node B (enodeb) is located at the center of a macro cell which is also covered by N R RRHs. The enodeb is interfaced with the BBU pool through a backhaul link to mitigate the cross-tier interference between RRHs and enodeb via centralized cloud-computing based cooperative processing techniques, while the RRHs operate as soft relays by compressing and forwarding the received signals from UEs to the centralized BBU pool through the wired/wireless fronthaul links [5]. The RRHs have their respective coverage which is much smaller than that of the macro cell. Here, the CUE in the coverage of the enodeb and RRHs are called MUE and RUE, respectively, and the transmitter and receiver of the DUE pairs are called DT and DR, respectively. PHY/MAC Fig. 1: A typical H-CRAN scenario with D2D pairs In this model, we consider the scenario where D2D pairs reuse the uplink (UL) resource of cellular networks. We assume the total number of RBs, which are also known as sub-channels, in the UL period is N. Then the set of RBs can be expressed as Φ = {CH 1, CH 2,, CH N }. When there is no ambiguity,we also use the simplified form Φ = {1, 2,, N}. Assuming each RB has been pre-assigned to one MUE, the number of MUEs in the system will also be N. Without loss of generality, we denote the MUE which occupies the i th RB as MUE i. Then the set of MUEs can be expressed as M = {MUE i i = 1,, N}. Assume there are K RUEs, and the set of RUEs is R = {RUE r,i r = 1,..., N R, i = 1,, N}. Here, RUE r,i is the RUE occupying the i th RB in the r th RRH. To avoid serious interference between the RUEs in the same RRH, different RUEs in

5 the same RRH can only access different RBs. Finally, we assume there are M D2D pairs, and the set of D2D pairs is defined as D = {D k k = 1, 2,, M}. In order to make full use of the spectrum resources as well as the potential gain the D2D communication brings, the DUEs are encouraged to share the sub-channels with the CUEs. We assume one D2D pair can only choose one of the sub-channels for reuse. However, one sub-channel can accommodate more than one D2D pair. When one DUE join a sub-channel, it will inevitably cause co-channel interference to the CUEs using the same sub-channel and at the same time, suffer the interference from these CUEs. Note that both of the CUEs and DUEs have acceptable ranges of interference, and the CUEs have the higher priority in interference protection. B. Interference Analysis The interference in a H-CRAN with D2D pairs is illustrated in Fig.2. Before the D2D pairs share DT DR DUE pair Interference from MUE/RUE to DUE Interference from MUE to RUE Interference from RUE to MUE Interference from DUE to MUE/RUE Interference between co-channel RUEs Desired Signal Fig. 2: Interference scenario in H-CRAN with D2D pairs the RBs with CUEs, for a given MUE,i.e., MUE i, the SINR received by the enb can be expressed as γ i = P ih i I i + σ 2, (1) where P i is the transmission power of MUE i, h i is the channel gain from MUE i to the enodeb, I i = r ϕ i P r,i h r,i,b is the total interference from the RUEs reusing the same sub-channel as MUE i, and σ 2 is the thermal noise power which is assumed to be the same for enodeb, all the RRHs and D2D receivers. Here, ϕ i is the set of RRHs which have RUEs in sub-channel i, P r,i is the transmission power of RUE r,i, and h r,i,b is the channel gain from RUE r,i to the enodeb.

