Multi-Resource Computing Offload Strategy for Energy Consumption Optimization in Mobile Edge Computing

Abstract

The energy consumption optimization of edge devices in the mobile edge computing environment is mainly based on computational offload strategy. Most of the current common computing offload strategies only consider a single computing resource and do not comprehensively consider different kinds of computing resources in mobile edge computing environments, which cannot fully reduce the energy consumption of edge devices under the condition of ensuring response time constraints. To solve this problem, a multi-resource computing unloading energy consumption model is proposed in the mobile edge computing environment, and a new fitness calculation method for evaluating the energy consumption of edge devices is designed. Combined with the workflow management system, a multi-resource computing offloading particle swarm optimization task scheduling algorithm for energy consumption optimization in mobile edge computing is proposed. The algorithm can fully reduce the energy consumption of mobile terminals under the condition of considering the response time constraint. Experiments show that, compared with the existing four algorithms, the task scheduling algorithm corresponding to the new strategy has stable convergence and optimal fitness. Under the constraint of user response time, the energy consumption of edge devices in the task scheduling scheme is better than the other four unloading strategies.

1. Introduction

With the rapid development of IoT technology and mobile communication technology, intelligent mobile terminals such as smartphones, tablet PCs, and smart home devices have entered millions of households, greatly improving people’s lives. At the same time, the massive increase in mobile terminals will inevitably lead to a dramatic increase in mobile traffic [1]. According to the latest forecast report released by Cisco, by the end of 2021, global mobile data traffic will reach 49 Exabyte/m, and mobile traffic will account for 84% of the total traffic in the network [2]. With the development of various information services and terminal applications, users have increasingly stringent requirements for network service quality and network request delay. At present, the information processing capacity of the central processing unit of mobile devices has been greatly improved, but it is still insufficient in the face of processing tasks with high time delay requirements and a huge amount of data [3,4]. In addition, limited battery capacity and storage space are also important factors restricting the processing of computing-intensive and time-sensitive applications by mobile devices [5,6]. The traditional method is to upload to a cloud server, or cloud computing. However, with the explosive growth of user data, this method reveals some shortcomings with problems of delay due to transmission bandwidth, data processing, physical distance, and higher transmission costs [7,8]. For these reasons, mobile edge computing came into being. It is a new computing model. However, mobile edge computing does not replace cloud computing but supplements and extends cloud computing [9]. The task scheduling algorithm in the workflow management system can efficiently manage and allocate a variety of virtual machine resources, making the workflow task execution more efficient.

The proliferation of terminal devices will inevitably lead to the explosive generation of data. Although the computing power and storage capacity of mobile devices are constantly improving, these terminal devices are incompetent in the face of new intensive tasks [10,11]. For these intensive tasks, cloud computing is proposed as a centralized computing model to offload local tasks to the cloud center for processing, and the powerful computing and storage capacity of the cloud center is a good solution to the lack of resources at the terminal, but because the cloud server is located at the core network layer, geographically far from the terminal and users, when the tasks are offloaded to the cloud server, the task data must flow through the entire access network and core network. When the task is offloaded to the cloud server, the task data must flow through the entire access network and core network and be transmitted through multiple base stations and other network devices, and the calculation results are also transmitted over long distances before being returned to the terminal devices, which brings additional transmission delay and also easily leads to data leakage and other security problems. As a result, this cloud-centric architectural solution struggles to meet the demands of low-latency, high-reliability, and high-security [11,12,13]. Mobile Edge Computing (MEC) is a new computing model that extends the services of cloud computing to the edge of the network by deploying lightweight cloud servers at the edge of the network, which can provide real-time and elastic computing supply to users [10,14,15]. With MEC technology, mobile end-users can offload computing tasks from local standalone machines to a Mobile Edge Computing Server (MECS) anytime, anywhere, and flexibly, enabling user devices to run demanding applications while meeting low latency requirements [16,17].

Different from cloud computing, the computing tasks of terminal devices in MEC are unloaded to the edge server closer to the terminal device, and the edge server can provide computing, content caching, and other functions [18].Servers distributed on the edge of the network (also known as computing nodes and edge nodes) can reduce the computing pressure of terminal devices and reduce the frequency of interaction with the centralized data center of cloud computing; it can also significantly reduce the waiting time in the message exchange. Because the edge server has certain storage space and computing power and is closer to the terminal device, the computing-intensive or delay-sensitive mobile terminal device can unload the computing task to the edge server for computing.

