CN114666339B - Edge unloading method and system based on noose set and storage medium - Google Patents

Abstract

The present invention relates to the technical field of edge computing offloading, and discloses an edge offloading method, system and storage medium based on neutrosophic sets. The second cost required when being executed on the edge server, the first edge server is the edge server closest to the mobile device among the N edge servers; when the first cost is greater than the second cost, the task is offloaded to the first edge Execute in the server; when the first edge server reaches the load threshold, lock the second edge server based on the context parameter of each edge server in the N edge servers, offload the task to the second edge server for execution; design a The multi-level edge computing offloading strategy solves the problem of whether tasks need to be offloaded and where, and reduces the cost of executing tasks.

Description

A method, system and storage medium for edge unloading based on CICI

Technical Field

The present invention relates to the field of edge computing offloading technology, and in particular to an edge offloading method, system and storage medium based on CISCO.

Background Art

With the rapid development of wireless communication technology and Internet technology, the era of the Internet of Everything has quietly arrived. It is estimated that by 2030, the number of mobile devices in China will reach 4 billion. At the same time, new applications such as VR, online games, and driverless cars have sprung up like mushrooms after rain. The amount of data generated by such applications has increased dramatically in size and complexity, and the requirements for latency have become increasingly stringent, otherwise it will be difficult to meet the user experience quality. Although the CPUs equipped in mobile devices are becoming more and more powerful, due to the size, storage space, and power consumption, mobile devices themselves are unable to cope with latency-sensitive and computationally intensive applications. Therefore, edge computing came into being.

Edge computing (EC) is a new computing model that deploys computing and storage resources, such as cloudlets (edge servers), micro data centers or fog nodes, at the edge of the network closer to mobile devices or sensors to solve the shortcomings of devices in terms of resource storage, computing performance and energy efficiency, so as to provide fast processing and low-latency services. However, edge computing still faces many technical challenges, such as computing offloading and mobility management. As one of the key technologies in edge computing, computing offloading refers to the technology in which terminal devices hand over part or all of the computing tasks to edge servers for processing, which plays an important role in minimizing latency and ensuring service quality. Mobility management refers to how users can choose appropriate edge nodes to provide services for themselves according to their mobile trajectories when they are in dense and complex network coverage areas. In response to the problems existing in edge computing, a large number of studies and algorithms have been conducted in academia, which aim to explore the formulation of optimal offloading decisions to minimize energy consumption while meeting execution delay constraints. However, these studies also have some shortcomings. For example, some algorithms assume that all tasks need to be offloaded by default, which is obviously unreasonable because some tasks may be processed more cheaply inside the mobile device. When there are multiple edge nodes near the mobile device, how to simultaneously consider the contextual factors of the edge nodes and the mobile device and select the optimal node by integrating multiple attributes is not sufficient. In some studies, the offloading decision model and mobility management are either carried out separately or the offloading decision is made entirely based on mobility. In fact, mobility is an important factor that characterizes the impact of context on decision-making. It can be seen that the offloading method of the existing edge offloading method is costly.

Summary of the invention

The present invention provides an edge unloading method, system and storage medium based on Zhongzhiset to solve the problem that the unloading method of the existing edge unloading method has a high cost.

In order to achieve the above object, the present invention is implemented by the following technical solutions:

In a first aspect, the present invention provides an edge offloading method based on a central intelligence set, which is applied to a computing offloading structure, wherein the computing offloading structure includes a cloud computing center, N edge servers and a mobile device, where N is a positive integer, and the method includes:

S1, calculating a first cost required when a task is executed on a mobile device and a second cost required when the task is executed on a first edge server, where the first edge server is the edge server closest to the mobile device among the N edge servers;

S2. When the first cost is greater than the second cost, offloading the task to the first edge server for execution;

S3. When the first edge server reaches a load threshold, lock a second edge server based on context parameters of each candidate edge server among the N edge servers, where the context parameters include user mobility, network conditions, load of each edge server, and CPU utilization;

S4. Offloading the task to the second edge server for execution.

In a second aspect, an embodiment of the present application provides an edge unloading system based on CITS, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method described in the first aspect when executing the computer program.

In a third aspect, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in the first aspect.