6 Similarly, for a given RUE, i.e., RUE r,i, the SINR received by its RRH can be expressed as : γ r,i = P r,ih r,i I r,i + σ 2, (2) where h r,i is the channel gain from RUE r,i to RRHr, and I r,i = P i h B,i,r + s ϕ i\r P s,ih s,i,r is the total interference from the MUE and RUEs sharing the same RBs as RUE r,i. Here, h B,i,r and h s,i,r are the channel gains from MUE i and RUE s,i to RRH r, respectively. To satisfy the QoS requirement, we assume that γ i γ th0 and γ r,i γ th0. Here, γ th0 is the SINR threshold to guarantee the communication quality for CUEs. After D2D pairs access sub-channel i, the interference suffered by the enodeb in this sub-channel can be rewritten as I i = r ϕ i P r,i h r,i,b + k i P D,k h DB,k,i, (3) where i refers to the set of D2D pairs that access sub-channel i, P D,k is the transmission power of D k, and h DB,k,i is the channel gain from the DT of D k to the enodeb. The SINR received by the enodeb thus becomes γ i = P ih i I i +σ 2. While the interference introduces to the RUE r,i should be I r,i = P i h B,i,r + s ϕ i \r P s,i h s,i,r + k i P D,k h DB,k,i, (4) and the SINR received by the RRH r should be γ r,i = P r,ih r,i I r,i+σ 2. The SINR of D k is γ D,k = P D,kh D,k I D,k,i + σ 2 γ thd, (5) where h D,k is the channel gain between the transmitter and receiver of D k, I D,k,i = P i h BD,i,k + r ϕ i P r,i h RD,r,k + t i,t k P D,th DD,t,k is the interference suffered by the DR, and h BD,i,k, h RD,r,k and h DD,t,k are the channel gains from MUE i, RUE r,i and DUE t to the DR of D k, respectively. III. PROBLEM FORMULATION As we have described in the previous section, when D2D pairs are allowed to reuse the RBs of cellular users, more interference is introduced. To maximize the system throughput as well as keep the interference to each user within the acceptable range, we first clarify the sub-channels that are suitable for the reuse of D2D pairs, then we study how to allocate the D2D pairs to these channels. Intuitively, sub-channels in which the CUEs have high SINR have better condition, thus are more suitable for D2D pairs to reuse. Let the set of MUE and RUEs in sub-channel i to be

7 G i = {MUE i, RUE r,i r ϕ i }. For any sub-channel i, a vector q i = {γ i, γ r,i r ϕ i } is used to describe the SINR of MUE and RUEs in it. Let β i = min q i and γ th1 be the SINR threshold. Then if β i γ th1, i.e., all the SINR of CUEs in sub-channel i is above the threshold γ th1, subchannel i can be regarded as a channel with good condition and thus can be selected to reuse. Let Π = {i, β i γ th1 } be the set of available channels. To ensure that CUE in Π will not suffer serious performance degradation after the DUEs access, we require that the minimum SINR of the CUEs in sub-channel i after DUEs access, i.e., β i, should not be lower than the threshold γ th2 (γ th0 < γ th2 ), i.e., β i γ th2 1. Let a k = i {1,, N} denotes D k being allocated to sub-channel i, and a k = 0 denotes D k being not allocated to any sub-channels. Therefore, how to allocate the D2D pairs to Π such that the system throughput is maximized is formulated as the following maximization problem. max (a k,k=1,...,m) N N R N W i log 2 (1 + γ i) + W r,i log 2 (1 + γ r,i) + W n log 2 (1 + γ D,k ) (6) i=1 r=1 i=1 n Π k n s.t. γ i γ th2, i Π (7) γ r,i γ th2, i Π, r ϕ i (8) γ D,k γ thd, k n, n Π (9) n1 n2 =, n 1, n 2 Π, n 1 n 2 (10) Our objective in (6) is to find RBs allocation results a k, k = 1,..., M for D2D pairs to maximize system throughput. Note that n = {D k a k = n}, and W i, W r,i are the bandwidth of the RB occupied by MUE i and RUE r,i, respectively. Constraints in (7) and (8) guarantee that the selected CUEs should have a SINR greater than γ th2. While (9) means that the SINR of the D2D pairs should meet their SINR threshold γ thd. (10) indicates that one D2D pair can be only access to one of the sub-channels. From the objective of (6) we can see that finding the solutions to a k, k = 1,..., M is in fact an integer programming problem. Meanwhile, the constraints in (7), (8) and (9) are non-linear functions of a k, k = 1,..., M. Therefore, the formulated problem in (6) to (10) is a mixed integer nonlinear programming (MINLP) problem which has been proved to be NP-hard. Inspired by the CG model and the GTM+ algorithm [12], in the following section, we propose Throughput Aware Maximum 1 In practice, value of γ th0 should be selected to achieve reasonable communication performance for users. While the values of γ th1 and γ th2 are selected as long as the condition γ th0 < γ th2 < γ th1 is satisfied.