Mobile edge computing solves the problem of insufficient computing power at user terminals to a certain extent, but the limitations of MEC servers are becoming increasingly evident for emerging applications with latency-sensitive, compute-intensive, and huge energy consumption [16]. The computing power of user terminal devices is limited, and they cannot process a large number of computing tasks in a short period. Although the computing power and storage capacity of current mobile devices are increasing, these end devices are not up to the task in the face of new types of intensive tasks such as virtual reality, natural language processing, and interactive games, so they need to be combined with devices with greater computing and storage power to perform tasks in a collaborative manner [12,19,20]. To meet the low latency and low energy consumption requirements of user terminal equipment for intensive task processing in the edge computing field, the latency and energy consumption of user terminal equipment in processing computing tasks can be reduced by unloading computing tasks to the edge server [21]. Therefore, it is of practical importance to design a reasonable computational offloading strategy to meet user requirements in edge computing systems. Computational offloading has been studied by many scholars at home and abroad as an effective means to save device energy and reduce latency in edge computing networks.

Woodward, L. [22] et al.’s present study conceptually shifts an energy-efficient DPFSP to a sustainable DPFSP, and a multi-objective optimization model is developed to approach a sustainable TBL-based DPFSP. Finally, sensitivity analyses are performed to assess the efficiency of the proposed model. In Deming Lei [23] et al., a mixed-integer mathematical programming model is proposed to minimize the cycle time and environmental cost, while a metaheuristic approach based on a fruit fly optimization algorithm (FOA) is developed to find a fuzzy disassembly scheduling scheme. Finally, experiments are conducted to compare with other multi-objective optimization algorithms. The computational results demonstrate that the proposed algorithm outperforms other algorithms in computational efficiency and applicability to different problems. As shown in

Table 1. Research of related papers.

Main Types of Offloading:	Authors	Methodologies	To Compare
No offloading	Simon, W.A. [24]	Introduces BLADE, a BitLine Accelerator for Devices on the Edge.	Compared with no offloading and complete offloading, the method adopted in this paper is partial offloading. We propose a multi-resource computing offload task scheduling algorithm for energy consumption optimization. This method is suitable for the complex situation of tasks and data.
	S. Bi, Y.J. Zhang [25]	A wireless-powered multi-user MEC system is considered, where each device either computes its task locally or offloads it entirely.
Complete offloading
	Salmani, M., Davidson, T.N. [26]	Provides a closed-form optimal solution to the energy minimization problem when a set of users with different latency constraints are completely offloading their computational tasks and a tailored greedy search algorithm for a good set of users.
Partial offloading	Kai Guo, Ruiling Zhang [27]	A fairness-oriented approach, including the determination of the application offloading strategy, the data transmission strategy, and the cloud–edge.

, relevant papers are studied.

There are three main types of offloading: no offloading, complete offloading, and partial offloading, as shown in Figure 1. This strategy can be applied to a situation where the amount of computation is small or the mobile edge computing resources are not available. In the complete offload mode, all tasks in the edge device are unloaded to the edge or cloud server. This strategy is suitable for scenarios with a small amount of data transmission between tasks. Part of the unloading method is to unload some of the computing tasks that meet the conditions to the edge or cloud server, while the remaining tasks are performed on the edge devices. This method is suitable for the complex situation of tasks and data.

However, at present, the use of a computing offload strategy can optimize the energy consumption of terminal devices in the mobile edge computing environment. Xu [28] et al., proposed a global cost model that takes time delay and energy consumption into account and formulated it into an optimization problem. Huang [29] et al., investigated low-complexity computation offloading policies to guarantee the quality of service of the MEC network and minimize WDs’ energy consumption and further proposed a heterogeneous DDLO to achieve better convergence performance than DDLO. Extensive numerical results show that the DDLO algorithms guarantee better performance than the LR-based algorithm. Mao [30] et al., investigated a green mobile edge computing system with energy harvesting devices and developed an effective computation offloading strategy using the complete unloading strategy. The paper [30,31] proposed a task execution energy optimization algorithm for a partial offload strategy.

