Joint Cooperative Computation and Interactive Communication for RelayAssisted Mobile Edge Computing
Abstract
To realize cooperative computation and communication in a relay mobile edge computing system, we develop a hybrid relay forward protocol, where we seek to balance the execution delay and network energy consumption. The problem is formulated as a nondifferentialbe optimization problem which is nonconvex with highly coupled constraints. By exploiting the problem structure, we propose a lightweight algorithm based on inexact block coordinate descent method. Our results show that the proposed algorithm exhibits much faster convergence as compared with the popular concaveconvex procedure based algorithm, while achieving good performance.
Xihan Chen, Qingjiang Shi, Yunlong Cai, Minjian Zhao, and Mingjie Zhao
\address\myfont Department of ISEE, Zhejiang University, Hangzhou, China
\myfont College of EIE, Nanjing University of Aeronautics and Astronautics, Nanjing
, China
\myfont School of Info. and Sci. Tech., Zhejiang SciTech University, Hangzhou, China
\myfont chenxihan@zju.edu.cn, qing.j.shi@gmail.com, ylcai@zju.edu.cn, mjzhao@zju.edu.cn, zhaomjx@163.com
\ninept
1 Introduction
The mobile edge computing (MEC) has been considered as a new network architecture that enables cloud computing capabilities and IT service environment at the edge of network. This architecture has the potential to significantly reduce latency, avoid congestion and prolong the battery lifetime of mobile devices, by pushing data intensive tasks towards the edge and locally processing data in proximate MEC server[1].
Recently, MEC has gained a lot of interest [4, 2, 3, 5, 6]. In [2] and [3], the authors derived the optimal resource allocation solution for a singleuser MECO system with multiple elastic tasks to minimize the average execution latency of all tasks under power constraints. You et al. [4] investigated the optimal resource and offloading decision policy to minimize the weighted sum mobile energy consumption under the computation latency constraint in a multiuser TDMA/OFDMA MEC system. The work [5] considered the joint optimization of radio and computational resources for computation offloading in a dense deployment scenario, in the presence of intercell interference. Wang et al. [6] presented an ADMMbased decentralized algorithm for computation offloading, resource allocation and internet content caching optimization in heterogeneous wireless cellular networks with MEC.
The aforementioned works focused on scenario where mobile terminals offload their computational tasks to MEC server and then the latter feedbacks the results to mobile terminals. In this paper, differently from the current MEC works, we consider a cooperative computing and interactive communication system as shown in Fig. 1 where user A wants to share his computational result with user B with the aid of a Relay equipped with an MEC Server (abbreviated as RMES). To reduce the execution delay and the energy consumption, we seek to minimize a weighted sum of the execution delay and network energy consumption subject to radio and computation resource constraints at users and RMES. The problem is a nonconvex nondifferentiable problem with highly coupled constraints. By exploiting the problem structure as fully as possible, we adopt smooth approximation and inexact block coordinate descent method to address the difficulties arising from the nondifferentiability, nonconvexity and constraint coupling. Simulation results show that the proposed algorithm significantly improves the speed of convergence while achieve good performance as compared with the concaveconvex procedure (CCCP)based algorithm.
2 System model and problem statement
As shown in Fig. 1, we consider a computational result sharing (CRS) system where user A wants to share its computational result with user B over a relay channel. It is assumed that the relay is equipped with an MEC server and it has the capability to help user A process the tasks. Given the computational tasks, user A decides whether to locally process the tasks or offload all its tasks or part of its tasks to the RMES. To support CRS, we propose a hybrid relay forward (HRF) protocol, where, the relay channels consist of an amplifyandforward (AF) relay subchannel and a decodeandforward (DF) relay subchannel over orthogonal frequency bands. The AF subchannel is used to deliver the computational result shared by user A while the DF subchannel is used to receive the computational tasks required by user . Apparently, user A’s offloading strategy could impact the endtoend delay and the system energy consumption. This work aims to balance these two system performance metric by properly allocating the system computational and communication resources.