8 Association (TAMA) algorithm and Channel Aware Maximum Association (CCAMA) algorithm to get a tradeoff between computational complexity and system performance. Besides, a Random Assignment Maximum Association (RAMA) algorithm is given as a benchmark. IV. RESOURCE ALLOCATION OPTIMIZATION ALGORITHM In this section, we first characterize the main idea of the RAMA algorithm. Then, we give a detail description of our TAMA algorithm and CCAMA algorithm. Here, the term maximum association (MA) means the number of DUE pairs admitted to resue the RBs of the CUEs can reach the maximum under the constraint of SINR thresholds. The essential difference of these three algorithms lies in the strategies with which we choose appropriate sub-channels for DUEs. A. RAMA algorithm In the RAMA algorithm, sub-channel for each DUE is first randomly chosen from the set of candidate channels Π. Then, for any sub-channel i Π, we sort in descending order the sum interference of MUE i, RUE r,i, r ϕ i, caused by each DUE which attempts to share this subchannel. Further, beginning with the DUE which has the largest sum interference to the CUEs, we check one by one whether the DUE can meet the SINR requirement. If not, we will remove the corresponding DUE from sub-channel i. We finally choose DUE pairs as many as possible until any of the SINR of MUE, RUEs below the SINR threshold γ th2, giving the first consideration to the remaining D2D pairs which have the smaller sum interference to the CUEs in the same sub-channel. In such a way, the DUEs in sub-channel i can not only satisfy their own SINR requirement but also can meet the SINR threshold of the CUEs. After repeating these steps for all the sub-channels which have been pre-allocated, we can get the resource allocation results for the DUE pairs. B. TAMA algorithm The main idea of our TAMA algorithm is shown in the pseudo code of Algorithm 1. First, we define a utility to help the D2D pairs choose one appropriate sub-channel from the set of candidate channels Π. The utility is defined as U Dk (i) = log 2 (1 + P i h i r ϕ i P r,i h r,i,b + P D,k h DB,k,i + σ 2 ) + r ϕ i log 2 (1 + P r,i h r,i P i h r,i,b + s ϕ i\r P s,ih s,i,r + P D,k h DR,k,r + σ 2 ) (11) + log 2 (1 + P D,k h D,k P i h BD,i,k + r ϕ i P r,i h RD,r,k + σ 2 )