However, there are many types of computing resources in the mobile edge computing environment, such as cloud resources, edge server resources, and terminal resources. The traditional computing soft offloading strategy usually only considers the above resources and does not consider the multiple computing resources in the mobile edge computing environment. The multi-resource computing offload task scheduling algorithm is more efficient than traditional methods.

Therefore, considering the types of resources in the mobile edge computing environment, we explore the computing offload strategy for energy consumption optimization in mobile computing. We consider the types of resources in the mobile edge computing networks and design a fitness calculation method suitable for evaluating the execution of energy consumption of task scheduling schemes in the mobile edge computing environment and finally, propose a multi-resource computing offload task scheduling algorithm for energy consumption optimization. The main contributions in this paper are summarized as follows.

• We propose a multi-computing unloading energy consumption model based on energy consumption optimization.
• We design a fitness calculation method suitable for evaluating the execution of energy consumption of task scheduling schemes in the mobile edge computing environment.
• We propose a multi-resource computing offload task scheduling algorithm for energy consumption optimization, which can be integrated considering a load of edge devices, edge servers, and stone data centers during unloading. The unloading mode is selected independently to reduce the energy consumption of mobile terminals under the premise of ensuring the user response time constraint. The experimental results show that the energy consumption of the task scheduling scheme obtained by this algorithm is the lowest, and the convergence of the algorithm is stable.

The rest of the paper is organized as follows. We first give the energy consumption and completion times for different offloading strategies for completing a workflow task in Section 2. Then, we give the multiple resource computation offloading energy consumption model in Section 3. After that, a multiple resource offloading particle swarm task scheduling algorithm based on energy consumption optimization is presented in Section IV. Next, experimental simulations are given, and the results are analyzed in Section 5. Finally, we conclude in Section 6.

2. Energy Consumption and Completion Time of Scheduling Schemes

This section first gives a workflow example and then compares the energy consumption and completion time of four scheduling schemes using different self-unloading strategies to illustrate the necessity of finding the task scheduling scheme with the optimal energy consumption under the condition of guaranteeing the user response time constraint.

2.1. Working Examples

We can visually display the priority dependency of all tasks in the workflow. A simple workflow is shown in Figure 2. There are six tasks in the workflow. The workload and data volume of each task are shown in

Table 2. Task load and data volume.

Task Number	1	2	3	4	5	6
Load	500	1000	1500	2000	2500	550
The amount of data/KB	1200	1200	700	850	60	70

. When task 1 runs, tasks 2 and 3 can start to run, while tasks 4 and 5 need to wait for tasks 2 and 3, respectively, to start running. Task 6 cannot be started until task 4 and task 5 are finished. When task 6 is finished, the whole workflow is finished.

2.2. Comparison of Scheduling Algorithms with Different Unloading Strategies

It is assumed that the operating speed of the edge server virtual machine, cloud data center server virtual machine, and edge device is 1.0 GHz, 1.6 GHz, and 0.7 GHz, respectively. The operating power, transmission power, and idle power of edge devices are 0.7 W, 0.1 W, and 0.0 W, respectively. The data transmission bandwidth is 20 Mbps, and the response time constraint is 5 s. The optimal task scheduling scheme corresponding to the four unloading strategies is shown in Figure 3.

In Figure 3, the energy consumption of the scheduling schemes generated by non-unloading, full unloading, partial unloading, and multi-resource unloading scheduling algorithms are 12.2 j, 0.474 j, 0.66 j, and 0.729 j, and the completion time is 17.5 s, 9.47 s, 7.63 s, and 4.8 s, respectively. From the results of energy consumption and completion time, it can be seen that the task execution energy consumption is the highest without unloading, and the energy consumption is the lowest in the case of complete unloading, but the task completion time is long and cannot meet the response time constraint. At the same time, the energy consumption values of partial unloading and multi-resource unloading are not significantly different, but the scheduling scheme generated by the multi-resource unloading scheduling algorithm can optimize the task execution energy consumption under the condition of meeting the response time constraint. To sum up, the scheduling algorithm based on multiple resources offloading can find the task scheduling scheme with optimal energy consumption under the constraint of user response time.

