DeepAI

# Optimal Task Assignment and Power Allocation for NOMA Mobile-Edge Computing Networks

• 23 publications
• 12 publications
• 75 publications
• 42 publications
• 18 publications
• 32 publications
09/11/2020

### Optimal Resource Allocation for Delay Minimization in NOMA-MEC Networks

Multi-access edge computing (MEC) can enhance the computing capability o...
07/17/2018

This paper considers non-orthogonal multiple access (NOMA) assisted mobi...
08/01/2021

### Delay Aware Secure Offloading for NOMA-Assisted Mobile Edge Computing in Internet of Vehicles

In this paper, a multi-vehicle multi-task nonorthogonal multiple access ...
01/14/2019

### On the Performance and Optimization for MEC Networks Using Uplink NOMA

In this paper, we investigate a non-orthogonal multiple access (NOMA) ba...
03/28/2020

Mobile Edge Computing (MEC) enables rich services in close proximity to ...
04/13/2022

### Rate Splitting Multiple Access Aided Mobile Edge Computing in Cognitive Radio Networks

In this paper, we investigate rate splitting multiple access (RSMA) aide...
03/31/2021

Both the mobile edge computing (MEC) based and fog computing (FC) aided ...

## I Introduction

Driven by the explosive emergence of new compute-intensive applications in the Internet of things (IoT), especially ultra-low-latency applications, such as virtual reality (VR) and augmented reality (AR), mobile edge computing (MEC) was proposed to enhance the computing capability of the mobile devices [1]. In order to meet extremely high data rate requirements, non-orthogonal multiple access (NOMA) has been considered as another promising technology in the next generation of wireless communications, due to its superior spectrum efficiency [2, 3, 4]. Specifically, the multiuser superposition transmission scheme, which is a special case of NOMA, has been adopted in the third generation partnership project long term evolution advanced (3GPP-LTE-A) networks [5]. Therefore, to support multiple users and lower the transmission latency and energy consumption, the combination of NOMA and MEC is an inevitable trend in the next generation of wireless networks [6].

### I-a Related literature

A very popular scheme is the power-domain NOMA, where the channel difference between users is exploited to multiplex different users on the same frequency band with different power levels [18, 2]. Thus multiple users can transmit signals simultaneously with lower interference than the orthogonal multiple access (OMA) system. This attracts lots of researchers’ attention in recent years. Resource optimization, i.e. subchannel allocation, power allocation and user association, has been investigated to enhance the performance of NOMA networks [19, 20, 21, 22]. Among these works, the weighted sum rate was maximized for the downlink full-duplex NOMA system via the proposed optimal power allocation and subchannel allocation in [19]. To meet the requirement of green communications, the energy efficient resource allocation was investigated for the perfect channel statement (CSI) [20] and the imperfect CSI [21, 22].

focused on the independent and non-separate task. In this work, an efficient heuristic algorithm of user clustering and frequency and resource blocks allocation was proposed to minimize the system energy consumption

[6]. The energy efficient power allocation, time allocation and task assignment were proposed to minimize the energy consumption for a NOMA MEC network [27]. Besides the computational resource, the successive interference cancellation (SIC) decoding order was optimized to reduce the task delay for NOMA enabled narrowband Internet of Things (NB-IoT) systems [29].

### I-B Contributions

Different from the existing works, which mainly focus on energy minimization for NOMA multiuser MEC networks, we focus on the completion time minimization by considering the partial offloading in NOMA multiuser MEC. The main contributions are listed as follows:

• In this paper, we apply NOMA into a multiuser MEC network where multiple users can offloading their task to the MEC server simultaneously via the same frequency band. Considering the partial task offloading scheme, the energy consumption limitation and offloading power limitation, the task completion time minimization problem is formulated as a non-convex problem. Thus it is polynomial time unsolvable. By analysing the properties of the formulated problem, some significant insights are revealed, and the corresponding propositions and Lemma are proposed to equivalently transform the original formulated problem to a simplified form. Based on those analytical results, the quasi-convexity of the transformed problem is proved. Therefore, a bisection searching (BSS) based algorithm is proposed to find the global optimal solution to the transformed problem. In the proposed algorithm, the original problem is solved by equivalently solving a series of feasibility subproblems. Moreover, we analyze the complexity of the proposed algorithm. The convergence and optimality of BSS algorithm are evaluated by simulation results.