9 U Dk (i) can be interpreted as the sum throughput of the MUE, RUEs and the D2D pair D k in subchannel i when only D k joins this sub-channel. For D k, we can get the sub-channel which gives it the largest U Dk. Then this sub-channel temporarily assigned to D k. Next, we sort the sub-channels by the descending order of the number of DUEs in it. Starting from the sub-channel which has largest number of D2D pairs, we first remove the D2D pairs whose SINR requirement cannot be met. Then, for each DUE remaining in this sub-channel, we calculate its sum interference to the CUEs in the same sub-channel, and we sort the DUEs by their sum interference in the ascending order. Giving the first consideration to the DUE with smaller sum interference to CUEs, we will choose D2D pairs as many as possible until any of the SINR of CUEs in the same sub-channel is below the SINR threshold γ th2. The rest few DUEs are then removed from this sub-channel and this sub-channel is marked. Algorithm 1: TAMA algorithm Input: N MUEs, K RUEs, M DUE pairs Output: a 1, a 2,, a M // Phase I: Initialization; Set a k = 0, k = 1,..., M, and set 1, 2,, N to be empty sets; Obtain candidate sub-channel set Π based on γ 1 and β i ; U Π = {c 1, c 2,, c Π }; for each DUE pair k {1, 2,, M} do Find the sub-channel which gives D k the largest utility from U. Denote this sub-channel as c n; c n c n k; // Phase II: Sub-channel allocation iterative process; while U do Sort the sub-channels in U by the descending order of cn. Denote the ordered sub-channel sequence as {m 1, m 2,, m U }; U {m 1, m 2,, m U }; for each k m1 do if P Dk h Dk < γ thd (I Dk + σ 2 ) then remove D k from m1 ; for each DUE m1 do Calculate his sum interference to the MUE, RUEs in sub-channel m 1 ; Set Γ to be an empty set, X = m1, x = 1; while every CUE in sub-channel m t satisfy the SINR threhold γ th2 and x X do Sort the DUEs m1 by their sum interference to the CUEs in ascending order. Denote the sorted DUEs as D a1, D a2,, D ax ; Γ = Γ D ax ; m1 = m1 D ax ; x = x + 1; m1 Γ; for each DUE pair removed from m1 do Set c n to be the sub-channel which gives the largest utility from U {m 1 }; c n c n {k}; U U {m 1 };

10 At last, for each DUE which has been removed from the current sub-channel, we choose it a new sub-channel from the unmarked sub-channels that can bring it the largest utility. The above steps are repeated until all of the sub-channels are allocated. C. CCAMA algorithm As shown in the Algorithm 2, CCAMA algorithm cares about the quality of the sub-channels directly. We first sort the sub-channels in Π according to the value of β i defined in section III. Then, from the sub-channel with the largest value of β i, we allocate D2D pairs join this sub-channel as many as possible, giving the first consideration to the D2D pairs which have the smaller sum interference to the CUEs in the same sub-channel. Further, we check the SINRs of the D2D pairs which have been admitted to join this sub-channel. For the D2D pairs which do not meet their SINR threshold, we remove them from this sub-channel. We repeat these steps for the rest sub-channels and DUE pairs until all the sub-channels are allocated. Algorithm 2: CCAMA algorithm Input: N MUEs, N R K RUEs, M DUE pairs Output: a 1, a 2,, a M // Phase I: Initialization; Set a k = 0, k = 1,..., M, and set 1, 2,, N to be empty sets; Obtain candidate sub-channel set Π based on γ 1 and β i ; Sort the sub-channels in Π by the descending order of β i. Denote the ordered sub-channel sequence as {m 1, m 2,, m U }; U {m 1, m 2,, m U }, t = 1; //Phase II: Sub-channel allocation iterative process; while U do for each DUE pair D do Calculate his sum interference to the each CUE in sub-channel m t ; Set x = 1, X = D ; while every CUE in sub-channel m t satisfy the SINR threhold γ th2 and x X do Sort the DUEs D by their sum interference to the CUEs in ascending order. Denote the sorted DUEs as D a1, D a2,, D ax ; mt = mt D ax ; D = D D ax ; x = x + 1; for each DUE pair k mt do if P Dk h Dk < γ thd (I Dk + σ 2 ) then remove D k from mt ; D = D D k U U {m t }; t = t + 1;