3. Energy Consumption Model of Multi-Resource Calculation Unloading

In the mobile edge computing environment, the existing computing offload strategy only considers a single computing resource and does not comprehensively consider different kinds of computing resources in the mobile edge computing environment, resulting in the limited energy consumption level of the task scheduling algorithm corresponding to the energy consumption model. Therefore, according to the reference, this paper improves the traditional unloading energy consumption model and proposes a multi-resource unloading energy consumption model. The model is divided into three parts:

• The energy consumption model of the edge device when the task is not unloaded;
• The energy consumption model is when the task is unloaded to the edge server;
• The energy consumption model is when the task is uploaded to the cloud server.

3.1. Edge Device Task Energy Consumption Model

When the task T_i is not unloaded, T_i is placed on the edge device for execution. At this time, we need to consider the energy consumption generated by the execution of task T at the edge device. Firstly, the execution time of the task on the edge device is calculated by the task load, and then the energy consumption of the task running on the edge device is calculated by the task execution time. The execution time of tasks on the edge device is calculated according to different task loads:

$${\mathrm{T}}_{\mathrm{edge}}={\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}\frac{{\mathrm{l}}_{\mathrm{i}}}{{\mathrm{f}}_{\mathrm{edge}}}}$$

(1)

where l_i is the task load of task I, f_edge is the task processing speed of the edge device.Tedge is the sum of the execution time of all tasks in the edge device.

According to the sum of the task execution time of the edge device, the execution energy consumption of the task on the edge device is calculated:

$${\mathrm{E}}_{\mathrm{edge}}{=\mathrm{P}}_{\mathrm{edge}}\times {\mathrm{T}}_{\mathrm{edge}}$$

(2)

where E_dge is the energy consumption of task execution at the edge device, P_edge is the Power when performing tasks for edge devices.

3.2. Edge Server Task Energy Consumption Model

When the task (${\mathrm{T}}_{\mathrm{i}}$) is unloaded to the edge server, the edge device has no execution energy consumption. Firstly, we need to consider the energy consumption of data transmission when the task is unloaded to the edge server and the energy consumption of receiving data returned to the edge device after the task is executed in the edge server. Finally, the idle energy consumption of the edge device when the task is unloaded to the edge server should be considered. Therefore, the energy consumption calculation formula of tasks unloaded to the edge server is as follows:

$${\mathrm{E}}_{\mathrm{ec}}{=\mathrm{P}}_{\mathrm{tr}}\times \frac{{\mathrm{c}}_{\mathrm{j}}}{{\mathrm{R}}_{\mathrm{l}}}{+\mathrm{P}}_{\mathrm{re}}\times \frac{d_j}{{\mathrm{R}}_2}{+\mathrm{P}}_{\mathrm{idle}}\times {\mathrm{T}}_{\mathrm{ec}}$$

(3)

The energy consumption of task unloading to the edge server is divided into three parts: data sending energy consumption, data receiving energy consumption, and edge device idle energy consumption. Where: (${\mathrm{P}}_{\mathrm{tr}}$) is the edge device data transmission power; (${\mathrm{C}}_{\mathrm{j}}$) is the task (${\mathrm{T}}_{\mathrm{j}}$), the size of the data sent; (${\mathrm{R}}_1$) is the data transmission rate in the mobile edge computing environment; (${\mathrm{P}}_{\mathrm{re}}$) is the edge device data receiving power; ($d_{\mathrm{j}}$)is the task (${\mathrm{T}}_{\mathrm{j}}$); (${\mathrm{R}}_2$) is the data receiving rate in the mobile edge computing environment; (middle) is the idle power of the edge device; (${\mathrm{T}}_{\mathrm{ec}}$) is the execution time of the task unloaded to the edge server,

$${\mathrm{T}}_{\mathrm{ec}}={\displaystyle \sum _{\mathrm{j}=1}^{\mathrm{n}}\frac{{\mathrm{l}}_{\mathrm{j}}}{{\mathrm{f}}_{\mathrm{ec}}}}$$

(4)

where (${\mathrm{l}}_{\mathrm{j}}$) is the task load of task (j), (${\mathrm{f}}_{\mathrm{ec}}$)is the running speed of the edge server, and (${\mathrm{T}}_{\mathrm{ec}}$) is the execution time of all tasks unloaded to the edge server.

3.3. Cloud Server Task Energy Consumption Model

When the task (${\mathrm{T}}_{\mathrm{k}}$) is uploaded to the cloud server, there is no task execution energy consumption in the edge device. Like offloading to the edge server, the task (${\mathrm{T}}_{\mathrm{k}}$) unloaded to the cloud server also needs to consider the energy consumption of data sending, data receiving, and edge device idle.

