The mobility of edge computing is the primary concern, but edge computing should not be limited to mobile edge computing (MEC). It is worth mentioning that some scholars “MEC” multi access edge computing.

The edge computing industry alliance has proposed a feasible architecture for edge computing, which is widely used in existing research. A simplified schematic diagram is shown in Fig. 1.

The terminal layer is located at the edge of the network, which is composed of various terminal devices and generates a large amount of heterogeneous data. Limited by the resources of terminal devices, most terminal devices cannot complete those tasks with a long duration and a large amount of computation.

The edge layer acts as a bridge between the cloud layer and the terminal layer, which is closer to the terminal devices at the edge of the network. The edge layer completes the processing of data generated by terminal devices and supports terminal layer services nearby.

Cloud is rich in computing, storage and other resources, and is good at dealing with long-term, large amounts of data and complex computing. However, with the rich types and number of terminals, the data generated shows an explosive growth trend. Due to network transmission costs, cloud computing has some deficiencies in supporting emerging low latency applications.

Edge computing and cloud computing are complementary technologies, and each has its advantages. The emergence of edge computing extends cloud services to the edge of the network. It will become a new development trend of cloud computing technology to process the data at the edge nearby and cooperate in the cloud center. Reference [19] propose that the value of edge computing and cloud computing can be magnified, and they also provide reference framework and application scenarios of “edge cloud” collaboration.

Most of the existing edge computing researches focus on the edge layer, and form the “Edge-Terminal” binary architecture with the terminal devices, which has some limitations in the global and integrity [6,7,11,12]. In this paper, a Cloud-Edge-Terminal collaborative framework is proposed. The cloud center is introduced into the traditional “Edge-Terminal” binary architecture, and the edge computing and cloud computing are organically combined to construct the “Cloud-Edge-Terminal” collaborative framework. This paper proposes a framework for bringing the advantages of cloud computing resources to the network edge layer, and at the same time synergizing the local advantages of edge computing and cloud computing. The diagram of Cloud-Edge-Terminal framework is shown in Fig. 2.

This paper focuses on the process of the terminal user unloading the task to the edge server. The edge computing server is closer to the data generation side, but its resources are still limited compared with those of the cloud center. Therefore, it is necessary to find a suitable edge computing server to allow the terminal user to offload the task.

Considering that in practice, the priorities for offloading requests of different user tasks is different, and the resources provided by the server at the edge layer are often different at different times. At the same time, considering the user demand and the load of the edge server can improve the efficiency of task offloading, as well as reduce service time and energy consumption. Taking into account the possibility of further task segmentation, the task is further divided into more fine-grained tasks, so as to explore the parallelism between subtasks.

This paper mainly focuses on resource collaboration in cloud edge collaboration. The task unloading of terminal devices first selects the edge server cluster closer to each other. The cloud is responsible for the scheduling of task unloading, making the unloading decision, processing the unloading request of the terminal, matching the best edge server for the unloading task, and participating in the collaborative computing when the edge device resources are insufficient. The process is shown in Fig. 3.

Definition 1 Center Cloud

The center cloud is represented as a 3-tuple cloud (m, c, b).(1)

“m” represents the memory resources that the center cloud can provide;

Definition 2 Edge Server

The edge server is represented as a 4-tuple E_i (m, c, b, t₀).(1)

“m” represents the memory resources that the edge server can provide;

Definition 3 Terminal

The terminal is represented as a 2-tuple Ter_i (m, c).(1)

“m” represents the memory resources that the terminal can provide;

Task division aims to divide task J into multiple subtasks ω_i(I = 1, 2, …). Then according to whether unloading is allowed or not, it can be divided into two categories: allowing unloading and not allowing unloading. In this paper, directed acyclic graphs are used represent the task after partition. D (V, E):V represents subtasks and E is the directed edge between nodes, as shown in the Fig. 4.

Definition 4 Task

The task is represented as a 4-tuple Task_i (m, c, tmax,Mob ).(1)

“m” represents the memory requirement of this task;

Definition 5 Subtask

The task is represented as a 6-tuple Subtask_j (m, c, tmax,taski , d_in, d_out).(1)

“m” represents the memory requirement of this task;

Subtasks should meet the two constraints:(1)

The starting subtask, that is, the task without degree must be executed locally;

According to the constraint rules, subtasks A and B in Fig. 5 must be executed locally at the terminal, and tasks C through H are allowed to be unloaded. The specific execution location depends on the remaining resources of each server and the network conditions.

Definition 6 Delay

The total task processing delay is defined as follows:(1) Tall=tc+t0+tw where t_c represents task task_i unload to edge computing server E_j transmission delay and total return delay of calculation results; t₀ represents the processing delay on the edge server; and t_w is the waiting delay of unloading process.

Definition 7 Energy Consumption

The total task processing energy consumption is defined as follows:(2) En=∑i=1n⁡Pi∗Ti where P_i represents the power of device i;  Ti represents the continuous running time of the device i.