11 D. Complexity analysis In this subsection, we estimate the computational complexity in the worst case of the above three algorithms. In Phase I, TAMA algorithm costs O(N 2 ) for the selecting candidate sub-channels from original N sub-channels and costs O(N M) for choosing the sub-channel which gives the largest utility for each D2D pair. In Phase II, the complexity of TAMA algorithm largely depends upon the the iteration loop and sort operation in each loop. The complexity is O(N 3 +2NM +NM 3 +NK) = O(NM 3 ). So the total computational complexity of TAMA algorithm is O(NM 3 ). Similarly, the total computational complexity CCAMA depends on Phase II which has a complexity of O(NM 3 + 2NM) = O(NM 3 ). So, CCAMA algorithm has the same complexity as TAMA algorithm. The RAMA algorithm has a lower time complexity of O(N 2 + NM + NM 2 ) = O(NM 2 ) for the reason that RAMA randomly chooses sub-channels for the DUE pairs and does not reassign channels for the removed D2D pairs. E. Implementation To implement our proposed algorithms in the H-CRAN scenario illustrated in fig.1, the signaling interaction process should be designed as follows. The receivers of CUE, RUE and DUE obtain the channel state information (CSI) from the corresponding transmitters and feed it back to the enodeb. The enodeb delivers all the CSI and the related SINR threshold to the BBU pool via interface X2. Using the proposed resource allocation algorithms, the BBU pool finally determines which channels should be reused and for a seclected sub-channel, which DUEs can access. The BBU pool sents the resource allocation results to the enodeb, and the enodeb informs each D2D pair whether it can join a sub-channel and which sub-channel it can join. The DUEs which have been assigned sub-channels adjust their transmission parameters and finish their communication. Through the above procedure, we can see that, through centralized management by the enodeb, the global optimization resource allocation and scheduling can be achieved, and the unnecessary handover of the DUEs can be avoided. On the other hand, the resource allocation algorithms are designed to be executed in the BBU pool, so the powerful cloud-computing capabilities can be fully utilized and signaling processing delay can be greatly reduced.

12 TABLE I: Simulation Parameters Parameters Value System bandwith 5 MHz (25 RBs) Radius of enodeb coverage 500m Radius of RRH coverage 150m Number of RRH (N R ) 2 Number of MUEs 25 Number of RUEs in each RRH 25 MUE/RUE/DUE transmission power 23dBm /16dBm /11dBm Noise spectral density -174dBm/Hz Path loss model for CUE and DUE log 10 (d[km]) Path loss model for DUE pair log 10 (d[km]) Antenna Gain for enodeb, RRH and UE 14dBi,12dBi, 0dBi Shadow fading for cellular link 10dB Shadow fading for D2D link 12dB Bandwidth per RB 12*15kHz=180kHz V. PERFORMANCE EVALUATION We perform extensive simulations to evaluate the efficiency of the proposed resource allocation algorithms. For illustration purposes, we provide simulation results for a 5 MHz LTE-A macro-cell with N R = 2 small cells as shown in Fig.3. The enodeb is located at (0, 0) of the plane with a radius of 500m, while the positions of the RRHs is located at (-350,0) and (350,0) with equal radius of 150m. To avoid the severe co-channel interference from the DTs to the CUEs which are far from enodeb or RRHs, we require that the distance from DTs to the enodeb can not exceed d 1, while the distance from DTs to the RRHs can not exceed d 2. We set d 1 = 30m, d 2 = 50m. The position of the DT is described by a uniform distribution in the permitted area, while the DR is inside a region at most L from DT. The MUEs are uniformly distributed in the macro cell and the RUEs are uniformly distributed in the small cell. The throughput is calculated as follows [21] : 0, if γ min > γ η = α log 2 (1 + γ), if γ min < γ < γ max 4.8, if γ max < γ, where η is the estimated spectral efficiency in bps/hz, γ is the SINR and α is the attenuation factor applied to the Shannon bound. According to [21], we set α = 0.75, γ min = 6.5dB, γ max = 19dB. So η achieves 4.8 bps/hz at γ = 19dB or higher. The SINR thresholds of γ th0, γ th1, γ th2 is set to be 6.5dB, 0dB, 3dB, respectively. Table I summarizes a list of simulation parameters and their default values. All results are averaged over 1000 simulations.