Therefore, the calculation formula of energy consumption for task execution unloaded to E_CC is as follows:

$${\mathrm{E}}_{\mathrm{cc}}{=\mathrm{P}}_{\mathrm{tr}}\times \frac{{\mathrm{c}}_{\mathrm{k}}}{{\mathrm{R}}_{\mathrm{l}}}{+\mathrm{P}}_{\mathrm{re}}\times \frac{d_{\mathrm{k}}}{{\mathrm{R}}_2}{+\mathrm{P}}_{\mathrm{idle}}\times {\mathrm{T}}_{\mathrm{cc}}$$

(5)

where (${\mathrm{P}}_{\mathrm{tr}}$)is the data transmission power of the edge device; (${\mathrm{c}}_{\mathrm{k}}$) is the data size sent by the task (${\mathrm{T}}_{\mathrm{cc}}$); (${\mathrm{R}}_1$) is the data transmission rate in the mobile edge computing environment; (${\mathrm{P}}_{\mathrm{tr}}$) is the data receiving the power of the edge device; ($d_{\mathrm{k}}$) is the received data size of the task (${\mathrm{T}}_{\mathrm{cc}}$); (${\mathrm{R}}_2$) is the data receiving rate in the mobile edge computing environment; (middle) is the idle power (${\mathrm{T}}_{\mathrm{cc}}$) of the edge device and is the execution time of the task unloaded to the cloud server,

$${\mathrm{T}}_{\mathrm{ec}}={\displaystyle \sum _{\mathrm{j}=1}^{\mathrm{n}}\frac{{\mathrm{l}}_{\mathrm{j}}}{{\mathrm{f}}_{\mathrm{ec}}}}$$

(6)

where (${\mathrm{l}}_{\mathrm{j}}$) is the task load of task (${\mathrm{T}}_{\mathrm{ec}}$) (${\mathrm{f}}_{\mathrm{ec}}$) is the running speed of ECs, (${\mathrm{T}}_{\mathrm{cc}}$) is the execution time of all tasks unloaded to ECs.

4. Particle Swarm Optimization Task Scheduling Algorithm Based on Energy Consumption Optimization

In this section, based on the energy consumption model of multi-resource unloading in the mobile edge computing environment, a new fitness calculation method to evaluate the energy consumption of edge devices is designed. Then, combined with the workflow management system, a multi-resource computing offloading particle swarm task scheduling algorithm is proposed to solve the problem of energy consumption optimization of edge devices in the mobile edge computing environment.

4.1. Fitness of Particle Swarm Optimization

In the multi-resource computing unloading particle swarm optimization (PSO) algorithm for energy consumption optimization, the fitness can comprehensively evaluate the execution time and cost of task scheduling schemes. It represents the physical state of particles corresponding to each task scheduling scheme in the particle swarm optimization algorithm. Based on the particle swarm optimization (PSO) scheduling algorithm in the cloud environment, a new fitness calculation method is designed to evaluate the energy consumption of edge devices in the mobile edge computing environment by combining the energy consumption modeling of multi-leisure resource unloading in the mobile edge computing environment.

$$\mathrm{fitness}=({\mathrm{f}}_1\times {\mathrm{E}}_{\mathrm{sum}})+({\mathrm{f}}_2\times 10\times {\mathrm{E}}_{\mathrm{sum}}\times \frac{{\mathrm{T}}_{\mathrm{sum}}}{{\mathrm{T}}_{\mathrm{response}}})$$

(7)

$${\mathrm{E}}_{\mathrm{ec}}={\displaystyle \sum _{\mathrm{i}=1}^{\max }{\mathrm{E}}_{\mathrm{edge}}+}{\displaystyle \sum _{\mathrm{j}=1}^{\max }{\mathrm{E}}_{\mathrm{ec}}+}{\displaystyle \sum _{\mathrm{k}=1}^{\max }{\mathrm{E}}_{\mathrm{cc}}}$$

(8)

$${\mathrm{T}}_{\mathrm{sum}}={\displaystyle \sum _{\mathrm{i}=1}^{\max }{\mathrm{T}}_{\mathrm{edge}}+}{\displaystyle \sum _{\mathrm{j}=1}^{\max }{\mathrm{T}}_{\mathrm{ec}}+}{\displaystyle \sum _{\mathrm{k}=1}^{\max }{\mathrm{T}}_{\mathrm{cc}}}$$

(9)

where (sum) represents the total energy consumption of edge devices. (${\mathrm{T}}_{\mathrm{sum}}$) represents the total task execution time of mobile edge computing environment, (response) represents the workflow execution response time constraint required by users.