• Motivated by the practical applications, we focus on the two-user case to reduce the decoding complexity of SIC. To further reduce the complexity of the proposed BSS iterative algorithm, a corresponding proposition is proposed to equivalently transform the original problem to a convex problem. The convexity of the transformed problem is proved. Moreover, the closed-form expressions of the task partition ratios and offloading power are derived by exploiting the Lagrangian approach, i.e., Karush-Kuhn-Tucker (KKT) conditions, for the two-user NOMA MEC network. The simulation results demonstrate the proposed BSS iterative algorithm matches with our derived optimal solution, which reveals that our proposed BSS algorithm can converge to the optimal solution.

### I-C Organization

The organization of this paper is as follows. We introduce the system model and formulate the task completion time minimization problem in Section II. The BSS algorithm is proposed in Section III. In Section IV, the optimal closed-form solution derived for two-user case. Simulation results are presented in Section IV, and Section V concludes this paper.

## Ii System Model and Problem Formulation

### Ii-a NOMA-enabled Multiuser MEC Networks

In the NOMA multiuser MEC network, users are randomly distributed in the single cell where one BS equipped with the MEC server is located in the cell centre. We assume that all users and the BS are equipped with single antenna. The indices of users are defined as . Denote the offloading task partial factor of User () by , thus is the task partial factor of the task to locally compute at , where . By implementing NOMA, the SIC technique is applied at the MEC based BS. Define as the channel gain from to the MEC server. Without loss of generality, users are sorted as . In NOMA offloading transmission, the SIC decoding order is assumed as the decreasing order of channel gains. It indicates that the MEC server first decodes the information transmitted by and then decodes the information of , until . Define as the offloading power of , then the signal-to-interference-plus-noise-ratio (SINR) of the received at the MEC server can be written by

 Γoffm=|hm|2pmm−1∑j=1|hj|2pj+σ2. (1)

where represents zero-mean complex additive white Gaussian noise (AWGN) power. Denote the bandwidth of this system by Hz, thus the achievable data rate of the can be written by

 Rm= Blog2(1+Γoffm)=Blog2⎛⎜ ⎜ ⎜ ⎜ ⎜⎝m∑i=1|hi|2pi+σ2m−1∑j=1|hj|2pj+σ2⎞⎟ ⎟ ⎟ ⎟ ⎟⎠. (2)

In this phase, each user will offload part of its task to the MEC server for remote executions. According the achievable offloading data rate (2) of the , the task offloading time from to the MEC server can be written by

 Toffm=βmLmRm. (3)

 Eoffm=Toffmpm. (4)

#### Ii-A2 Mobile Execution Time

For , partial task is offloaded to the MEC server for remote computation, and the remaining task is computed locally. Denote the CPU frequency at is (in cycles per second). The local computation time of is given by

 Tlocm=(1−βm)LmCmflocm. (5)

Thus the energy consumption of computing at the mobile device can be written by

 Ecm=κm(1−βm)LmCm(flocm)2 (6)

where denotes the effective capacitance coefficient for each CPU cycle of the local user .

### Ii-B Problem Formulation

 Tm=max{Toffm, Tlocm}. (7)

Therefore, the task completion time minimization problem can mathematically formulated as

 min max{β,p} {Toffm,Tlocm,∀m} (8a) s.t. 0≤βm≤1,∀m, (8b) 0≤pm≤Pmax,∀m, (8c) Ecm+Eoffm≤Emax,∀m, (8d)

where and . Constraint (8b) specifies the range of the offloading task ratio. Constraint (8c) describes the range of the offloading transmit power. Constraint (8d) guarantees that the energy consumed at each mobile user is limited to the maximum energy consumption . Since problem (8) is nonconvex, we propose the following transformation and solution to solve this problem.

## Iii The Optimal Solution for the Multiuser case

### Iii-a Significant Observations

In this section, we consider the multiuser case and attain the minimum computation time for NOMA multiuser MEC networks. An bisection searching based iterative algorithm is proposed to find the global solution to problem (8). Before solving this problem, we propose the following Lemma and proposition to simplify the optimization problem (8).

###### Lemma 1

In NOMA multiuser MEC networks, if users offload their signals to the BS equipped with MEC server within the same transmit time ,

 T=Toffm=Toffm′, ∀m≠m′. (9)

Then (9) can be equivalently transformed to

 ~Toffm=m∑i=1βiLiBlog2(m∑i=1|hi|2pi+σ2),∀m. (10)
###### Proof.