Beneficial effects:

The edge offloading method based on the central intelligence set provided by the present invention, when the task required to be executed on the mobile device is large, first considers offloading the task to the first edge server closest to the mobile device for execution, and when the first edge server reaches the load threshold, locks the second edge server based on the context parameters of each edge server in N edge servers, and offloads the task to the second edge server for execution. In this way, a multi-level edge computing offloading strategy is designed to solve the problem of whether the task needs to be offloaded and where to offload, and reduce the cost of executing the task; on this basis, in this application, the second edge server is determined according to user mobility, network conditions, the load of each edge server and CPU utilization, fully considering the real-time nature of user movement within different cloud coverage ranges, and using the central intelligence set to process the high variability of context parameters over time, which can not only save energy and time, but also reduce the number of failed tasks.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of an edge offloading method based on a central intelligence set according to a preferred embodiment of the present invention;

FIG2 is a schematic diagram of a three-layer computing offloading structure according to a preferred embodiment of the present invention;

FIG3 is a schematic diagram of a cloud model according to a preferred embodiment of the present invention;

FIG4 shows the change of user mobility over time in a preferred embodiment of the present invention;

FIG5 is a graph showing the average number of task failures of NSCO according to a preferred embodiment of the present invention and a method of a comparative experiment;

FIG6 is an average consumption time of NSCO of a preferred embodiment of the present invention and two comparative methods when processing different numbers of tasks;

FIG. 7 shows the average energy consumption of NSCO according to the preferred embodiment of the present invention and two comparative methods when processing different numbers of tasks.

DETAILED DESCRIPTION

The technical solution of the present invention is described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

Unless otherwise defined, the technical terms or scientific terms used in the present invention shall have the usual meanings understood by persons with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, words such as "one" or "one" do not indicate quantity restrictions, but indicate the existence of at least one. Words such as "connect" or "connected" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the object being described changes, the relative positional relationship also changes accordingly.

Referring to FIG. 1 , an embodiment of the present application provides an edge offloading method based on a central intelligence set, which is applied to a computing offloading structure. The computing offloading structure includes a cloud computing center, N candidate edge servers, and a mobile device, where N is a positive integer. The method includes:

S1, calculating a first cost required when the task is executed on the mobile device and a second cost required when the task is executed on the first edge server, where the first edge server is the candidate edge server closest to the mobile device among the N candidate edge servers;

The edge offloading method based on CICI provided in this application is applicable to the three-layer computing offloading structure as shown in Figure 2, in which the cloud computing center can provide stable and powerful computing power, suitable for processing computing-intensive and delay-tolerant tasks. Cloudlet (edge server) is a server with computing and network resources deployed at the edge of the network, with computing power superior to mobile devices and lower latency than cloud computing. Four offloading destinations are considered in this application: local mobile device, the nearest cloudlet (first edge server), the best cloudlet (second edge server), and the cloud computing center, which are represented by subscripts m, nea, Opt, and Cl in the following calculations.

In a three-layer cloud-edge hybrid environment, the offloading problem is described as how to choose where to execute tasks and how to select an optimal cloudlet based on contextual factors to minimize the overall completion time and energy consumption. Specifically, suppose there are n tasks in total. If W tasks are executed locally, x tasks are executed on the nearest cloudlet, y tasks are executed on the best cloudlet, and Z tasks are executed on the cloud server, then the total cost of completing this set of tasks is:

W+x+y+Z=n (1)

C _Site consists of two parts, namely the task completion time T and the consumed energy E:

C _site =α·T _site +β·E _site ;

site∈{m, nea, opt, cl}; (2)

α and β are weight factors for adjusting the time and energy consumption parts of the cost, which can be adjusted according to the user's preference for these factors. The calculation of completion time T and energy consumption E is introduced in detail later. Their values are normalized and then used in the calculation of the cost.

S2. When the first cost is greater than the second cost, offload the task to the first edge server for execution.

It should be noted that not all tasks need to be offloaded, and mobile devices can also provide a good operating environment for some small tasks. This application determines whether a task needs to be offloaded by the offload cost.

The latency and energy consumption of completing tasks on mobile devices are calculated as follows:

T _m =I/S _m (3)

_Em = _Pmo ·(I/ _Sm ) (4)

Where I refers to the task size, which is described by the number of instructions, _Sm is the number of instructions executed by the mobile device per unit time, and _Pmo is the power consumption per unit time when the mobile device executes the task.

If the task request is answered by the nearest cloudlet, the corresponding calculation is as follows:

In this embodiment, the explanations of the parameters involved include: D _nea / _Sp represents the propagation time, _Du / _Bu and D _d / _Bd represent the upload and return link times, I/S _nea represents the time spent by the nearest cloudlet to process the task, Q _nea represents the waiting time in the nearest cloudlet, _Du represents the amount of uploaded data, D _d represents the amount of returned data, Bu represents the upload data rate, _Bd represents the return data rate, P _ts represents the power consumption per unit time when the mobile device sends data, P _tr represents the power consumption per unit time when the mobile device receives data, P _mo represents the power consumption per unit time when the mobile device performs task calculation, _Pi represents the power consumption per unit time when the mobile device is in an idle state (waiting for the result), _I represents the number of instructions of the task to be executed, and the task size is represented by the number of instructions, S _m represents the number of instructions executed by the mobile device per unit time, S _nea represents the number of instructions executed by the nearest cloudlet per unit time, S _opt represents the number of instructions executed by the best cloudlet per unit time, S _cl represents the number of instructions executed by the cloud computing center cloud per unit time, S _p represents the propagation speed, D _cl represents the distance between the nearest cloudlet and the cloud, D _nea represents the distance between the nearest cloudlet and the mobile device, D _opt represents the distance between the nearest cloudlet and the best cloudlet, Q _nea represents the queuing time in the nearest cloudlet, Q _opt represents the queuing time in the best cloudlet, and T _timer represents the set timer time.

