 Research
 Open Access
 Published:
Joint rate control and power allocation for lowlatency reliable D2Dbased relay network
EURASIP Journal on Wireless Communications and Networking volume 2019, Article number: 111 (2019)
Abstract
Emerging 5G applications impose stringent requirements on network latency and reliability. In this work, we propose a lowlatency reliable devicetodevice (D2D) relay network framework to improve the cell coverage and user satisfaction. Particularly, we develop a crosslayer lowcomplexity resource allocation algorithm, which jointly optimizes the rate control and power allocation from a longterm perspective. The longterm optimization problem is transformed into a series of shortterm subproblems by using Lyapunov optimization, and the objective function is separated into two independent subproblems related to rate control in network layer and power allocation in physical layer. Next, the KarushKuhnTucher (KKT) conditions and alternating direction method of multipliers (ADMM) algorithm are employed to solve the rate control subproblem and power allocation subproblem, respectively. Finally, simulation results demonstrate that the proposed algorithm can reach 99.9% of the optimal satisfaction of D2D pairs with lower average network delay compared to the baseline algorithm. Furthermore, the convergence time of the ADMMbased power allocation algorithm is only about 1.7% of that by using the CVX toolbox.
Introduction
With the explosive growth of mobile applications, it is predicted that approximately 50 billion devices will interconnect to the network by 2020 [1, 2]. Celledge devices are likely to experience poor quality of service (QoS) and quality of experience (QoE) due to the long distance and timevarying channel states between devices and the base station (BS). Devicetodevice (D2D)based relay communication, as a key technology of 5G, can improve data transmission and network coverage by assisting users with inferior channel conditions via multihop transmissions. Specifically, the transmitters (TXs) of D2D pairs can act as relay nodes to reduce the transmission distance and conditions. improve the channel The required data of D2D receivers (RXs) is transmitted from the BS to the nearby TXs, which is stored in the queue buffers of TXs before being transmitted to RXs. Compared to conventional multihop relay network, the D2Dbased relay network can be deployed underlying conventional cellular networks, which enables centralized resource management and coordination.
However, despite the advantages described above, the widespread deployment of D2Dbased relay network still faces some challenges.
Firstly, there lacks a crosslayer resource allocation scheme to guarantee the reliability of network operation as well as to satisfy the lowlatency requirements of applications. Since both the arrival rate of required data at TXs and channel conditions among D2D pairs are varying over time, there requires a joint optimization of the rate control in network layer and transmission rate in physical layer. Traditional schemes, which only consider the optimization of the physicallayer transmission rate while ignoring the arrival rate in network layer, will result in the data imbalance between the arrival rate and the transmission rate. The data imbalance will cause queue backlog and packet drop at TXs due to the limited queue buffering capability of TXs, which leads to intolerable latency and network unreliability.
Second, there lacks an effective online resource allocation scheme which optimizes the network performance from a longterm perspective. Conventional shortterm optimization cannot satisfy the longterm optimization objective and constraints, which will lead to severe performance degradation since resources are only allocated based on instantaneous states and constraints. In addition, it is difficult to obtain accurate future information in practical applications due to the casuality constraint.
Last but not least, the computational complexity of traditional resource allocation algorithms increase dynamically as the number of D2D pairs increases. The reason is that numerous optimization variables in the network are coupled with each other, e.g., the constraint of sum rate and sum power consumption, which leads to prohibitive computational complexity. On the other hand, the optimization problem with coupled constraints is carried out in a time slot basis, which further increases computational complexity.
In this paper, to solve the abovementioned challenges, we propose a crosslayer online joint resource allocation algorithm to optimize the longterm satisfaction of D2D pairs while maintaining network reliability and reduce transmission delay. The main contributions of this work are summarized as follows:

1.
We transform the longterm joint rate control and power allocation problem into a series of shortterm optimization problems by using Lyapunov optimization [3]. At each time slot, the crosslayer joint optimization problem can be decomposed into two separate subproblems and solved independently. The proposed scheme guarantees a \(\left [\mathcal {O}(\frac {1}{V}),\mathcal {O}(V)\right ]\) tradeoff between queue stability and D2D pair satisfaction.

2.
The power allocation problem has high computational complexity due to the coupling of optimization variables among different D2D pairs. To provide a tractable solution, we developed an alternating direction method of multipliers (ADMM)based lowcomplexity power allocation scheme, which decomposes the large problem into a series of smaller subproblems and coordinates the solutions of these subproblems to find the solution of the original problem.