The proof is shown in Appendix A. ∎

According to Lemma 1, we propose the following proposition to further simplify problem (8).

###### Proposition 1

For any two users ( and ) in NOMA multiuser MEC networks, to minimize the maximum task completion time of different users, i.e.,

 min max{βm,βm′,pm,pm′}{Toffm, Toffm′}, (11)

the optimal solution can be only obtained if and only if the offloading time of each user equals to each other .

###### Proof.

The proof is shown in Appendix B. ∎

### Iii-B BSS Iterative Algorithm

Based on Lamma 1 and Proposition 1, we have . The energy consumed at can be written as

 Em=κmβmLm(flocm)2+~Toffmpm,∀m. (12)

Thus, the task completion minimization problem for NOMA multiuser MEC can be reformulated as:

 minmax{β,p} {~Toffm,Tlocm,∀m} (13a) s.t. βm∈[0,1],∀m, (13b) 0≤pm≤Pmax,∀m, (13c) Em≤Emax,∀m. (13d)

This problem is non-convex due to the nonconvexity of with respect to . Thus it is challenging to obtain its global optimum within polynomial time. However, after investigating properties of this problem, we first propose Proposition 2 to demonstrate the quasi-convexity of the objective function in (13), which can be solved by a series of convex feasibility problems.

###### Proposition 2

Given by the expressions (5) and (11), the objective function in problem (13) is strictly quasi-convex.

###### Proof.

The proof is shown in Appendix C. ∎

To obtain the optimal solution, problem (13) can be equivalently transformed to the following problem:

 min{β,p,αT} αT (14a) s.t. m∑i=1βiLilog2(1+M∑i=1|hi|2pi)≤αT, ∀m (14b) (1−βm)LmCmflocm≤αT, ∀m (14c) 0≤βm≤1, ∀m (14d) 0≤pm≤Pmax, ∀m (14e) κm(1−βm)LmCm(flocm)2+αTpm≤Emax, ∀m. (14f)

Note that the inequality constraint set is not convex set in due to the quasi-convexity of (14b) and non-convexity of (14f). However, problem (14) becomes a feasibility problem if we fix . Thus we can solve the above problem by solving a serious of convex feasibility subproblems. For a given , the feasibility problem can be formulated as:

 find {β,p} (15a) s.t. (15b)

Note that the convex constraint set can be denoted as

 CαT={{β,p}|(???)−(???)},∀αT. (16)

Define the optimal solution of this feasibility problem is . Bisection search method can be utilized to find [31]. Therefore, Algorithm 1 is proposed to find the minimum task completion time of user tasks. In Algorithm 1, we first initialize by its lower bound and upper bound. For given , we say that problem (15) is feasible , and we have . Problem (15) is infeasible , and we have . Algorithm 1 converges to the unique global optimal solution to problem due to its strictly quasi-convexity [32].

The computational complexity of Alg. 1 mainly comes from the BSS to find the optimal . For given accuracy , and , the computational complexity of Alg. 1 is given by .

## Iv Closed-Form Optimal Solution Derivation for the Two-user case

To further reduce the complexity of the proposed iterative algorithm, in this section, we derive the closed-form solution for two-user case based on the insights and propositions obtained from problem (13). To further simplify the problem, the following proposition can be obtained.

###### Proposition 3

 min max{βm,pm}{Toffm, Tlocm} ∀m, (17)

the optimal solution can be only obtained if and only if its offloading time equals to its local computing time, i.e., .

###### Proof.

Proof by contradiction is shown in Appendix D. ∎

Based on Proposition 3, we conclude that the optimal solution to problem (13) can be obtained when for each user. According to Proposition 1, the optimal solution can only be obtained when the offloading time equals to each other. Considering two-user case , problem (13) can be rewritten by

 min{β1,β2,p1,p2} β1L1+β2L2log2(1+|h1|2p1+|h2|2p2) (18a) s.t. β1∈[0,1],β2∈[0,1], (18b) 0≤p1≤Pmax,0≤p2≤Pmax, (18c) κ1(1−β1)L1C1(floc1)2 +β1L1+β2L2log2(1+|h1|2p1+|h2|2p2)p1≤Emax, (18d) κ2(1−β2)L2C2(floc2)2 +β1L1+β2L2log2(1+|h1|2p1+|h2|2p2)p2≤Emax (18e) (1−β1)L1C1floc1=β1L1+β2L2log2(1+|h1|2p1+|h2|2p2) (18f) (1−β2)L2C2floc2=β1L1+β2L2log2(1+|h1|2p1+|h2|2p2) (18g) β1L1log2(1+|h1|2p1)=β1L1+β2L2log2(1+|h1|2p1+|h2|2p2). (18h)