Therefore, when C _m ≤ C _nea , it is less costly to process the task locally on the mobile device. If C _m ＞ C _nea , it is more appropriate to offload the task to the nearest cloudlet.

S3. When the first edge server reaches a load threshold, locking the second edge server based on context parameters of each edge server in the N edge servers, where the context parameters include user mobility, network conditions, load of each edge server, and CPU utilization;

S4. Offload the task to the second edge server for execution.

The above-mentioned edge offloading method based on Zhongzhiji, when the task is executed on a mobile device and the required time is large, first considers offloading the task to the first edge server closest to the mobile device for execution, and when the first edge server reaches the load threshold, locks the second edge server based on the context parameters of each edge server in the N edge servers, and offloads the task to the second edge server for execution. In this way, a multi-level edge computing offloading strategy is designed to solve the problem of whether the task needs to be offloaded and where to offload, and reduce the cost of executing the task; on this basis, in this application, the second edge server is determined according to the user mobility, network conditions, server load and CPU utilization of each candidate edge server, and the real-time movement of users within different cloud coverage areas is fully considered.

Optionally, S3 specifically includes:

S31, considering the candidate edge servers other than the first edge server among the N candidate edge servers as candidate edge servers, and constructing a time-varying context matrix of each candidate edge server according to the context parameters of each candidate edge server in the most recent q moments;

S32, using the reverse cloud generator algorithm to convert the time-varying context matrix of each candidate edge server into a single-valued neutral context matrix;

S33, using a single-valued neutral intelligence set weighted average aggregation operator to aggregate the single-valued neutral intelligence context matrix into a single-valued neutral intelligence number of the candidate edge server;

S34: Calculate the score of each candidate edge server using a score function of a single-valued neutral number, and use the candidate edge server with the highest score as the second edge server.

In this optional implementation, the high variability of context parameters over time is processed by using a neutral intelligence set, which can not only save energy and time, but also reduce the number of failed tasks.

In some scenarios, the nearest cloudlet reaches the load threshold and cannot provide services for newly arrived tasks. At this time, the nearest cloudlet needs to act as a proxy, broadcast requests, and seek help from other nearby cloudlets. Specifically, in one example, the steps of locking the second edge server (the nearest cloudlet) are as follows.

The first edge server broadcasts the task request message, searches for other nearby cloudlets, and sets a timer T _timer , T _timer ＜＜T _nea . When there are cloudlets nearby that can meet the task request, these cloudlets respond to the proxy with their CPU utilization, current load, and network connection status with other cloudlets. The proxy cloudlet selects the best cloudlet based on the information of the candidate cloudlets and forwards the task to the cloudlet for execution. In this case, the delay T _opt , energy consumption E _opt and cost are calculated as follows:

Optionally, the above method further includes: setting a preset time threshold, and if the second edge server is not successfully locked within the preset time threshold, offloading the task to the cloud computing center.

In this optional implementation, if no cloudlet responds to the task request when the timer T _timer reaches 0, it means that there is no suitable cloudlet nearby. At this time, the task is unloaded to the cloud computing center, and the cloud server provides the service. In this way, it can be ensured that regardless of whether the second edge server is successfully locked, the task on the mobile device can be unloaded, reducing the pressure on the mobile device. In this case, the delay T _opt , energy consumption E _opt and cost are calculated as follows:

It should be noted that in different time periods, the network status presents different levels of congestion, the available server resources also change dynamically over time, and the terminal devices are in different locations at different times due to their mobility. That is, the network status, server resources and terminal device locations are all dynamically updated over time. These contextual factors that affect the offloading decision show time-varying characteristics.