3.
Simulation results demonstrate that the proposed joint optimization scheme can converge quickly and approximate the optimal solution. Moreover, we analyze the tradeoff between satisfaction of D2D pairs and network delay, which proves that significant performance can be improved by the proposed algorithm.
The remaining parts of this paper are organized as follows. Section 2 describes the related works. Section 3.1 introduces the power allocation model, the queue backlog model and satisfaction model. The problem formulation is presented in Section 3.2. The online joint rate control and power allocation algorithm based on Lyapunov optimization and ADMM and the performance analysis of the proposed algorithm are described in Sections 3.3 and 3.4, respectively. Simulation results are presented in Section 4. The conclusion and future works are summarized in Section 5.
Related works
Due to the advantages of D2Dbased relay network such as enlarging cell coverage, reducing network delay and enhancing network reliability, it has aroused widespread concern in both academia and industry. A multidimensional optimization algorithm was proposed to solve a content distribution problem in multihop D2D relay networks [4], which can effectively reduce the average delay in the network. In [5], Zhou et al. studied the D2D communication underlying cellular networks and proposed a joint channel selection and power allocation optimization algorithm to improve the energy efficiency subject to various QoS constraints. In [6], Dang et al. proposed a fullduplex based D2D multihop communication framework, where the data forwarded between D2D transmitters and receivers are assisted by using multiple relays. However, these works mainly focus on the optimization of shortterm network performance, e.g., instantaneous network capacity, energy efficiency, and transmission latency, while ignoring the optimization of timeaverage performance.
Lyapunov optimization is a powerful methodology for studying longterm optimization problems, which is able to transform longterm objective function into a series of shortterm subproblems and transform the longterm constraints into queue stability constraints. It has been applied in various application scenarios such as D2D networks [7], edge computing [8], and OFDMAbased cellular networks [9]. In [3], Sheng et al. proposed a resource allocation algorithm to maximize the energy efficiency of D2D communication underlaying cellular networks subject to the timeaverage and network stability constraints by combining fractional programming and Lyapunov optimization. In [10], Guo et al. proposed a crosslayer joint rate control and resource allocation scheme, which can maximize the timeaverage user satisfaction based on Lyapunov optimization. In [11], Peng et al. considered the energy efficiency optimization problem in multimedia HCRANs subject to individual fronthaul capacity as well as multiple interference constraints to sense queue and proposed an online resource allocation algorithm based on Lyapunov optimization. However, when optimizing the performance of the overall network, the computational complexity increases dramatically with the number of devices due to the coupling of optimization variables and constraints across devices.
ADMM algorithm can solve some specific convex optimization problems with a much lower complexity, because both the primal and dual variables are updated in an alternative direction to increase the convergence speed [12]. It employs a decompositioncoordination procedure, in which the global optimization problem is firstly decomposed into numerous small subproblems, and then the solutions to these subproblems are calculated, updated, and coordinated to find a solution to the global problem. ADMM has been widely adopted in addressing largescale optimization problems in various application scenarios. In [13], Li et al. decoupled the power constraint and objective function by employing ADMM and proposed a robust design of transceiver multicell distributed antenna network with numerous remote radio heads. In [14], Ling et al. proposed a weighted ADMM algorithm to solve the consensus optimization problem in decentralized networks, which is able to minimize the communication cost of optimization. In [15], Chen et al. combined the ADMM algorithm with the convexconcave procedure to reduce the complexity and improve system performance for largescale multigroup multicast beamforming problems.
Different from the abovementioned works, we propose a longterm crosslayer joint optimization of rate control and power allocation scheme for D2D relay networks by combining Lyapunov optimization and ADMM. Various constraints of network reliability, transmission delay, and power consumption have been taken into consideration. The difference between [10] and our work is that instead of solving the largescale network optimization problem directly, we develop a lowcomplexity power allocation algorithm based on ADMM.
The proposed algorithm is not constrained to D2Drelay networks. It can be extended to solve similar joint rate control and power allocation problems in different application scenarios such as task offloading [8] and energy harvesting [16].
Theoretical method
System model