Without loss of generality, we assume that the AF relay subchannel and the DF relay subchannel occupy the same width bandwidth . Let denote the AF relay subchannel between user A and relay, and denote the AF relay subchannel between the relay and user B. Likely, let denote the DF relay subchannel between user A and relay, and denote the DF relay subchannel between the relay and user B. Moreover, suppose that partial offloading is implemented with the data partitioned oriented tasks [7, 8], and the results after computation are proportional to the input size of the tasks. We characterize computation tasks at user A by the tuple (), where (in bits) is the size of the tasks before computation, is the number of required CPU cycles in order to execute each bit, and denotes the data compressed ratio. Let denote the percentage of computation tasks allocated to the DF relay channel, i.e., bits are offloaded to RMES and bits are computed locally. In the AF relay subchannel, user A first processes the task locally and transmits the computational results to the RMES. Hence, the received signal at the relay is
(1) 
where denotes the the transmit signal after local computing, denotes the complex additive white Gaussian noise (AWGN) at relay, and denotes the transmit power of user A. Next, the RMES amplifies the received signal and forwards it to user B. Therefore, the received signal at user B is given by
(2) 
where and denotes AWGN at user B in the AF relay subchannel and the transmit power of RMES allocated to the AF relay subchannel, respectively. According to (2), the rate and delay in AF relay subchannel can be expressed as
(3)  
(4) 
On the other hand, the energy consumption of the AF subchannel is given by
(5) 
Differently from the AF relay subchannel, in the DF relay subchannel user A first offloads the computational tasks to the RMES, and then the RMES decodes the message. Similarly as the AF relay subchannel, we have
(6)  
(7)  
(8) 
where denotes the the transmit signal from user A in DF relay subchannel, denotes the AWGN at the RMES, and denotes the transmit power of user A in the DF relay subchannel.
Then, the RMES executes edge computing and reencodes computational results by using the same or a different codebook and forwards the message to user B. After decoding and retransmission user B receives
(9) 
where denotes the the transmit signal after edge computing, denotes the complex AWGN at destination in the DF relay subchannel, and denotes the transmit power of RMES in the DF relay subchannel. According to (9), the rate and delay from RMES to destination in DF relay channel can be expressed as
(10)  
(11) 
Furthermore, the energy consumption of DF relay subchannel communication is given by
(12) 
As in [8], we model the power consumption of CPU as , where and are the CPU’s computational speed and coefficient depending on chip architecture, respectively. As is equal to cycles per second, the energy consumption per cycle is thus . For local computation, its computation energy consumption can be minimized by optimally configuring computational speed via dynamic voltage and frequency scaling (DVFS) technology [8]. When the amount of data bits processed at user A is , the execution time is
(13) 
where is computational speed of user A. The energy consumption is given by
(14) 
Similarly, the execution time and the energy consumption of edge computation are given by
(15)  
(16) 
where is computational speed of RMES.
Considering both the AF subchannel and the DF subchannel, the total latency due to executing the whole task is given by
(17) 
and the total system energy consumption is expressed as
(18) 
Consequently, the efficient resource allocation problem can be formulated as
(19a)  
s.t.  (19b)  
(19c)  
(19d)  
(19e)  
(19f)  
(19g)  
(19h) 
where denote the set of decision variables . The objective function in is a weighted sum of execution delay and system energy consumption with (in J) as the weighting factor, which could tradeoff the execution delay and the energy consumption. Constraints (19b) and (19c) are the maximum computational speed constraints imposed by user A and RMES CPU, respectively. Constraints (19e), (19f), (19g) and (19h) specify the transmission power budgets at user A and MRES.
Problem is nonconvex and nondifferentiable. In addition, the constraint coulping due to (19h) further complicates problem . The popular CCCP algorithm [11] can be applied to address problem but incurs a very high computational burden because the CCCP algorithm requires solving a sequence of complex convex optimization problems[12]. By exploiting the problem structure, we propose a lightweight iterative algorithm, which is presented in the next section.