In problem (18), the objective function in (18) is quasiconvex based on Proposition 2. However, the constraint (18d) and (18e) are not convex set with respective to . To simplify this problem, we first deal with equality constraints (18f)-(18h). To solve the above problem and obtain the global optimum, we first equally transform this problem to an equivalent convex form via equality constraints. By using the equation (18h), we can replace the right sides of (18f) and (18g) with the left side of (18h). Then we have

 (1−β1)L1C1=β1L1log2(1+|h1|2p1)floc1 (19a) (1−β1)L1C1/floc1=(1−β2)L2C2/floc2 (19b) (1−β1)L1C1/floc1=β1L1+β2L2log2(1+|h1|2p1+|h2|2p2). (19c)

After a series of calculations, the objective function in (18) can be rewritten by

 L1+L2floc1C1+floc2C2+Blog2(1+|h1|2p1+|h2|2p2). (20)

The transformation can be found in Appendix E. Therefore, problem (18) can be rewritten by

 min{p1,p2} L1+L2floc1C1+floc2C2+Blog2(1+|h1|2p1+|h2|2p2) (21a) s.t. (21b)
###### Proposition 4

Problem (21) is convex problem.

###### Proof.

The convexity proof is shown in Appendix F. ∎

According to equations (18f)-(18h), once the optimal and are obtained, the optimal and can be calculated by the following expressions.

 β∗1=log2(1+|h1|2p∗1)floc1C1+log2(1+|h1|2p∗1) (22a) β∗2=1−(1−β∗1)L1C1floc2L2C2floc1 (22b)

In the following, we focus on deriving the optimal closed-form expressions of and . Since problem (21) is convex and satisfy Slater’s condition, the KKT conditions can be exploited to derive the closed-form optimal solution. The optimal solution can be concluded obtained by following four cases.

Case 1: When

 {P1,w(p∗2=Pmax)≥PmaxP2,w(p∗1=Pmax)≥Pmax (23)

Define

 P1,w(p∗2)=−W0⎛⎜⎝−B1log(2)|h1|22(−B1|h1|2+A1)⎞⎟⎠B1log(2)−1+|h2|2p∗2|h1|2, (24)

where , , , and where is Lambert function, which is single value function.

 P2,w(p∗1)=−W0⎛⎜⎝−B2log(2)|h2|22(−B2|h2|2+A2)⎞⎟⎠B2log(2)−1+|h1|2p∗1|h2|2 (25)

where and . Thus we have

 {p∗1=Pmaxp∗2=Pmax. (26)

Case 2: When

 {P1,w(p∗2=Pmax)≥PmaxP2,w(p∗1=Pmax)≤Pmax, (27)

thus we have

 {p∗1=Pmaxp∗2=P2,w(p∗1=Pmax). (28)

Case 3: When

 {P1,w(p∗2=Pmax)≤PmaxP2,w(p∗1=P1,w)≥Pmax, (29)

thus we have

 {p∗1=Pmaxp∗2=P2,w(p∗1=Pmax). (30)

Case 4: When

 {P1,w(p∗2)≤PmaxP2,w(p∗1)≤Pmax, (31)

thus we have

 ⎧⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪⎨⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪⎩p∗1=κ2(floc2)3−κ1(floc1)3+p∗2p∗2=−W0⎛⎜⎝−B2log(2)(|h1|2+|h2|2)2(−B2(|h1|2+|h2|2)+A2)⎞⎟⎠B2log(2)−1+|h1|2(κ2(floc2)3−κ1(floc1)3)(|h1|2+|h2|2). (32)

where and .

###### Proof.

The derivation is provided in Appendix G. ∎

From the above solution, we can directly obtain the optimal solution for two-user case NOMA MEC and the minimum task completion time can be obtained. The complexity is largely reduced compared with the proposed Algorithm 1.