The fitness calculation method can meet the requirements of the response time of the edge device users to the workflow task execution and balance the energy consumption of the edge device generated by the execution of the task scheduling scheme. The greater the fitness of the scheduling scheme, the higher the energy consumption of the corresponding edge devices. On the contrary, the energy consumption required is lower. Fitness mainly includes two parts.

The energy consumption of the edge equipment of the scheduling scheme when the total task execution time is less than the response time constraint (f₁ = 1, f₂ = 0), and the fitness is expressed as the execution energy consumption of the task scheduling scheme. The algorithm focuses on optimizing the energy consumption of the scheduling scheme.

The total execution time of the task is greater than the response time constraint (f₁ = 1, f₂ = 0). In this case, the fitness is expressed as the product of the energy consumption of the scheduling scheme and the ratio of constraint when the response is exceeded. The algorithm focuses on optimizing the energy consumption of the scheduling scheme under the response time constraint.

4.2. Unloading Particle Swarm Optimization Task Scheduling Algorithm

In this section, according to the energy consumption model of multi-resource unloading in the mobile edge computing environment and the calculation method of adaptability, and energy optimization multi-resource computing and unloading particle swarm optimization task scheduling algorithm emo are proposed. Algorithm 1 is as follows.

Algorithm 1: The energy consumption optimization algorithm for multi-resource computing and unloading particle swarm optimization task scheduling.

Input: algorithm iteration, tasks, VMS, SES, EDS, response.

Output: energy best.

Step 1 For i = 1 to k do

Step 2 Random initialization task scheduling scheme (Si), search speed (Vi), and unloading decision (Di)

Step 3 End for

Step 4 For i = 10 to k do

Step 5 Calculate the energy consumption and time of task execution of three resources of unloading inkstone from the scheduling scheme (formula 1–6)

Step 6 Calculate the total energy consumption, time, and fitness value of the scheduling scheme (formula 7–9)

Step 7 End for

Step 8 The global optimal scheduling scheme with the best adaptability is selected from (k) task scheduling schemes

Step 9 For i = 1 to Iteration do

Step 10 Update all scheduling schemes according to search speed, and well update unloading decisions for each task

Step 11 For j = 1 to k do

Step 12 Calculate the energy consumption and time of task execution (formula 1–6) for tasks unloaded to three resources in the scheduling scheme (Si)

Step 13 Calculate the total energy consumption, time, and fitness value of the scheduling scheme (formula 7–9)

Step 14 Ends for

Step 15 The global optimal task scheduling scheme with the lowest adaptability is selected from (k) task scheduling schemes

Step 16 Update the inertia weight value according to the adaptive hate weight strategy

Step 17 Update the search speed of each scheduling scheme

Step 18 End for

Step 19 Return Energy-best

The emo algorithm firstly initializes the task scheduling scheme, particle search speed, and task unloading decision (Steps 1–8), then enters the algorithm iteration process. Firstly, the scheduling scheme is updated according to the search speed of the scheduling scheme, and the unloading decision is updated for each task (Step 10). Then, the unloading to the scheduling scheme (Si) is calculated according to the multi-resource unloading energy consumption model, the task execution energy consumption, and time of three resources (Step 12); then calculate the total energy consumption, time, and adaptability of edge equipment (Step 13), and select the global optimal task scheduling scheme with the lowest adaptability according to the calculation results (Step 15); further, update the inertia weight value and the search speed of the scheduling scheme according to the adaptive inertia weight strategy (Steps 16–17); finally, the optimal task scheduling scheme of energy consumption under the user response time constraint is obtained after the iteration times are reached (Step 19). The algorithm combines the multi-resource unloading energy consumption model in the mobile edge computing environment and a new method to evaluate the adaptability of edge equipment energy consumption and obtains the task scheduling scheme with optimal energy consumption of edge equipment under the constraint of user response time. Assuming that the number of scheduling schemes is (n), the number of iterations of the algorithm is (d), and the number of tasks is (T), the time complexity of the algorithm is (O (NDT)).