In the traditional cellular network, the QoS of some celledge devices cannot be well satisfied due to the long distance and timevarying channel gain between the devices and BS. Thus, D2D relay networks can be utilized to enhance cell coverage via a twohop communication manner. As shown in Fig. 1, we consider a D2Dbased relay network in a single cell, which consists of one BS and M D2D pairs. In this work, we assume that the relay selection has been finished, which has been studied in many papers [17, 18] so that it is left out of consideration in this work. That is, there exists a onetoone mapping between D2D TXs and D2D RXs. The BS operates in a timeslotted manner and collects the queue state information (QSI) and channel state information (CSI) at each time slot [19]. The set of time slots is defined as \(\mathcal {T}=\{0,\cdots, t, \cdots, T1\}\). Taking the D2D pair m as an example, the data requested by the RX is firstly transmitted from the BS to the TX. The TX maintains a queue temporarily to store the arrival data, which is then delivered to the RX. At each time slot, assume that A_{m}(t) Mbits of data arrive at the TX of D2D pair m from the BS, which is assumed to be independently and identically distributed (i.i.d.) over time slots with the maximum arrival rate A_{m,max}. In addition, assume that D_{m}(t) Mbits of data depart from the TX of D2D pair m in each time slot, which is related to the channel state and transmission power between the D2D pairs.
Power allocation model
D2Dbased relay communication can be divided into inband communication (also known as LTE direct) and outband communication. Inband D2D communication can be further classified into the categories of underlay and overlay D2D communication [20]. In the circumstances of inband overlay D2Dbased relay network, the D2D communication occupies the licensed spectrum owned by the cellular operators. The cellular operators are able to employ complex interference mitigation techniques to provide higher satisfaction for D2D pairs compared to the use of unlicensed spectrum [21]. Hence, we assume that the transmission data at TXs is transmitted to the RXs through a series of orthogonal channels in the LTE direct system [22], which means there is no interference among D2D pairs. The transmission rate of D2D pair m is expressed as:
where B_{m}(t) represents the channel bandwidth allocated to the D2D pair m, p_{m}(t) is the transmission power of D2D pair m, h_{m}(t) is the channel gain of D2D pair m, and \({\sigma _{0}^{2}}\) is the power of additive white Gaussian noise. Without loss of generality, assume that h_{m}(t) is i.i.d. over time slots, and takes values in a finite state space. Moreover, h_{m}(t) keeps constant during one time slot but varies across different time slots.
In order to reduce the power consumption of the network, the longterm timeaverage power consumption for arbitrary D2D pair m is defined as:
and instantaneous transmission power for each time slot t is defined as:
where P_{m,ave} and P_{max} are the timeaverage and instantaneous power consumption constraints, respectively.
Queue backlog model and satisfaction model
Due to the fact that the requested data cannot be transmitted instantaneously to the RX, it has to be stored at the queue of the TX temporarily. The queue backlogs at the TXs are denoted as Q(t)\(\buildrel \Delta \over = \{Q_{1}(t), \cdots, Q_{m}(t), \cdots, Q_{M}(t) \}\) at each time slot t, which are determined by the arrival rate and the transmission rate. Hence, the dynamic queue consists of the arrival data and the departure data. The data arrival process of the queue is determined by the rate control policy, which affects the amount of data enters into the queue and the satisfaction of D2D pairs. The data departure process of the queue is determined by the power allocation policy, which affects the amount of data leaves the queue and the network latency and stability. Thus the queue Q_{m}(t) at TX of D2D pair m evolves in accordance with the following expression:
There exists no data overflow if the longterm average transmission data of the queue is larger than or equal to the longterm average arrival data of the queue. Thus, the queue Q_{m}(t) is mean rate stable [23] if
Equation (5) implies that the data in the stable network should be transmitted within finite delay and the stability of the network is guaranteed if the queue length is finite.
In addition, we define the satisfaction of D2D pair m as a nondecreasing concave function [10, 24]:
where γ_{m} is a predefined parameter related to the service of RX in D2D pair m. The logarithmic function indicates that the marginal increment of the satisfaction declines gradually with A_{m}(t).
Problem formulation
The objective of this paper is to maximize the longterm timeaverage satisfaction of D2D pairs. The optimization problem can be formulated as follows:
where C_{1} represents that the arrival rate cannot exceed the maximum tolerance rate of TX. C_{2}∼C_{4} are the nonnegative transmission power, instantaneous transmission power, and timeaverage transmission power constraints, respectively. C_{5} denotes the queue stability constraint in Eq. (5).
Next, we propose a crosslayer online optimization algorithm to solve P1 based on the Lyapunov optimization algorithm and ADMM algorithm.
Joint rate control and power allocation optimization
In this section, we firstly transform the longterm timeaverage optimization objective into a series of online subproblems by using the Lyapunov optimization. Then, we describe the detailed procedures of crosslayer joint optimization problem of rate control and power allocation.
Problem transformation
It is noticed that there exist longterm constraints in original problem P1. To handle the longterm timeaverage power consumption constraint, C_{4} can be transformed into queue stability constraint by employing virtual queue [25]. The virtual queue is defined as Z(t)\(\buildrel \Delta \over = \{Z_{1}(t), \cdots, Z_{m}(t), \cdots, Z_{M}(t) \}\), and Z_{m}(t) evolves as follows:
It is worth noting that there is no actual queue data in queue Z_{m}(t), which is only proposed to satisfy constraint C_{4}.
Theorem 1
If virtual queue Z_{m}(t) is mean rate stable, then C_{4} holds automatically.
Proof
The detailed proof can be found in [10]. □
According to Theorem 1, P1 can be rewritten as:
Lyapunov optimization
Let Θ(t) =[Q(t), Z(t)] be the concatenated vector of queue length in the network. Define the Lyapunov function as a measure of total queue length at each time slot t:
At each time slot, the conditional Lyapunov drift is expressed as:
At each time slot, it can be observed that the Lyapunov function is able to control Lyapunov drift’s ultimate value by adjusting the final queue length. According to the Little’s Theorem [26], the average delay is proportional to the average queue length, which is expressed as:
where D_{net} is the timeaverage delay, which can be adjusted by minimizing Lyapunov drift.
To minimize the network delay and maximize the longterm timeaverage satisfaction of all D2D pairs, the driftminusreward term is defined as:
where V is a nonnegative control parameter that is chosen to affect the relative performance of the network delay and satisfaction of D2D pairs, i.e., the tradeoff between “network delay minimization” and “satisfaction maximization of D2D pairs”.
Theorem 2
At each time slot, under any possible Θ(t) with given V≥0, the driftminusreward term is upper bounded by:
where C is a positive constant which satisfies:
Proof
The detailed proof can be found in Appendix 1. □
According to the principle of Lyapunov optimization, P2 can be transformed into optimizing the driftminusreward with the constraints C_{1}∼C_{3}. The second term of the right hand side (RHS) in (14) involves only the rate control variables {A_{m}(t)}, while the third and the fourth term of the RHS in (14) involves only the power allocation variables {p_{m}(t)}. Therefore, P2 can be decoupled into two independent rate control subproblem and power allocation subproblem.
Rate control
The rate control subproblem can be formulated as:
The secondorder derivative of P3 is greater than zero, which indicates that it is a convex function with respect to {A_{m}(t)} and can be solved by KKT conditions [27]. The Lagrangian associated with Q_{m}(t)A_{m}(t)−VS_{m}(A_{m}(t)) is expressed as:
where λ is the Lagrange multiplier.
The firstorder conditions of (17) with respect to A_{m}(t) is expressed as:
Considering the primal constraint A_{m}(t)≤A_{m,max}, dual constraint λ≥0 and complementary slackness constraint λ(A_{m}(t)−A_{m,max})=0, the optimal rate is given by:
It is noticed that the optimal rate is inversely proportional to the queue length Q_{m}(t). Thus the online algorithm can adjust the arrival rate based on the queue length.
ADMMbased power allocation algorithm
The power allocation subproblem is expressed as:
It can be proved that P4 is a convex function regarding to p_{m}(t) by calculating the corresponding secondorder derivative. However, the calculating time by using toolbox to solve the optimization objective is very large due to the coupled power variables and the dynamically increasing of D2D pairs. Hence, we develop a low complexity ADMMbased power allocation algorithm to solve the problem P4.
The ADMM algorithm is simple but powerful to research distributed convex optimization problems, which had been successfully applied in many aspects, i.e., statistical learning problems, timeseries analysis and scheduling [28]. In general, the diverse application domains are characterized with the largescale problems, highdimensional data processing and distributed collection of large scale data in stochastic process [29]. The basic produces of ADMM algorithm is to alternatively update primal variables and dual variables in an iterative manner [12]. It is beneficial to find the optimal solution with low computational complexity by using decompositioncoordination procedure.
In order to obtain the optimal solution, we rewrite the power variables into p1={p_{1},p_{2},⋯,p_{n}} and p2={p_{n+1},p_{n+2},⋯,p_{M}}. Then let x=p1 and z=p2. Therefore, we can transform P4 into the following ADMM format [30]:
where x∈R^{n×1},z∈R^{(M−n)×1},J∈R^{1×n},K∈R^{1×(M−n)}, and c=P_{max}. J and K are unit vectors. f(x) and g(z) satisfy:
There exist two basic forms of the ADMM algorithm, such as the unscaled form and the scaled form [16]. For the sake of simplicity, the scaled ADMM algorithm is employed in this paper. The augmented Lagrangian of P5 is expressed as:
where ρ>0 represents the penalty parameter, which is related to the convergence speed of ADMM algorithm. β is the vector form of the Lagrange multipliers.
The ADMM algorithm in scaled form consists of the following iterations regarding to the primal variables and Lagrange multipliers:
where i denotes the index of iteration.
Next, based on the analysis of optimality conditions [31], the primal residual is expressed as:
and the dual residual is expressed as:
Therefore, the reasonable termination criteria satisfies:
where ε^{pri}>0 and ε^{dual}>0 denote feasibility tolerances with respect to primal conditions and dual conditions. Consequently, the ADMMbased power allocation algorithm is summarized in Algorithm 1.
Performance analysis
In this subsection, we analyze the performance of Lyapunov optimization algorithm and ADMM algorithm, respectively.
Performance of Lyapunov optimization algorithm
Due to the fact that all physical quantities cannot be infinitely large in the practical network, we consider that the arrival rate, transmission rate, power consumption, and satisfaction of D2D pairs are all bounded, i.e., \(\mathbb {E}\{A_{m}(t)\}\le \theta, \mathbb {E}\{D_{m}(t)\}\le \theta \), \(\mathbb {E}\{p_{m}(t)\}\le \theta \), \(S_{min} \le \mathbb {E}\{S_{m} (A_{m}(t))\}\le S_{max}\), where θ, S_{min} and S_{max} are finite nonnegative constants.
Assume that there is at least one feasible solution to problem P1 which satisfies constraints C_{1}∼C_{5} and the bounded values mentioned above. For arbitrary small positive real number ε and ζ, the following expressions hold [32]:
where \(A^{*}_{m}(t), D^{*}_{m}(t), p^{*}_{m}(t)\), and \(S^{*}_{m} (A^{*}_{m}(t))\) are the corresponding resulting values, and S_{opt} denotes the theoretical optimal value.
Theorem 3
Suppose that the problem P1 is feasible, h_{m}(t) is i.i.d. with slotted time, and that \(\mathbb {E} \{{L(\Theta (0))}\}<\infty \). For arbitrary V≥0, the following properties corresponding to the proposed algorithm hold:

1.
Q_{m}(t) and Z_{m}(t) are mean rate stable, which guarantee the constraint C_{6}.

2.
The timeaverage satisfaction of D2D pairs satisfies:
$$\begin{array}{*{20}l} \underset{T \to \infty}{\lim} \frac{1}{T}\sum_{t =0}^{T1} \mathbb{E}\{S_{m} (A_{m}(t))\} \ge S_{opt}  \frac{C}{V} \end{array} $$(34) 
3.
The timeaverage queue length satisfies:
$$\begin{array}{*{20}l} \underset{T \to \infty}{\lim} \frac{1}{T}\sum_{t =0}^{T1} \sum_{m=1}^{M} \mathbb{E}\{Q_{m}(t)\} \le \frac{C+V(S_{max}S_{opt})}{\varepsilon} \end{array} $$(35)
Proof
The detailed proof can be found in Appendix 2. □
Based on the above analysis, we conclude that the proposed rate control and power allocation algorithm can satisfy the queue stability constraint and achieve the tradeoff between network delay and satisfaction of D2D pairs by adjusting parameter V.
Convergence of ADMM algorithm
The objective function of P5 is closed, proper, and convex, and the Lagrangian L_{0}(x,z,β) has a saddle point. Thus, the iterations satisfy the following convergence properties.
Theorem 4
The residual convergence, objective convergence and dual variable convergence are expressed as follows:

1.
Residual convergence: The primal and dual residuals converge to 0 as i→∞, which implies that the iterations approach feasibility.

2.
Objective convergence: The objective function of P5 eventually converges to the optimal value under the stopping criterion as i→∞.