3 Proposed Algorithm
Observing that the constraints are separable w.r.t the four block variables, i.e., , , , and , we apply inexact block coordinate descent (BCD) algorithm[13] (a variant of BCD algorithm[14]) to problem P1. This requires the objective function to be differentiable in general. To address the nondifferentiability issue, we first approximate the objective function of P1 as a smooth function using logsmooth method. Specifically, using the logsumexp inequality [9, pp. 72]
we can approximate as
with a large . Hence, the objective of problem can be approximated as a differentiable function, which is given by
(20) 
Now we are ready to use inexact BCD method to solve the smoothed problem, i.e., minimizing subject to (19b–19h). In the BCD method, each time we update one block variable while fixing the others, leading to four subproblems. It can be easily checked that the subproblem w.r.t , , or is convex, and thus all these three subproblems can be easily solved using Bisection method[9]. Therefore, our main efforts are devoted to the update of .
Let us consider the subproblem given by
(21) 
Obviously, (19h) is a nonconvex constraint. This makes it difficult to solve the problem. To efficiently update while decreasing the objective value, we apply the concept of CCCP to tackle the nonconvexity of (19h). First, (19h) can be expressed as a DC program:
(22) 
By linearizing the nonconvex term at the current point , we approximate (22) as a convex constraint
(23) 
As a result, we can approximate problem (21) as
(24) 
Now the constraints are all convex. So we can apply onestep projected gradient (PG) method[14] to problem (24), updating according to
(25)  
(26) 
where can be determined by Armijo rule, denotes the gradient of , denotes the constraint set of problem (24), and denotes the projection of the point onto , which is the optimal solution to problem P2
(27) 
We summarize the proposed BCD algorithm in Table I, where the four block variables are sequentially updated. It can be shown that the algorithm can keep the objective of problem (21) nonincreasing and finally reach a KKT point of problem (21). The proof is omitted due to space limitation.

Next we show how problem can be globally solved using efficient Bisection method. Note that problem is convex. Thus, it can be solved by dealing with its dual problem[9]. To this end, by introducing Lagrange multiplier for the constraint (3), we define the partial Lagrangian associated with problem as
(28) 
Thus, the dual problem of problem can be expressed as
(29) 
where is the dual function given by
(30) 
Note that problem (30) can be decomposed into two independent linearly constrained convex quadratic optimization subproblems w.r.t and , respectively, both of which can be globally solved in closedform. As a result, efficient Bisection method can be applied to problem (31) (and thus ). The details are omitted due to page limitation.
4 Numerical Results
In the simulations, all channel gains are modeled as Rayleigh fading with average power loss . The variance of complex white Gaussian channel noise in all subchannels is set to [4]. The maximum CPU speed of user A and the RMES is 200 MHz and 600 MHz, respectively. Let bits, cycles/bit, and [10]. The maximum transmit power of user A and RMES is set to 1 Watts and 5 Watts. The smoothness factor is set to .
First, we show the convergence performance of the proposed algorithm compared with the CCCP method^{1}^{1}1The CCCPbased algorithm developed for problem (19) is omitted due to space limitation. [11]. Figure 2 shows that, despite a gap between the objective values of (19) and (21) due to smooth approximation (see the green line Vs. the red line), the proposed algorithm can monotonically converge to the same value as that achieved by the CCCP method (see the green line Vs. yellow line). Furthermore, the proposed algorithm can achieve faster convergence than the CCCP method in terms of the number of iterations. In fact, since the CCCP method requires solving a sequence of complex convex problems, it takes s on average for convergence in our simulations while the proposed algorithm takes only s. Hence, the proposed algorithm performs much more efficient than the CCCP method.