5. Experimental Analysis

5.1. Simulation Experiment Environment and Algorithm Parameter Setting

The simulation experiments are all run in MATLAB r2017b, and the operating environment is Intel Core i7 3.6 GHz CPU and 16 GB memory PC. The workflow is generated randomly by DAG, with the number of tasks of 50–300 for each workflow, and a load of each task in the workflow is 30–3000 megacycles, which is subject to the random value of the normal distribution. Tang [32] et al.’s evaluation of random DAGs with different extended scheduling lengths.

The data amount of each task is a random value of normal distribution between 5 MB and 300 MB. The power of data transmission and receiving of edge equipment is 100 MW and 25 MW, respectively, and its working and idle power are 700 MW and 3O MW, respectively. The CPU processing capacity is 1.0 GHz. The data upload and download rates in the mobile edge computing environment are 20 Mbps and 40 Mbps. The CPU processing capacity of the edge server is 2 1.3 GHz virtual machines, and the CPU processing capacity of the cloud server is 4 virtual machines with 1.0 GHz, 1.3 GHz, and 1.6 GHz. The user response time constraint is twice the average execution time of workflow tasks on a 1.4 GHz virtual machine.

5.2. Experimental Results and Analysis

This section compares the task scheduling algorithms based on five kinds of computing unloading strategies from the following aspects: adaptability value, mobile power consumption, and task completion time. The comparison strategies include non-unloading strategy (log as LO), full unload to cloud policy (remember as FOC), full unload to edge server policy (remember as a foe), and partial unloading strategy (recorded as SPO). Then, the adaptive value and energy consumption optimization effect of different inertia weight particle swarm optimization algorithms based on a multi-resource unloading strategy is compared. Finally, the convergence rate of the algorithm is compared with the different inertia weights.

(1)

Comparison of adaptability and energy consumption of different load-carrying strategies.

The adaptability value and energy consumption of the five kinds of unloading strategies are compared. It has been proved that the emo algorithm based on a multi-resource unloading strategy can reduce the energy consumption of mobile terminals fully under the premise of considering the time constraints of user response.

The PSO algorithm is a kind of evolutionary algorithm. It starts from the random solution, searches for the optimal solution through iteration, evaluates the quality of the solution through fitness, and searches for the global optimal value by following the currently searched optimal value. The fitness values of PSO task scheduling algorithms corresponding to the five unloading strategies are compared as shown in Figure 4, in which the Lo algorithm corresponds to the right coordinate axis, and the other algorithms correspond to the left coordinate axis. When the number of tasks changes from 50 to 300, the fitness value of the scheduling scheme based on the emo algorithm is the lowest; when the number of tasks is 50, the other 4 unloading strategies (FOC, foe, EMO) except lo, however, with the increase in the number of tasks, the fitness gap of the 5 unloading strategies is increased. When the number of tasks is 50, the fitness value of emo is 3% lower than that of SPO, and when the number of tasks is 300, the fitness of emo is 8% lower than that of SPO. This shows that the partial offload strategy proposed in this paper is suitable for large-scale task scheduling and management in the mobile edge computing environment.

The completion time of the task scheduling schemes generated by the five unloading strategies is shown in Figure 5, in which the emo algorithm has the lowest completion time for all tasks. When the number of tasks is 300, the completion time of the scheduling schemes obtained by the emo algorithm is 12%, 60%, and 58% lower than that of FOC, FOE, and SPO, respectively. This is because of the emo proposed in this paper. The algorithm can optimize the task scheduling scheme considering the user’s time constraints, while the rest of the computing offloading strategies offload all the computing to the edge or cloud server for execution, which reduces the energy consumption of the mobile terminal, but also increases the transmission delay of tasks and data in the transmission process, resulting in the completion time of the scheduling algorithm corresponding to these unloading strategies is inferior to that of the emo algorithm. Efficient task scheduling can improve the throughput of the mobile edge computing system, ensure the quality of data processing, reduce the delay caused by data processing, and solve the performance of the whole computing system to a large extent. See

Table 3. Comparison analysis table of throughput and latency metrics.

Index	Latency (s)
Throughput (kbit)	FOC	FOE	SPO	LO	EMO
50	0.1	0.2	0.24	6.3	0.06
100	0.05	0.1	0.12	3.15	0.03

for a comparative analysis of throughput and latency metrics.