This paper uses an optimized genetic algorithm to solve the problem. The flowchart of the algorithm is shown in Fig. 6. Genetic algorithm uses roulette selection, and fitness function selects the reciprocal of time. The mutation selection is random point crossover.(1)

Coding guidelines

The goal of the task scheduling method is to match the most suitable destination server for the task, and Dests is used to represent the final unloading method. For the sake of simplicity, the local processing task code is 0, the cloud server number is 1, and the edge servers are coded from 2 in turn. The target server uses binary encoding, and the final result is the final scheduling method.

DestS = [0000 0001 0002 0003 …](2)

Selection

In this paper, roulette algorithm is used as the generation method, and the reciprocal of the system cost is defined as the fitness function.(5) Fitness =  1Cost_Func (3)

Crossover

The process of cross operation and chromosome exchange to produce new offspring is one of the processes of population producing new offspring. The common crossover operations include single point crossover, multi-point crossover and extended crossover.(4)

Mutation

Single point mutation and multiple point mutation are common methods for mutation operation and chromosome individual gene mutation. Single point mutation is used in this paper.(5)

Annealing process

In order to reduce the probability of genetic algorithm falling into local optimal solution in the early stage, a simulated annealing factor is introduced.

Disturbance probability is introduced in Eq. (6).(6) p=exp(−ΔFitness)

If P/(K*T) > rand, accept the suboptimal population. The “rand” is a rand number, where “K” is constant, and “T” is the current temperature. The temperature drop follows the rule Eq. (8), “σ” annealing coefficient.(7) rand ∈(0,1) (8) T(t + 1)=σT(t)

In this section, the above is tested and verified. The experiment is based on the Python programming language, and the simulation platform parameter information is shown in Tab. 2.

The experiment adopts the current mainstream method, of using Python for data simulation, and generating experimental parameters through random methods [11,12]. The parameter information is shown in Tabs. 3 and 4.

This paper verifies for experimental sensitive applications. The experimental design consists comparing task localization processing with task random offloading processing in order to prove the necessity of task offloading; then verifying the task scheduling method proposed in this paper, and finally comparing the algorithm proposed in this paper with the task average distribution algorithm.

The equipment energy consumption parameters are shown in Tab. 4.

In order to determine the effectiveness of the algorithm, we first compared the performance of the improved genetic algorithm. The optimal scheduling results were analyzed under the condition that the number of edge server nodes is 10 and the number of cloud center nodes is 1. The algorithm parameters are shown in Tab. 5.

The experimental results are shown in Tab. 6.

The experimental results show that under the conditions of this experiment setting, as the number of tasks increased from 100 to 700, the improved algorithm becomes more effective.

Next the relationship between the optimal iteration results of the algorithm and the number of iterations were compared. The experiment was carried out under the conditions that the number of tasks was 600, the number of edge computing servers was 10, and the number of cloud centers is 1. To demonstrate the optimization performance of the algorithm, the experiment was repeated 50 times, and the comparison between the average optimal result and the optimal result was conducted. The results are shown in Tab. 7.

The experimental results show that the improved genetic algorithm has a smaller adaptation value and better searching effect under the same number of iterations.

Fig. 7 shows the fitness of the improved genetic algorithm when the number of tasks was 600 and the number of iterations was 500.

Because genetic algorithms have relatively few parameters, they are easy to implement. Therefore, for to analyze the influence of crossover and variation factors on the parameters of the improved genetic algorithm in term of the optimization results, the experiment was carried out using 500 iterations and 600 tasks. The results are shown in Tab. 8

Next, the unload necessity proof experiment was carried out to compare the unload processing with local task processing and unscheduled task processing. The experimental results are shown in Fig. 8.

Fig. 8 shows the comparison of task localization processing and the total time delay of task randomly unloading task processing under the “Cloud-Edge-Terminal” framework proposed in this paper. Analysis of the experimental results, show that random task offloading can reduce the total task processing time by 91.4%. Therefore, edge computing for task offloading can reduce the total task duration.

Then, for the number of tasks from 100 to 700, under the condition of the number of edge servers is 10 and the number of cloud centers is 1.

The delay weighting factor ω₁ = 1 was selected in experiment, and the longest processing time of the device was regarded as the duration of the method. The longest processing time of the device was selected as the processing delay, the experiment was verified, and the task local terminal processing, average offloading and the offloading method designed in this paper were compared. The experimental results shown in Fig. 9, confirm that the offloading algorithm proposed in this paper can reduce the task processing time by 43.35% on average compared with the average offloading algorithm, and reduce the task processing time by 77.67% with task localization.

Task local processing means that the tasks generated by the terminal are processed on the terminal, and the task processing time is the genetic algorithm-based scheduling algorithm parameters discussed in this paper, as shown in Tab. 5.

When the number of tasks is 500, verify the relationship between the number of tasks and the number of edge servers. The experiment selects the number of edge servers [10, 100], and the specific results are shown in Fig. 10. The experimental results show that when the number of edge servers is small, the average computational delay of the algorithm proposed in this paper is decreased significantly.

Finally, the experiment vertifies the relationship between the task delay and the number of algorithm iterations in an environment where the number of tasks is 100, 200, and 500, and the number of edge servers is 50. The experimental results show that the task processing time of the 100 to 300 generations of iterations is significantly reduced. Starting with the 300th generation, the time has been reduced slowly. The experimental results are shown in Fig. 11.