3.
Dual variable convergence: The dual variable β[i+1] converges to dual optimal value as i→∞.
Proof
The detailed proof can be found in [33]. □
Results and discussion
In this section, we verify the system performance of the proposed algorithm through simulation results. Assume that there are M=4 D2D pairs for data transmission in each time slot, and the corresponding number of subchannels is set as 4. Detailed parameters are summarized in Table 1 [10, 34, 35].
Figure 2 shows the queue length of the data queue Q_{m}(t) and the virtual power queue Z_{m}(t) versus time slots, respectively. It can be observed that both the data queue and the virtual queue are bounded after a period of time, which guarantees the stability of network. The phenomenon can be well explained by the first property in Theorem 3.
Figure 3 shows the transmission rate and arrival rate versus time slots, respectively. It can be observed that both the transmission rate and arrival rate are stable, which guarantees the longterm timeaveraged power constraint. In addition, the fluctuation of arrival rate in Fig. 3a is smaller than that of transmission rate in Fig. 3b. The reason is that the arrival rate is only related to the stable queues Q(t) and Z(t), while the transmission rate is not only related to the stable queues Q(t) and Z(t), but also related to the transmission power and channel gain, which is varying across time slots. Therefore, the transmission rate remains stable over a larger range of values compared with the arrival rate.
Figure 4 shows the satisfaction of D2D pairs and the average network delay versus the control parameter V, respectively. It can be observed that the satisfaction of D2D pairs and the average network delay increase as V increases. The reason is consistent with the second and third properties of Theorem 3. Furthermore, the snapshot based algorithm [10] is denoted as the baseline algorithm to maximize instantaneous satisfaction of all D2D pairs for the purpose of comparison. Typically, the baseline algorithm only considers the service demands of users in shortterm while ignoring the power constraint and stability of the network in longterm. It can be observed that the proposed algorithm can approximate to 99.9% of the optimal satisfaction of D2D pairs with lower average network delay. The reason is that the proposed algorithm can simultaneously optimize the arrival rate in network layer and the power allocation in physical layer from a global longterm perspective. Figure 5 shows the residual convergence of the ADMM algorithm versus the number of iterations. The stopping criterion constraints ε^{pri} and ε^{dual} are represented by the dotted lines in Fig. 5a and b, respectively. It can be observed that the stopping criterion can be satisfied after 57 iterations in Fig. 5a and 4 iterations in Fig. 5b. The iterations will stop if and only if both the primal residual and the dual residual conditions are satisfied simultaneously, i.e., after 57 iterations, which is consistent with the first property of Theorem 4.
Figure 6 shows the the optimal convergence of the ADMM algorithm versus the number of iterations. It shows the convergence of objective function, which demonstrates that the proposed algorithm can obtain the optimal solution of the objective function, i.e., 8.75. The reason can be explained by the second property of Theorem 4. In addition, the dual variables will converge after multiple iterations due to the derivation from the primal variables, which is consistent with the third property of Theorem 4.
Table 2 shows the simulation analysis by using CVX toolbox approach and ADMMbased power allocation algorithm. On one hand, it can be observed that the convergence time of ADMMbased power allocation algorithm requires 0.021236 seconds, which is only about 1.7% of the convergence time by using the CVX toolbox approach. Due to the fast convergence of the ADMM algorithm, the ADMMbased power allocation algorithm can get the optimal solution with lower computational complexity within error tolerance at each time slot. On the other hand, 4 TXs share the total power of 0.8 W at each time slot, which proves the feasibility of the proposed power allocation algorithm. The table shows that the power allocation varied with the channel states and the service demands of different TXs.
Conclusion
In this paper, we propose the D2Dbased relay framework to improve the reliability and reduce delay of the network. Based on the Lyapunov optimization, the crosslayer joint optimization problem is separated into two independent rate control subproblem and power allocation subproblem, which can be solved by using KKT conditions and ADMM algorithm, respectively. The tradeoff between network delay minimization and satisfaction maximization of D2D pairs with \([\mathcal {O}(\frac {1}{V}),\mathcal {O}(V)]\) can be obtained by the proposed algorithm, which has been verified by the simulation results.
In the future work, we will improve the performance of the network while considering device mobility from a longterm perspective. Furthermore, we will study how to combine Lyapunov optimization and ADMM algorithm with big data or machine learning to deal with high complexity problems.
Appendix
Appendix 1: Proof of Theorem 2
According to (max{Q−b,0}+A)^{2}≤Q^{2}+b^{2}+A^{2}+2Q(A−b), where Q, b and A are nonnegative real numbers, we obtain the following expression:
where
Add and subtract \(V\mathbb {E}\left \{ \sum _{m=1}^{M} S_{m} (A_{m}(t))\Theta (t) \right \}\) to the both sides of Eq. (36), then merge the same variables to prove (14).
Appendix 1: Proof of Theorem 3
In order to minimize the RHS of (14) with constraints C_{1}∼C_{3}, we obtain the following expression:
Plugging (31), (32) and (33) into (38), and taking ζ→0, the following expression satisfys:
By applying the rules of iterated expectation and telescoping sums, we obtain:
Since \(\mathbb {E} \{{L(\Theta (0))}\}=0, \mathbb {E} \{{L(\Theta (t))}\}\ge 0\) and Q_{m}(t)≥0, rearrange terms and we obtain:
Dividing by VT at both sides and taking T→∞, (34) can be proved.
Similarly, by rearranging terms, (40) can be rewritten as:
Dividing by εT at both sides and taking T→∞, (35) can be proved.
According to the definition of Lyapunov function, taking expectation and rearranging terms yield:
We obtain the following expression by plugging (43) into (40) and rearranging terms:
According to inequality \(\mathbb {E}^{2}\{Q_{m}(t)\} \le \mathbb {E}\left \{{Q^{2}_{m}}(t)\right \}\), we obtain:
Dividing by T and taking T→∞, we can obtain:
which proves the stability of queues by changing the order of taking expectation and limit. Queue Q_{m}(t) is mean rate stable according to (5), thus C_{5} can be satisfied. Similar proof can be applied to Z_{m}(t) with slight modification.
Abbreviations
 ADMM:

Alternating direction method of multipliers
 BS:

Base station
 CSI:

Channel state information
 D2D:

Devicetodevice
 i.i.d.:

Independent and identical distribution
 KKT:

KarushKuhnTucher
 QoE:

Quality of experience
 QoS:

Quality of service
 QSI:

Queue state information
 RHS:

Right side hand
 RX:

D2D receiver
 TX:

D2D transmitter
References
 1
Z. Zhou, H. Liao, B. Gu, K. M. S. Huq, S. Mumtaz, J. Rodriguez, Robust mobile crowd sensing: when deep learning meets edge computing. IEEE Netw.32(4), 54–60 (2018).
 2
Z. Zhou, H. Yu, C. Xu, Y. Zhang, S. Mumtaz, J. Rodriguez, Dependable content distribution in D2Dbased cooperative vehicular networks: a big dataintegrated coalition game approach. IEEE Trans. Ind. Informat.19(3), 953–964 (2018).
 3
M. Sheng, Y. Li, X. Wang, J. Li, Y. Shi, Energy efficiency and delay tradeoff in devicetodevice communications underlaying cellular networks. IEEE J. Sel. Areas Commun.34(1), 92–106 (2016).
 4
C. Xu, J. Feng, Z. Zhou, J. Wu, C. Perera, Crosslayer optimization for cooperative content distribution in multihop devicetodevice networks. IEEE Internet Things J.99:, 1–1 (2017).
 5
Z. Zhou, M. Dong, K. Ota, G. Wang, L. T. Yang, Energyefficient matching for resource allocation in D2D enabled cellular networks. IEEE Internet Things J.3(3), 428–438 (2016).
 6
S. Dang, G. Chen, J. P. Coon, Multicarrier relay selection for fullduplex relayassisted OFDM D2D systems. IEEE Trans. Veh. Technol.67(8), 7204–7218 (2018).
 7
L. Pu, X. Chen, J. Xu, X. Fu, D2D fogging: An energyefficient and incentiveaware task offloading framework via networkassisted D2D collaboration. IEEE J. Sel. Areas Commun.34(12), 3887–3901 (2016).
 8
L. Chen, S. Zhou, J. Xu, Computation peer offloading for energyconstrained mobile edge computing in smallcell networks. IEEE/ACM Trans. Netw.26(4), 1619–1632 (2018).
 9
Y. Guo, Q. Yang, J. Liu, K. S. Kwak, Crosslayer rate control and resource allocation in spectrumsharing OFDMA smallcell networks with delay constraints. IEEE Trans. Veh. Technol.66(5), 4133–4147 (2017).
 10
Y. Guo, Q. Yang, K. S. Kwak, Qualityoriented rate control and resource allocation in timevarying OFDMA networks. IEEE Trans. Veh. Technol.66(3), 2324–2338 (2017).
 11
M. Peng, Y. Yu, H. Xiang, H. V. Poor, Energyefficient resource allocation optimization for multimedia heterogeneous cloud radio access networks. IEEE Trans. Multimedia. 18(5), 879–892 (2016).
 12
C. Liang, F. R. Yu, H. Yao, Z. Han, Virtual resource allocation in informationcentric wireless networks with virtualization. IEEE Trans. Veh. Technol.65(12), 9902–9914 (2016).
 13
N. Li, Z. Fei, C. Xing, D. Zhu, M. Lei, Robust lowcomplexity MMSE precoding algorithm for cloud radio access networks. IEEE Commun. Lett.18(5), 773–776 (2014).
 14
Q. Ling, Y. Liu, W. Shi, Z. Tian, Weighted ADMM for fast decentralized network optimization. IEEE Trans. Signal Process.64(22), 5930–5942 (2016).
 15
E. Chen, M. Tao, ADMMbased fast algorithm for multigroup multicast beamforming in largescale wireless systems. IEEE Trans. Commun.65(6), 2685–2698 (2016).
 16
G. Zhang, Y. Chen, Z. Shen, L. Wang, Distributed energy management for multiuser mobileedge computing systems with energy harvesting devices and QoS constraints. IEEE Internet Things J., 1–1 (2018).
 17
Z. Zhou, F. Xiong, C. Xu, Y. He, S. Mumtaz, Energyefficient vehicular heterogeneous networks for green cities. IEEE Trans. Ind. Informat.14(4), 1522–1531 (2018).
 18
C. Xu, J. Feng, B. Huang, Z. Zhou, S. Mumtaz, J. Rodriguez, Joint relay selection and resource allocation for energyefficient D2D cooperative communications using matching theory. Appl. Sci.7(5), 491–515 (2017).
 19
J. Liu, Q. Yang, S. G, Congestion avoidance and load balancing in content placement and request redirection for mobile CDN. IEEE/ACM Trans. Netw.26(2), 851–863 (2018).
 20