(2)

Comparison of fitness values and energy consumption of different inertia weights.

In particle swarm optimization (PSO), the setting method of inertia weight will affect the optimal solution of particle search. Appropriate inertia weight can provide reasonable global or local search ability for particles and avoid falling into local optimization. Therefore, based on the multi-resource unloading strategy, this section compares the fitness values of different inertia weight particle swarm optimization task scheduling algorithms and shows the advantages of the emo algorithm corresponding to the new adaptive inertia weight in solving the task scheduling problem in the mobile edge computing environment. The comparison algorithms include traditional adaptive PSO (denoted as EA), linear PSO (denoted as EL), random PSO (denoted as ER), exponentially reduced PSO (denoted as EI), and constant value PSO (denoted as EC).

The fitness values of the scheduling algorithm corresponding to the six inertia weights are shown in Figure 6. For example, when the number of tasks is 50 to 300, the fitness value of emo is 14% lower than that of the EA algorithm. Therefore, the emo algorithm in this paper is more suitable for task scheduling and management in the mobile edge computing environment.

(3)

Convergence speed of the algorithm

The convergence of the six algorithms is shown in Figure 7. Among them, the emo algorithm finds that the number of iterations of the optimal task scheduling scheme is basically stable at about 50 times, which is better than the other 5 scheduling algorithms with inertia weight.

When the number of tasks is 50~300, the optimal scheduling scheme iterations of the 6 inertia weight scheduling algorithms are shown in

Table 4. The optimal scheduling scheme iterations.

The Number of the Tasks	Iteration Times of the Algorithm
	EA	EL	ER	EI	EC	EMO
100	55	29	18	13	60	55
150	56	32	15	9	51	53
200	61	36	23	13	39	51
250	62	37	19	10	49	54
300	60	38	19	11	53	51
350	64	51	22	11	49	52

. Combined with Figure 7 and Table 4, it can be found that the convergence of the emo algorithm is also stable while finding the optimal solution with a lower fitness value. The convergence of the other five algorithms is not stable enough or easy to fall into local optimum. Therefore, the emo algorithm is better than the other five algorithms in terms of shrinking the optimal solution and convergence.

6. Conclusions and Future Research

The existing traditional computing unload strategy only considers the single computing unload resource and cannot be applied to the complex multi-resource environment in the mobile edge computing environment. In this paper, the traditional computational offload energy consumption model is improved, and a multi-resource offloading energy consumption model is proposed. A new fitness calculation method for evaluating the energy consumption of edge devices is designed, and the computing unloading based on multiple resources is proposed as an energy consumption optimization particle swarm optimization task scheduling algorithm emo. This paper, under a specific experimental environment, compares the task scheduling algorithms based on five kinds of computing unloading strategies. The experimental results show that the convergence speed of emo is relatively stable; when the number of tasks is 50 to 300, the fitness value of emo is 14% lower than that of the traditional adaptive PSO algorithm. The emo algorithm has the lowest completion time for all tasks. When the number of tasks is 300, the completion time of the scheduling schemes obtained by the emo algorithm is 12%, 60%, and 58% lower than that of FOC, FOE, and SPO, respectively. The emo algorithm finds that the number of iterations of the optimal task scheduling scheme is basically stable at about 50 times, which is better than the other 5 scheduling algorithms with inertia weight. This is because the emo algorithm proposed in this paper can optimize the task scheduling scheme considering the user’s time constraint, while other computing offloading strategies unload all computing to the edge or cloud server, which reduces the energy consumption of mobile terminals but also increases the transmission delay of tasks and data in the transmission process. As a result, the completion time of the scheduling algorithm corresponding to these unload policies is not as good as that of the emo algorithm.

Therefore, the emo algorithm in this paper is more suitable for task scheduling and management in the mobile edge computing environment, and the energy consumption of edge devices of the task scheduling scheme is better than the other four unloading strategies under the constraint of user response time. However, in this paper, the power of edge devices is set to a fixed value. In the process of task execution, with the change in CPU utilization, the energy consumption of edge devices may change accordingly. Therefore, it is necessary to study the impact of dynamic CPU utilization on device energy consumption. In the future, we will also consider applying this strategy to specific business process scenarios to test the effect of energy consumption optimization.