Tasks can be offloaded to different cloudlets at different times and locations. These time-varying contextual factors play an important role in offloading decisions, but few existing studies consider the time-varying dynamics of context parameters. This chapter will use a single-valued neutral intelligence set to characterize the time-varying nature of four context parameters: user mobility, network conditions, server load, and CPU utilization.

i∈{1，2，...，p}，j∈{1，2，...，q}。其余三个上下文参数包括候选cloudlet的负载、CPU利用率及其与代理cloudlet之间的网络条件，可以通过相应API获取对应的数值，分别记为

考虑到移动性是极大型指标(越大越好)，其余三个是极小型指标，为了避免尺度混乱，采用y′＝max_y-y极小型指标进行正向化处理，max_y为指标y取值的最大值。同时，为使不同量纲的特征处于同一数值量级，减少方差大的特征的影响，采用y＝(y-min_y)/(max_y-min_y)据进行归一化处理。其中，min_y为指标y取值的最小值，max_y为指标y取值的最大值，因此，对用户服务范围内的p个候选cloudlet，在最近q个时刻内其归一化的上下文属性数据可表达为矩阵

每个矩阵元素

即cloudleti在时刻j的归一化上下文属性数据，为一个四元组

i∈{1，2，...，p}，j∈{1，2，...，q}。则有：When the current user is within the service range of p candidate cloudlets, in order to select the best cloudlet, the values of the four contexts of the p candidate cloudlets in q moments are obtained to facilitate decision making. This application characterizes mobility by estimating the user's stay time in cloudlets, denoted as

i∈{1, 2, ..., p}, j∈{1, 2, ..., q}. The other three context parameters include the candidate cloudlet’s load, CPU utilization, and the network conditions between it and the proxy cloudlet. The corresponding values can be obtained through the corresponding API and are denoted as

Considering that mobility is an extremely large indicator (the larger the better), and the other three are extremely small indicators, in order to avoid scale confusion, the extremely small indicator y′=max_y-y is used for positive processing, and max_y is the maximum value of the indicator y. At the same time, in order to make the features of different dimensions at the same numerical order and reduce the impact of features with large variance, y=(y-min_y)/(max_y-min_y) is used for normalization processing. Among them, min_y is the minimum value of the indicator y, and max_y is the maximum value of the indicator y. Therefore, for the p candidate cloudlets within the user service range, their normalized context attribute data in the last q moments can be expressed as a matrix

Each matrix element

That is, the normalized context attribute data of cloudleti at time j is a four-tuple

i∈{1, 2, ..., p}, j∈{1, 2, ..., q}. Then:

The time-varying context information is uncertain, and the Neutrosophic Set (NS) has independent membership functions, uncertain membership functions, and non-membership functions, which can well characterize the inconsistent and uncertain information caused by time-varying. However, it is defined on a non-standard unit subinterval]0-,1+[,.

First, the definition of single-valued neutrosophic set (SVNS) is given. Let X be a given domain. A single-valued neutrosophic set A on X is represented by a membership function _TA (x), an uncertain membership function _IA (x) and a non-membership function _FA (x):

A＝{<X, T _A (x), I _A (x), F _A (x)>lx∈X}

有0≤T_A(x)+I_A(x)+F_A(x)≤3。对论域X上的单值中智集A中的一个元素，称为单值中智数(Single-Valued NeutrosophicNumber，SVNN)，简记为<T_A，I_A，F_A>。Where T _A (x), I _A (x), F _A (x)∈[0,1], satisfying

_0≤TA (x)+ _IA (x)+ _FA (x)≤3. An element in a single-valued neutrosophic set A on the domain X is called a single-valued neutrosophic number (SVNN), abbreviated as < _TA , _IA , _FA >.

建立三种隶属度函数，即完成

的转换。为此，本申请选用一种基于概率统计和模糊集理论、实现定性概念与定量数据双向转换的认知模型——云模型(Cloud Model，CM)来完成这一转换。In order to establish an SVNS model for the time-varying context attributes of cloudlet and facilitate the subsequent multi-attribute decision-making based on the neutrino-intelligence set, it is necessary to provide context sequence data of cloudleti within q moments.

Establish three membership functions, that is, complete

To this end, this application uses a cognitive model based on probability statistics and fuzzy set theory to achieve bidirectional conversion between qualitative concepts and quantitative data, the Cloud Model (CM), to complete this conversion.

It is worth explaining that a cloud is composed of cloud droplets. A cloud droplet is a realization of a qualitative concept. A certain number of cloud droplets can express a cloud. In this application, the value of a time-varying context of cloudlet _i at time j can be regarded as a cloud droplet, and its data sequence at time q can represent the cloud model of the context, which is recorded as

熵

超熵

三个数值来表征。期望

是云滴在论域空间分布的期望，代表定性概念的基本确定性；熵

代表定性概念不确定性的度量，反映了在论域中可被这个概念所接受的数值范围，在图3中反映云的跨度。超熵

是熵

的熵，是熵的不确定性，表达云模型的偏离程度，在图3中反映云的厚度。因此，取期望

当作单值中智数的隶属度T_A，熵

当作不确定隶属度I_A，超熵

当作非隶属度F_A，可完成到SVNS的转换。即：As shown in Figure 3, the digital features of the cloud model are composed of expected values

entropy

Super Entropy

Three values are used to represent it.

is the expectation of the distribution of cloud droplets in the domain space, representing the basic certainty of qualitative concepts; entropy

A measure of uncertainty in a qualitative concept, reflecting the range of values that can be accepted by the concept in the domain, and in Figure 3, the span of the cloud.