13 MacroBS RRH -300 MUE RUE DT DR Fig. 3: The Simulation scenario without DUE RAMA TAMA CCAMA cdf CUE date rate bit/s/hz Fig. 4: The cdf of sum date rate of CUEs in the same sub-channel when M = 40 and L = 25m We first study the cumulative distribution function (cdf) of sum date rate of CUEs in a specific sub-channel. The results when M = 40 and L = 25m are shown in Fig.4. From Fig.4, we can see the CCAMA algorithm performs better than the RAMA algorithm and TAMA algorithm. This is because CCAMA aims to select the sub-channels in which the RUEs and CUEs have higher SINR and thus are more tolerant to the interference caused by D2D pairs. Moreover, the TAMA algorithm performs better than the RAMA algorithm when the date rate is below 9.6bps/Hz, but the gap becomes smaller with the increase of the date rate. When the date rate exceeds 9.6bps/Hz, the percent of the CUEs with TAMA is below that of RAMA. The reason is that RAMA aims to select the sub-channels with the larger utility for a given DUE pair. Larger utility means larger sum throughput of the related

14 RAMA TAMA CCAMA cdf DUE date rate bit/s/hz Fig. 5: The cdf of DUE date rate when M = 40 and L = 25m DUE, RUE and MUE. As a result, the date rate of CUEs and DUEs are balanced. The cdf of average date rate of the DUE pairs are presented in Fig.5. From Fig, 4 we can see that the curve of TAMA nearly overlaps with the curve of RAMA, both at the lower right of the curve of the CCAMA. It illustrates that TAMA improves the performance of the CUEs without damaging the date rate of DUEs. While CCAMA greatly improves the performance of CUEs by limiting the date rate of the DUEs which is consistent with Fig.4. Next, to evaluate the performance of our proposed algorithms, we compare them with the Heuristic algorithm in [?] in terms of throughput and computational complexity, and the results are shown in Fig.6 to Fig.13. System Throughput (bps/hz) TAMA CCAMA RAMA Heuristic withoutdue DUE Number Fig. 6: The system throughput vs. the number of DUE when L = 25m

15 TAMA CCAMA RAMA Heuristic DUE Throughput (bps/hz) DUE Number Fig. 7: The DUE throughput vs. the number of DUE when L = 25m CUE Throughput (bps/hz) TAMA CCAMA RAMA Heuristic DUE Number Fig. 8: The CUE throughput vs. the number of DUE when L = 25m Fig.6 to Fig.8 illustrate the throughput of the system, DUE and CUE vs. the DUE number. From Fig.6, we can see the system throughput increases with the number of DUE pairs linearly with all the three proposed algorithms and the Heuristic algorithm. Moreover, these four algorithms can greatly improve the system throughput when compared with the case without DUE pairs. When the DUE number is 57, the system throughput of TAMA and CCAMA respectively increase 17.9% and 8.7% when compared with RAMA algorithm, and respectively increase 35.1% and 24.6% when compared with the Heuristic algorithm. Similar trends can be seen in Fig.7 when the DUE throughput is considered. From Fig.7, we can observe that TAMA performs the best due to its effectiveness to allocate appropriate sub-channels for DUE pairs. The Heuristic performs the worst since it does not

16 consider the interference from RUEs. In contrast with system throughput and DUE throughput, CUE throughput decreases with the number of the DUEs, as shown in Fig.8. This is because that with the number of the DUE pairs increasing, the CUEs will suffer stronger interference. The CCAMA shows a slow decline since the DUE pairs much prefer the RBs of the CUEs with higher SINR when the CCAMA algorithm is used. The performance of the proposed three schemes are inferior to the Heuristic algorithm in terms of the throughput of CUE. This is because in the Heuristic algorithm only one D2D pair is allowed to access one sub-channel and less interference is introduced to the CUEs System Throughput (bps/hz) TAMA CCAMA RAMA Heuristic withoutdue Distance limitation of DUE pair L(m) Fig. 9: System throughput vs. DUE distance limitation L when M = TAMA CCAMA RAMA Heuristic DUE Throughput (bps/hz) Distance limitation of DUE pair L(m) Fig. 10: DUE throughput vs. DUE distance limitation L when M = 40