A. Asadi, Q. Wang, V. Mancuso, A survey on devicetodevice communication in cellular networks. Commun. Surveys Tuts.16(4), 1801–1819 (2014).
 21
H. Kim, J. Na, E. Cho, in Proceedings of the International Conference on Information Networking 2014. Resource allocation policy to avoid interference between cellular and D2D links/ and D2D links in mobile networks (IEEEPhuket, 2014), pp. 588–591.
 22
P. Mach, Z. Becvar, T. Vanek, Inband devicetodevice communication in OFDMA cellular networks: A survey and challenges. Commun. Surveys Tuts.17(4), 1885–1922 (2015).
 23
M. J. Neely, Stochastic Network Optimization with Application to Communication and Queueing Systems (Morgan and Claypool, San Rafael, 2010).
 24
M. Chen, S. S. M. Ponec, J. Li, P. A. Chou, Utility maximization in peertopeer systems with applications to video conferencing. IEEE/ACM Trans. Netw.20(6), 1681–1694 (2012).
 25
W. Bao, H. Chen, Y. Li, B. Vucetic, Joint rate control and power allocation for nonorthogonal multiple access systems. IEEE J. Sel. Areas Commun.35(12), 2798–2811 (2017).
 26
D. Bertsekas, R. Gallager, Data Networks (Prentice hall, NJ,USA, 1987).
 27
L. Duan, T. Kubo, K. Sugiyama, J. Huang, T. Hasegawa, J. Walrand, Motivating smartphone collaboration in data acquisition and distributed computing. IEEE Trans. Mobile Comput.13(10), 2320–2333 (2014).
 28
C. Lu, J. Feng, S. Yan, Z. Lin, A unified alternating direction method of multipliers by majorization minimization. IEEE Trans. Pattern Anal. Mach. Intell.40(3), 527–541 (2018).
 29
Y. Li, G. Shi, W. Yin, L. Liu, Z. Han, A distributed ADMM approach with decompositioncoordination for mobile data offloading. IEEE Trans. Veh. Technol.67(3), 2514–2530 (2018).
 30
Y. Wang, L. Wu, S. Wang, A fullydecentralized consensusbased ADMM approach for DCOPF with demand response. IEEE Trans. Smart Grid. 8(6), 2637–2647 (2017).
 31
G. Chen, Q. Yang, An ADMMbased distributed algorithm for economic dispatch in islanded microgrids. IEEE Trans. Ind. Informat.14(9), 3892–3903 (2018).
 32
L. Georgiadis, M. J. Neely, L. Tassiulas, Resource allocation and crosslayer control in wireless networks. J. Found. Trends Netw.1(1), 1–144 (2006).
 33
S. Boyd, N. Parikh, E. Chu, B. Peleato, J. Eckstein, Distributed optimization and statistical learning via the alternating direction method of multipliers. Found. Trends Mach. Learn.3(1), 1–122 (2011).
 34
Z. Zhou, J. Feng, B. Gu, B. Ai, S. Mumtaz, J. Rodriguez, M. Guizani, When mobile crowd sensing meets UAV: Energyefficient task assignment and route planning. IEEE Trans. Commun.66(11), 5526–5538 (2018).
 35
Z. Zhou, K. Ota, M. Dong, C. Xu, Energyefficient matching for resource allocation in D2D enabled cellular networks. IEEE Trans. Veh. Technol.66(6), 5256–5268 (2017).
Acknowledgements
Thanks to Professor Liangrui Tang of North China Electric Power University for his guidance of this research.
Funding
This work was partially supported by the National Natural Science Foundation of China (NSFC) under grant number 61601181; the Fundamental Research Funds for the Central Universities under grant number 2017MS001.
Availability of data and materials
Not applicable.
Author information
Affiliations
Contributions
YW wrote the manuscript and made a part of simulations. YH gave some suggestions and made a part of simulations. ZZ proposed the idea and revised this paper. All authors took an active role in the writing process of the document, and read and approved the final manuscript.
Corresponding author
Correspondence to Zhenyu Zhou.
Ethics declarations
Competing interests
The authors declare that they have no competing interests.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricteduse, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
About this article
Received
Accepted
Published
DOI
Keywords
 D2Dbased relay network
 Rate control
 Power allocation
 Lyapunov optimization
 ADMM algorithm
 Lowlatency and reliability