It is entropy

The entropy is the uncertainty of entropy, which expresses the degree of deviation of the cloud model and reflects the thickness of the cloud in Figure 3. Therefore, the expected

As the membership degree of the single-valued neutrosophic number T _A , entropy

As uncertain membership _IA , super entropy

Assuming it as non-membership degree F _A , the conversion to SVNS can be completed. That is:

Furthermore, this application uses the reverse cloud generator algorithm in cloud model theory to complete the conversion:

i∈{1，2，...，p}；输出可以是cloudlet_i某一上下文云模型的数字特征值，①

②

③

Specifically, the input of the reverse cloud generator algorithm can be a time-varying context sequence of cloudlet _i , such as

i∈{1，2，...，p}; the output can be the digital feature value of a context cloud model of cloudlet _i , ①

②

③

In an example, the above process is explained by taking the load data of cloudlet ₁ in 10 moments as an example. After the data is normalized and normalized, the time-varying load data of cloudlet ₁ is (0.98, 1.00, 0.54, 0.00, 0.60, 0.51, 0.26, 0.82, 0.80, 0.50), and the corresponding cloud model is:

熵

超熵

因此，cloudlet₁的负载单值中智数表示为

Input the data into the reverse generator to obtain the expected digital features corresponding to the cloud model

entropy

Super Entropy

Therefore, the load single-valued neutrinodic number of cloudlet ₁ is expressed as

使用云模型的逆向生成器算法进行中智转化后，建立的cloudlet_i的单值中智上下文模型分别为：So far, if the time-varying data series of cloudlet _i ’s network conditions, load, and CPU utilization are

After using the cloud model's reverse generator algorithm for neutrosophic transformation, the single-valued neutrosophic context models of cloudlet _i are established as follows:

To establish a single-valued neutrosophic mobility model, mobility must be measured first.

Please refer to Figure 4, which illustrates the change of user mobility over time. Assume that the user starts to move along the direction of α, at a certain moment, the user changes the moving direction to α1, and at the next moment, the user moves along α2. This paper measures mobility by predicting the residence time of the mobile device in the cloudlet to provide a reference for the selection of cloudlets.

和速度v可通过GPS获得，D为用户当前位置和cloudlet的直线距离，

为用户当前位置到cloudlet的方向向量，设用户当前位置为(A，B)，cloudlet位置为(a，b)，有

则用户在某一cloudlet内的停留时间可通过如下公式计算：Assume that at a certain moment the user is in the service area of two cloudlets at the same time. Taking cloudlet2 as an example, R represents the service range of the cloudlet, S is the distance the user is away from the coverage area of the cloudlet along the moving direction, v represents the user's moving speed, and the user's moving direction

and speed v can be obtained through GPS, D is the straight-line distance between the user's current location and the cloudlet,

is the direction vector from the user's current position to the cloudlet. Assume that the user's current position is (A, B) and the cloudlet's position is (a, b).

The user's stay time in a cloudlet can be calculated using the following formula:

The distance S can be calculated by trigonometric function:

利用逆向云生成器算法进行中智转化后，所建立的移动性单值中智集模型为

Therefore, the mobility data sequence of the user in cloudlet _i at q moments is

After using the reverse cloud generator algorithm to transform the neutral intelligence, the established mobility single-valued neutral intelligence set model is:

i∈{1，2，...，p}，行表示p个候选cloudlet，列为移动性、网络条件、负载、CPU利用率四个上下文因素的SVNS表示如下：In summary, for cloudlet _i , its single-valued neutrosophic context model is as follows:

i∈{1, 2, ..., p}, the rows represent p candidate cloudlets, and the columns represent the four contextual factors of mobility, network conditions, load, and CPU utilization as follows:

After obtaining the single-valued neutral-intelligence context model of cloudlet _i , a decision is made based on the four relevant contextual factors of the candidate cloudlets to select the best cloudlet. Obviously, this is a multi-attribute decision-making problem. In multi-attribute decision-making problems, the attributes of each alternative are often complex, and different contextual attributes contribute to the decision to different degrees, so they should be given different weight values. Because the attribute weights are completely unknown, they conform to the entropy characteristics of fuzzy theory. Therefore, this application uses the neutral-intelligence entropy theory to calculate the optimal weights of each context attribute in the SVNS environment.