17 CUE Throughput (bps/hz) TAMA CCAMA RAMA Heuristic Distance limitation of DUE pair L(m) Fig. 11: CUE throughput vs. DUE distance limitation L when M = 40 Fig.9 to Fig.11 illustrate the throughput of the system, DUE and CUE vs. the distance limitation L. From Fig.9 and Fig.10 we can see that the system throughput and DUE throughput decrease as L increases. This is due to the fact that when L becomes larger, the DUE pair will suffer larger pass loss. As a result, the DUE throughput displays a declining trend which further leads the decline of the system throughput. However, both the TAMA and the CCAMA algorithm achieve higher system throughput and DUE throughput than RAMA algorithm and the Heuristic algorithm. In Fig.11, the CUE throughput with the Heuristic algorithm, CCAMA and RAMA increases L. The reason is that as L increases, the SINR threshold of some DUE pairs can not be satisfied and thus the DUEs can not reuse the RBs of CUEs. As a result, CUEs suffer less interference from DUEs. While CUE throughput with TAMA algorithm displays a different result. The CUE throughput fluctuates slowly when L increases. This is because TAMA aims to allocate sub-channels for DUEs in a greedy way. Even though the distance between a D2D pair increases, appropriate sub-channels can be found for most of the DUEs to satisfy their SINR requirement. Thus the interference suffered by CUEs largely keep unchanged. Therefore, CUE throughput can keep a slowly fluctuant trend. Finally, we verify the computational complexity of TAMA, CCAMA, RAMA and the Heuristic algorithms, and the results are shown in Fig. 12 and Fig. 13, respectively. From Fig. 12 and Fig. 13 we can see that as the number of MUE or DUE increases, the computational complexity of TAMA increases much faster than the other three algorithms. Among the four algorithms, CCAMA has the similar performance to the Heuristic algorithm which has the lowest complexity.

18 0.15 Computational Complexity (s) TAMA CCAMA RAMA Heuristic MUE Number Fig. 12: The computational complexity vs. the number of MUEs. Computational Complexity (s) TAMA CCAMA RAMA Heuristic DUE Number Fig. 13: The computational complexity vs. the number of DUEs. VI. CONCLUSION In this paper, we investigate the resource allocation problem of D2D communications in 5G H- CRAN networks where more than one D2D pair can share the resource of one CUE. A simple channel pre-screening strategy to select the appropriate channels in which both the MUEs and RUEs have higher SINR is proposed. Then, the resource allocation problem is formulated as a maximization problem under the constraints that all the CUEs and DUEs should satisfy their corresponding SINR requirements. To solve the NP-hard problem, we propose the TAMA and CCAMA algorithms which have relatively low complexity. To take full advantage of the cloud computing abilities of BBU pool, the proposed algorithms are designed to be executed in the BBU pool. Extensive simulation results