## V Simulation Results

In this section, the performance of our proposed resource allocation schemes is evaluated by simulation results. In the simulation results, our proposed BSS algorithm and the optimal analysed solution are compared with three benchmark schemes, i.e., orthogonal frequency-division multiple access (OFDMA) based partial scheme, NOMA based full offloading scheme and fully local computing scheme. In the simulation setting up, the users are randomly distributed in a single cell with the radius of 500 m. The channel gain from the -th user to the BS is denoted by in which is a Rayleigh fading channel coefficient, and is the distance from this user to the BS. The path loss factor is 3.76. The AWGN power is where the AWGN spectral density is dBm/Hz. For the computational resource at each user, we set cycles/bit.

In Fig. 1, we provide the convergence and optimality of the proposed BSS algorithm. In figure 1 (a), we set the bandwidth to MHz. The length of computation tasks for each user is set to bits. The maximum energy consumption for each user is Joule, and the maximum power for each user is W, and the effective capacitance coefficient for each CUP cycle of the local users are . In Fig. 1 (b), we set the bandwidth to Hz, and the length of computation tasks for each user is set to bits. Since the offloading data rate is small due to small bandwidth, most the tasks are computed by local users. Therefore, the minimum completion time in BSS algorithm keeps increasing by each iteration until its optimal. From this figure, we can see that our proposed BSS algorithm converges within 10 iterations. The convergence point is perfectly matched with the optimal analytical solution, which is derived by Lagrangian approach.

Figure 5 presents the convergence BSS algorithm with multiuser case . In this figure, we set MHz and bits, Joule, W, and . From this figure, we can see that our proposed BSS algorithm converges within 10 iterations with different user numbers. The scheme with 8 users has the highest completion time than the schemes with 3 and 8 users.

## Vi Conclusions

In this paper, we have investigated the optimal task partition and power allocation to minimize task completion time for NOMA multiuser MEC networks. The optimization problem has been formulated as quasi-convex problem. Then BSS algorithm has been proposed to achieve the minimum task completion time, which is the global optimum. To further reduce the complexity of the proposed algorithm, the closed-form optimal power allocation and offloading task ratio expressions have been derived for two-user case via Lagrangian approach based on analytic insights obtained from the analysis. Simulation results have demonstrated the convergence and optimality the proposed schemes, which can provide an effective solution to minimize task completion time for NOMA multiuser MEC networks.

## Appendix A Proof of Lemma 1

We assume that users transmit their tasks in the same period , which indicates:

 T=β1L1R1=β2L2R2=⋯=βMLMRM. (33)

Since can be written by where . Thus . Therefore,

 T=β1L1+β2L2+⋯+βmLmR1+R2,n+⋯+Rm=m∑i=1βiLim∑i=1Rm,∀m. (34)

 m∑i=1Ri= Blog2(σ2+|h1|2p1)+Blog2(σ2+|h1|2p1+|h2|2p2σ2+|h1|2p1)+⋯ (35) +Blog2⎛⎜ ⎜ ⎜ ⎜⎝σ2+m∑i=1|hi|2piσ2+m−1∑i=1|hi|2pi⎞⎟ ⎟ ⎟ ⎟⎠=Blog2(σ2+m∑i=1|hi|2pi).

Define the number of transmitted bits as , and . Then the offloading time can be written by

 T=m∑i=1βiLiBlog2(σ2+m∑i=1|hi|2pi),∀m. (36)

Now let us prove from (36) to (33). When , we have . When , we have . Since , we can have . By deduction, we can have (33). We finish the proof of Lamma 1.

## Appendix B Proof of Proposition 1

According to Lemma 1, the offloading time minimization problem of different users can be represented by

 minmax{β,p}{Toff1,Toff2,⋯,ToffM}. (37)

Note that the minimum latency is with the optimal solution . This optimal solution is only obtained when . Proof by contradiction can be exploited to prove this proposition.

Assume that decodes its signal firstly, and the optimal solution is obtained when . Thus the minimum latency is with the optimal solution . In this case, if we increase to , then will be decreased. Since is fixed, and will be increased. Therefore, there must exist satisfying . Therefore, should be the optimal time consumption since it has lower value than . This contradicts the assumption that is the optimal solution to problem (37).

Assume that the optimal solution is obtained when . Thus the minimum latency is with the optimal solution . In this case, if we increase to , then will be decreased. Since is fixed, and will be increased. Therefore, there must exist satisfying