是论域X＝{x₁，x₂，…，x_n}上的一个单值中智集，则中智熵E(A)的定义为

A^c为A的补集。在

中，每一个元素

表示cloudleti的某一上下文的SVNN，每一列，代表每一个上下文属性的SVNS。因此，每个属性对应的权重计算如下：set up

is a single-valued neutrosophic set on the domain X = {x ₁ , x ₂ , …, x _n }, then the neutrosophic entropy E(A) is defined as

A ^c is the complement of A.

In each element

It represents the SVNN of a context of cloudleti. Each column represents the SVNS of each context attribute. Therefore, the weight corresponding to each attribute is calculated as follows:

where crtt∈{M, D, L, C} represents four context attributes.

Finally, using formula (13) - Single-valued Neutrosophic Set Weighted Algorithm (SVNSWA), the context attributes SVNS of cloudlet _i are aggregated into the S VNN of candidate cloudlet _i , denoted as SVNN _i = {T _i , I _i , F _i }, where cnt∈{M, D, L, C}.

After obtaining SVNN _i , use formula (14) - SVNN score function to calculate the score of each candidate cloudlet. The score function is an important indicator in SVNN sorting. The larger the membership T, the larger the SVNN; the smaller the uncertainty I, the larger the SVNN; similarly, the smaller the non-membership F, the larger the SVNN. After obtaining the score list of candidate cloudlets, the one with the highest score is the best cloudlet.

score(SVNN _i )=(T _i +1-I _i +1-F _i )/3 (14)

In summary, the time-varying context-aware edge offloading method based on the Neutrosophic Set (NSCO) of the present application can be described by the following algorithm:

{Input: Task = <I, _Du >, I is the number of instructions of the task to be executed, _Du is the amount of data uploaded during offloading.

Output: The execution location of task t.

1. Use formulas (2)(3)(4)(5)(6) to calculate the cost _Cm and _Cnea of executing the task locally and on the nearest cloudlet

2.if C _m ≤C _nea :

3. Execute locally;

4.else:

5. The cloudlet recently received the task;

6. If the cloudlet meets the task requirements recently:

7. Execute the task on the nearest cloudlet;

8.else:

9. The nearest cloudlet acts as a proxy and broadcasts the task request message to other available cloudlets nearby;

10. If within the time T _timer , there are other cloudlets that can satisfy the task's offloading request:

11. Obtain the user's mobility, network conditions, load, and CPU utilization in the last q moments of all candidate cloudlets to form the time-varying context matrix of cloudlet _i

转化为单值中智上下文矩阵：12. Use the reverse cloud generator algorithm to

Transformed into a single-valued neutrosophic context matrix:

聚合成候选cloudlet_i的SVNN。13. Use formula (13) to

Aggregate into SVNN of candidate cloudlet _i .

14. Use formula (14) to compare SVNN _i . The one with the highest score is the best cloudlet, and the task is transferred from the nearest cloudlet to the best cloudlet.

15.else:

16. If there is no nearby cloudlet that can perform the task within _{T timer} , the task will be offloaded to the cloud.

}

The main process of the algorithm is to determine the execution location of the task: first, steps 1-7 indicate whether to offload the task to the nearest cloudlet based on the cost of executing the task _Cm and _Cnea . Then steps 8-14 indicate that when the nearest cloudlet cannot meet the task request, the single-valued neutral-intelligence set is used to characterize the time-varying context information of the candidate cloudlets, and the cloudlets time-varying context information is used to select an optimal cloudlet to execute the task. Finally, steps 15-16 indicate that if there is no suitable candidate cloudlet within T _timer , the task is offloaded to the cloud computing center. The time complexity of the algorithm is O(n), among which the time complexity of the single-valued neutral-intelligence set weighted average aggregation operator is O(n), and the time complexity of using the SVNN comparison function to compare the candidate cloudlets is O(n). The two are in parallel, the time complexity of other algorithm processes is O(1), and the comprehensive time complexity is O(n).

Next, the performance of the edge offloading method based on the central intelligence set provided by the present application is verified through experiments.

(1) Dataset: This paper uses the Stanford Drone dataset and the Alibaba cluster dataset. The Stanford Drone dataset continuously records the movement trajectories of pedestrians in a certain area on the Stanford University campus, and contains the specific location information of users. The Alibaba cluster dataset provides cluster tracking from actual production, recording relevant data of 4,000 servers in 8 days. The experiment extracts the CPU utilization and load data of the servers. The network condition is measured by the communication delay, which is calculated according to (2) simulation parameters.