19 show that compared to the RAMA and the Heuristic algorithm, the TAMA and the CCAMA have better performance in terms of date rate cdf and throughput. REFERENCES [1] A. Asadi, Q. Wang, and V. Mancuso, A survey on device-to-device communication in cellular networks, IEEE Commun. Surveys Tuts., vol. 16, no. 4, pp , Apr [2] J. Liu, N. Kato, J. Ma, et al, Device-to-device communication in LTE-advanced networks: a survey, IEEE Commun. Surveys Tuts., vol. 17, no. 4, pp , Dec [3] M. Tehrani, M. Uysal, and H. Yanikomeroglu, Device-to-device communication in 5G cellular networks: Challenges, solutions, future directions, IEEE Commun. Mag., vol. 52, no. 5, pp , May [4] L. Wei, R. Q. Hu, Y. Qian, and G. Wu, Enable device-to-device communications underlaying cellular networks: Challenges and research aspects, IEEE Commun. Mag., vol. 52, no. 6, pp , Jun [5] M. Peng, Y. Li, J. Jiang, J. Li, and C. Wang, Heterogeneous cloud radio access networks: A new perspective for enhancing spectral and energy efficiencies, IEEE Wireless Commun., vol. 21, no. 6, pp , Dec [6] M. A. Abana, M. Peng, Z. Zhao, et al, Coverage and Rate Analysis in Heterogeneous Cloud Radio Access Networks with Device-to-Device Communication, IEEE Access., vol. 4, pp , May [7] M. Peng, Y. Li, Z. Zhao, and C. Wang, System architecture and key technologies for 5G heterogeneous cloud radio access networks, IEEE Network, vol. 29, no. 2, pp. 6 14, Apr [8] K. Doppler, M. Rinne, C. Wijting, C. Ribeiro, and K. Hugl, Device-to-device communication as an underlay to LTE-Advanced networks, IEEE Commun. Mag., vol. 47, no. 12, pp , Dec [9] K. Doppler, M. P. Rinne, P. Janis, C. Ribeiro, and K. Hugl, Device-to-device communications; functional prospects for LTE-Advanced networks, in Proc. IEEE ICC Workshops, pp. 1 6, Dresden, Germany, Aug [10] T. Han, R. Yin, Y. Xu, and G. Yu, Uplink channel reusing selection optimization for device-to-device communication underlaying cellular networks, in Proc. IEEE PIMRC, pp , Sydney, Australia, Nov [11] M. Zulhasnine, C. Huang, and A. Srinivasan, Efficient resource allocation for device-to-device communication underlaying LTE network, in Proc. IEEE 6th Int. Conf. WiMob Comput., pp , Niagara Falls, Canada, Oct [12] H. Sun, M. Sheng, X. Wang, Y. Zhang, J. Liu, and K. Wang, Resource allocation for maximizing the device-to-device communications underlaying LTE-Advanced networks, in IEEE/CIC International Conference on Communications in China (ICCC) Workshops, pp , Xi an, China, Nov [13] S. A. Ciou, J. C. Kao, and C. Y. Lee, Multi-sharing resource allocation for device-to-device communication underlaying 5G mobile networks, in IEEE International Symposium on Personal Indoor and Mobile Radio Communications (PIMRC), pp , Hong Kong, China, Dec [14] P. Phunchongharn, E. Hossain, and D. I. Kim, Resource allocation for device-to-device communications underlaying LTE-advanced networks, IEEE Wireless Commun. Mag., vol. 20, no. 4, pp , Aug [15] R. Zhang, L. Song, Z. Han, et al., Distributed resource allocation for device-to-device communications underlaying cellular networks, in Proc. IEEE International Conference on Communications (ICC), pp , Budapest, Hungary, Nov

20 [16] C.-H. Yu, K. Doppler, C. B. Ribeiro, and O. Tirkkonen, Resource sharing optimization for device-to-device communication underlaying cellular networks, IEEE Trans. Wireless Commun., vol. 10, no. 8, pp , Aug [17] P. Janis et al., Interference-aware resource allocation for device-todevice radio underlaying cellular networks, in Proc. IEEE VTC-Spring, pp. 1 5, Barcelona, Spain, Jun [18] H. Min, W. Seo, J. Lee, S. Park, and D. Hong, Reliability improvement using receive mode selection in the deviceto-device uplink period underlaying cellular networks, IEEE Trans. Wireless Commun., vol. 10, no. 2, pp , Feb [19] Z. Zhang, R. Q. Hu, Y. Qian, et al., D2D Communication Underlay in Uplink Cellular Networks with Fractional Power Control and Fractional Frequency Reuse, in Proc. IEEE Global Communications Conference (GLOBECOM), pp. 1 7, San Diego, USA, Feb [20] P. Bao, G. Yu, and R. Yin, Novel frequency reusing scheme for interference mitigation in D2D uplink underlaying networks, in Proc. IEEE IWCMC, pp , Cagliari, Italy, Jul [21] 3GPP TR v13.0.0, Evolved Universal Terrestrial Radio Access(E-UTRA);Radio Frequency (RF) System Scenarios(Release 13), Jan.2016,