Then, the tradeoff between the system energy consumption and the execution delay is shown in Fig. 3 with , , and . It can be observed that the energy consumption increases while the execution delay decreases as increases. Furthermore, when is relatively large, our design focuses more on delay minimization. As a result, when increases to some extent, our algorithm could achieve the minimum execution delay (see the dashed line). Conversely, when is relatively small, our design focuses more on energy consumption minimization and particularly yields the same energy consumption irrespective of the value of . This is because that the minimum energy consumption is achieved when is very small [cf. (16)].
5 Conclusion
This paper have considered joint cooperative computation and interactive communication for achieving the minimum weighted sum of the execution delay and the energy consumption in relay MEC systems. For future investigation, it would be interesting to extend this work for twoway relay MEC systems, enabling simultaneous computational results exchange between two users.
References
 [1] K. Kumar, J. Liu, Y.H. Lu, B. Bhargava, ”A survey of computation offloading for mobile systems”, Mobile Netw. Appl., vol. 18, no. 1, pp. 129140, 2013.
 [2] Y. Mao, J. Zhang, and K. B. Letaief, ”Dynamic computation offloading for mobileedge computing with energy harvesting devices,” IEEE J. Sel. Areas Commun., vol. 34, no. 12, pp. 35903605, Dec. 2016.
 [3] J. Liu, Y. Mao, J. Zhang, and K. B. Letaief, ”Delayoptimal computation task scheduling for mobileedge computing systems,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT), Barcelona, Spain, Jul. 2016, pp. 14511455.
 [4] C. You, K. Huang, H. Chae, B.H. Kim, ”Energyefficient resource allocation for mobileedge computation offloading”, IEEE Trans. Wireless Commun., vol. 16, no. 3, pp. 13971411, Mar. 2017.
 [5] S. Sardellitti, G. Scutari, and S. Barbarossa, ”Joint optimization of radio and computational resources for multicell mobileedge computing,” IEEE Trans. Signal Info. Process. Over Networks, vol. 1, no. 2, pp.89103, Jun. 2015.
 [6] C. Wang, C. Liang, F. Richard Yu, Q. Chen and L. Tang, ”Computation offloading and resource allocation in wireless cellular networks with mobile edge computing,” IEEE Trans. Wireless Commun., vol. 16, no. 8, pp. 49244938, Aug. 2017.
 [7] C. Cao, F. Wang, J. Xu, R. Zhang and S. Cui, ”Joint computation and communication cooperation for mobile edge computing,” [Online]. Available:[arXiv.org/abs/1704.06777].
 [8] Y. Wang, M. Sheng, X. Wang, L. Wang, and J. Li, ”Mobileedge computing: Partial computation offloading using dynamic voltage scaling,” IEEE Trans. Commun., vol. 64, no. 10, pp. 42684282, Oct. 2016.
 [9] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge University Press, 2004.
 [10] Y. Mao, J. Zhang, and K. B. Letaief, ”Joint task offloading scheduling and transmit power allocation for mobileedge computing systems,” in Proc. IEEE Wireless Commun. Networking Conf. (WCNC), San Francisco, CA, Mar. 2017.
 [11] A. L. Yuille and A. Rangarajan, The concaveconvex procedure , Neural Comput., vol. 15, pp. 915936, Apr. 2003.
 [12] M. Zhao, Q. Shi, Y. Cai, M. Zhao, “Joint transceiver design for fullduplex cloud radio access networks with SWIPT,” IEEE Trans. Wireless Commun., vol. 16, no. 9, pp. 56445658, Sept. 2017.
 [13] Q. Shi, H. Sun S. Lu, M. Hong, and M. Razaviyayn, “Inexact block coordinate descent methods for symmetric nonnegative matrix factorization,” IEEE Trans. Signal Process., vol. 65, no. 22, pp. 59956008, Nov. 2017.
 [14] D. Bertsekas, Nonlinear Programming, 2nd ed. Belmont, MA: Athena Scientific, 1999