(2) Simulation parameters: This paper uses Advantech EIS-D210 (3846MIPS, 1.5GHz, 4GB RAM) as cloudlet parameters and Del PowerEdge (31790MIPS, 3.0GHz, 768GB RAM) as cloud server parameters. The service range of cloudlet is set to 0-50m, the bandwidth between user and cloudlet is set to 100Mbps, and the bandwidth between user and cloud server is set to 1Gbps. The task size is distributed uniformly, with an average of 4600M instructions. The data transmission size of each task is also distributed in the same way, with an average of 750 kilobytes.

(3) Comparative experiments: The proposed algorithm NSCO is compared with Application-aware cloudlet selection for computation offloading in multi-cloudlet environment (appAware) and mCloud: AContext-Aw are Offloading Framework for Heterogeneous Mobile Cloud (mCloud).

appAware: Different Cloudlets can execute different types of applications. They are assigned to different Cloudlets based on the requested application type, balancing workloads, reducing system latency, and lowering power consumption.

mCloud: Considers changes in mobile device context (network conditions) to provide assistance in selecting wireless media and cloud resources, thereby making better offloading decisions to provide better performance and lower battery consumption.

(4) Evaluation indicators: This paper uses three evaluation criteria: average number of task failures, response time, and energy consumption. A failed task is defined as the time a user spends in a cloudlet that is less than the completion time of offloading the task to the cloudlet, because in this case, the user has left the service area of the cloudlet and cannot receive the result.

In a case study, 20 cloudlets were defined to simulate the task offloading experiment, and the original data of the context parameters within 10 moments were converted into a neutral intelligence set. The results are shown in Table 1. Then, using formula (13) for weighted average aggregation, the single-valued neutral intelligence number of each candidate cloudlet can be obtained, as shown in Table 2. Finally, using formula (14), a score list of candidate cloudlets can be obtained:

cloudletl5>cloudlet5>cloudletl3>cloudletT>cloudlet3...>cloudletl4.

Cloudlet15 becomes the best candidate because it exhibits higher membership and lower non-membership and uncertainty in each context parameter, while the worst cloudlet14 is the opposite.

Table 1. Neutral numbers of various context parameters

Table 2. Aggregate single-valued neutrinocity of candidate cloudlets

Then, a comparative analysis is performed as follows:

(1) Analysis of the average number of task failures

When the number of tasks is 25, 50, 75, and 100, the average number of task failures is measured. The average number of task failures of the time-varying context-aware edge offloading method based on neutral intelligence set (abbreviated as NSCO) proposed in this paper and the comparative experimental method is shown in Figure 5.

There are two reasons for the above advantages: (1) NSCO takes into account the high mobility of users and captures the user's mobility trend through the time they stay within the cloudlet range. Predicting the user's stay time can help filter out cloudlets with less available time, thereby avoiding potential task offloading failures. (2) NSCO uses historical data to infer the future. The most suitable cloudlet in the past period of time should also be the best in the near future. When the nearest cloudlet cannot meet the task requirements and turns to look for the best cloudlet around, NSCO uses relevant contextual historical moment data and combines cloud models, neutrinocular aggregation and other algorithms to select the nearest optimal cloudlet. This method can play an important role in reducing the number of task offloading failures.

In the comparative experiments appAware and mCloud, neither the dynamics of time nor the mobility of users were considered in selecting cloudlets, so there were many failed tasks. In addition, in the comparative experiment appAware, only whether there is a certain type of cloudlet dedicated to processing a certain type of task was considered. If there is no cloudlet of that type, it will choose to offload to the cloud center, but transferring tasks to the cloud will increase the response time, and long response time often leads to task failure. In the comparative experiment mCloud, only the context condition of the network interface was considered, and the task was processed only based on whether the network interface was available. At some point, a large number of tasks were processed by local devices. As we all know, the processing capacity of local devices is limited, and processing failures are prone to occur.

(2) Analysis of time and energy consumption of task offloading

Figures 6 and 7 show the average time and energy consumption of NSCO and the two comparison methods when processing different numbers of tasks. It can be seen from the figure that the average response time of the proposed solution is reduced by about 28.9% and 54.7% compared with appAware and mCloud, and the average energy consumption is reduced by about 33.2% and 56.8% compared with appAware and mCloud, respectively.

In addition to the analysis in (1) that the long response time leads to task failure, thus increasing the system response time, in the appAware method, if there is no dedicated cloudlet to handle a certain task type, it will choose to offload to the cloud center, but transferring the task to the cloud will increase the propagation delay, so the response time and energy consumption are affected. In mCloud, due to the unreasonable task ratio distribution, too many tasks are processed locally, which increases the response time and energy consumption.

In the scheme proposed in this paper, the nearest Cloudlet is selected first. If it cannot offload the task requirements, the nearest Cloudlet acts as a proxy server and selects the best Cloudlet from its nearby Cloudlets to handle the task. If none of the nearby Cloudlets respond, the task is offloaded to the cloud. Moreover, when selecting the optimal Cloudlet, this paper takes a global perspective and fully considers four contextual factors: user mobility, network conditions, cloudlet CPU utilization, and cloudlet load, which correspond to the propagation time, communication time, processing time, and queuing time in the response time, respectively. Therefore, the time spent on offloading tasks using this scheme is short, the energy consumed is low, and the effect of improving the overall user experience is achieved.

In summary, this paper studies the problem of task offloading under edge computing. A computing offloading strategy that considers multiple contextual factors such as user mobility is proposed. When the nearest cloudlet cannot handle the offloaded task, the problem is transformed into a multi-attribute decision-making problem, and an optimal cloudlet is selected from the vicinity to handle the task. A neutral intelligence set is used to handle the highly dynamic changes in context data over time. Simulation experimental results show that using our proposed strategy, latency and power consumption are reduced by 28.9%-54.7% and 33.2%-56.8%, respectively.

In this embodiment, the edge server may also refer to an edge cloud.

The present application also provides an edge unloading system based on a central intelligence set, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the computer program. The edge unloading system based on a central intelligence set can implement various embodiments of the above edge unloading method based on a central intelligence set, and can achieve the same beneficial effects, which will not be described in detail here.

The present application also provides a computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, the method steps described above are implemented. The computer-readable storage medium can implement various embodiments of the above-mentioned edge offloading method based on the central intelligence set, and can achieve the same beneficial effects, which will not be described here.

The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims (7)

1. An edge offloading method based on a central intelligence set is applied to a computing offloading structure, wherein the computing offloading structure includes a cloud computing center, N edge servers and a mobile device, where N is a positive integer, and is characterized in that the method includes: S1, calculating a first cost required when a task is executed on a mobile device and a second cost required when the task is executed on a first edge server, where the first edge server is the edge server closest to the mobile device among the N edge servers; S2. When the first cost is greater than the second cost, offloading the task to the first edge server for execution; S3. When the first edge server reaches a load threshold, locking a second edge server based on context parameters of each of the N edge servers, where the context parameters include user mobility, network conditions, load of each edge server, and CPU utilization; S4, offloading the task to the second edge server for execution; The S3 specifically includes: S31, considering edge servers other than the first edge server among the N edge servers as candidate edge servers, and constructing a time-varying context matrix of each candidate edge server according to context parameters of each candidate edge server in the most recent q moments; S32, using the reverse cloud generator algorithm to convert the time-varying context matrix of each candidate edge server into a single-valued neutral context matrix; S33, using a single-valued neutral intelligence set weighted average aggregation operator to aggregate the single-valued neutral intelligence context matrix into a single-valued neutral intelligence number of the candidate edge server; S34: Calculate the score of each candidate edge server using a score function of a single-valued neutral number, and use the candidate edge server with the highest score as the second edge server. 2. According to the edge unloading method based on Zhongzhiji according to claim 1, it is characterized in that the method also includes: setting a preset time threshold, if the second edge server is not successfully locked within the preset time threshold, the task is unloaded to the cloud computing center. 3. The edge offloading method based on central intelligence set according to claim 1, characterized in that the step S31 comprises: The estimated user's residence time within the range of the candidate edge server is taken as the user mobility of the candidate edge server, denoted as

Obtain the network conditions corresponding to the candidate edge server through the candidate edge server API

Server Load

And CPU utilization

Normalize the context parameters of the candidate edge server to obtain a time-varying context matrix

as follows:

Where q represents a total of q time points, p represents a total of p candidate edge servers, i∈{1, 2, ..., p}, j∈{1, 2, ..., q}. 4. The edge offloading method based on neutral intelligence set according to claim 3 is characterized in that the single-valued neutral intelligence number of the candidate edge server in S33 satisfies the following relationship:

Wherein, each element represents the neutral number representation of the context of the candidate edge server i, _Ti represents the membership of the candidate edge server i, _Ii represents the uncertain membership of the candidate edge server i, and _Fi represents the non-membership of the candidate edge server i. 5. The edge offloading method based on central intelligence set according to claim 4, characterized in that the S34 comprises: The score function score(SVNN _i ) of the single-valued neutron number is established as follows: score(SVNN _i )=(T _i +1-I _i +1-F _i )/3; The score function is used to calculate the score of each candidate edge server, and the candidate edge server with the highest score is used as the second edge server. 6. An edge unloading system based on Zhongzhiset, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any of the methods described in claims 1 to 5 when executing the computer program. 7. A computer-readable storage medium having a computer program stored thereon, wherein when the program is executed by a processor, the method steps as claimed in any one of claims 1 to 5 are implemented.

Images