Cloud edge end resource cost minimum task offloading method satisfying deadline requirement

Abstract

The application relates to a cloud edge terminal resource cost minimum task unloading method meeting a deadline requirement, time and energy consumption of all devices completing all tasks are calculated and stored in a time matrix and an energy consumption matrix respectively; all elements in the time matrix and the energy consumption matrix are normalized respectively; a target weight vector of each task is set, a weighted arithmetic average operator is adopted to perform weighted summation on a target cost value, a cost value paid by the device for completing the task is solved, and the solved cost value is stored in a resource cost matrix; through a Hungarian algorithm, based on the device matching the minimum cost for the task, the best resource is matched for all tasks, the best unloading scheme of all tasks is obtained, the conflict of task resource demand is solved continuously, the unloading scheme of all tasks meets the minimum total cost of the system, the minimum total cost of the system can be met under the premise of meeting the task processing deadline, and good processing performance is shown.

Description

TECHNICAL FIELD

[0001] The present application relates to a cloud edge end resource cost minimum task offloading method meeting the deadline requirement. BACKGROUND

[0002] With the large-scale growth of computationally intensive and delay-sensitive mobile tasks such as online games, real-time image signal processing, augmented reality, and real-time translation services, mobile mobile terminals (MD) such as mobile phones, tablets, and VRs have brought huge computing demands, but their power and computing capacity are limited, and it is difficult to meet the needs of these tasks for long battery life and low latency. Cloud servers have strong computing power and can complete a large number of computing problems that mobile terminals cannot complete, but when a large amount of data is uploaded to the cloud server, it is inevitable to generate a large amount of energy consumption and transmission delay, making it difficult to meet the high service quality requirements of users.

[0003] Mobile edge computing (MEC) arranges servers at the network edge near mobile terminals, shortens the ability of remote cloud services and computing to a location closer to mobile terminals, thereby reducing the transmission of data with low latency and servers, improving server processing efficiency. When a mobile terminal submits a task, it can choose to process data on an edge server with strong computing power, without the need to transmit to the cloud for processing, reducing data transmission time and improving task processing efficiency. Therefore, the mobile edge computing computing mode solves the problem of limited mobile device capability and high latency in cloud computing mode. Compared with cloud computing, the computing, network, and storage resources provided by edge computing are still limited. When a large number of users simultaneously submit task offloading requests to the edge server, it is inevitable that users will compete for edge server resources. Unreasonable task offloading may lead to insufficient utilization of computing, storage, network, and other resources, and there may be uneven busy situations, which seriously reduces the service efficiency of edge computing and causes the quality of service to decline. Therefore, the edge computing resource allocation problem of multiple users simultaneously submitting task offloading requests to resource-constrained edge servers is very worthy of study.

[0004] In the prior art, the mobile task offloading problem is converted into a Markov decision process, and the mobile task is scheduled according to the queuing state of the mobile task buffer, the execution state of the local processing unit and the state of the transmission unit. By analyzing the average time to complete each task and the average energy consumption of the mobile device, the power-constrained mobile task delay minimization problem is modeled, and the optimal offloading strategy of the problem is found by a one-dimensional search algorithm. The prior art also proposes a multi-user collaborative mobile edge computing system that minimizes the computing delay of local users by optimizing task allocation, time allocation and power allocation. The above work is from the theoretical level, and the effectiveness of the results is verified by a large number of simulation experiments. In real environment conditions, if only the minimum delay offloading is considered, due to the limited battery capacity of the mobile device, when the battery energy is exhausted, the corresponding task processing will be interrupted, and the corresponding decision cannot continue. In real environment conditions, by optimizing the transmission power of the device and the quality of the network link to reduce energy consumption, although the energy consumption is reduced, but due to the change of network environment with time, the device in low power state may cause high delay.

[0005] The prior art also proposes a "terminal-edge-cloud" collaborative 5G vehicle networking edge computing model, and a deep reinforcement learning and deep reinforcement learning collaborative distributed service offloading method is designed for the system model, which can reduce the user average service delay and energy consumption by 0.4%-20.4%. Although the reduction of delay and the reduction of energy consumption are considered comprehensively, the dependency relationship between tasks is not considered, however, in reality, each mobile terminal will generate multiple tasks, and there may be a certain dependency relationship between these tasks, such as face recognition tasks, target tracking tasks, etc. At this time, if only single task computing offloading is considered, the offloading of tasks with dependency relationship may fail, thereby causing a chain reaction to cause a large deviation between the offloading result and the expectation. SUMMARY

[0006] In order to solve the above technical problems, the present application provides a cloud edge terminal resource cost minimum task offloading method meeting the deadline requirement.

[0007] The present application adopts the following technical scheme:

[0008] A cloud edge terminal resource cost minimum task offloading method meeting the deadline requirement, comprising:

[0009] Calculate the time and energy consumption of all devices to complete all tasks, and store them in the time matrix and the energy consumption matrix respectively;

[0010] All elements in the time matrix and the energy consumption matrix are normalized respectively;

[0011] Set the target weight vector of each task, the target weight vector including the target weight of the time and energy consumption of the device to the task;

[0012] The time matrix and the energy consumption matrix after normalization are weighted and summed by using a weighted arithmetic average operator to obtain the value of the objective function, and the value of the objective function is stored in the resource cost matrix;

[0013] The optimal resource for all tasks is matched by using the Hungarian algorithm based on the device with the minimum cost for the task, and the optimal offloading scheme for all tasks is obtained.

[0014] In one embodiment, the optimal resource for all tasks is matched by using the Hungarian algorithm based on the device with the minimum cost for the task, and the optimal offloading scheme for all tasks is obtained, including:

[0015] An offloading strategy is initialized, and a mobile task is randomly selected from the task node set, the offloading strategy is initialized as an empty set, and the offloaded task set is initialized as an empty set;

[0016] The current task offloadable device resource node set is traversed, and the minimum cost offloading is selected to update the offloading strategy; the task is deleted from the task node set and inserted into the offloaded task set;

[0017] If offloading to the device resource makes the superposition value of the decision value greater than or equal to 1, it indicates that there is device resource competition, the Hungarian algorithm is used to solve the resource competition and select the optimal offloading scheme and update the offloading strategy;

[0018] When the task node set is an empty set, the maximum matching number is reached, and the optimal offloading decision is obtained.

[0019] In one embodiment, the calculation process of the time and energy consumption of all devices for all tasks is as follows:

[0020] The time required by each mobile terminal to process the mobile task generated by it, the time required by each mobile terminal to process the mobile task generated by it, and the time required by the virtual machine to process the task when the mobile task generated by each mobile device is offloaded to the virtual machine on the cloud server are obtained; and the weighted sum operation is performed in combination with the corresponding weight to obtain the time required by all devices to complete all tasks;

[0021] The energy consumption of each mobile terminal processing the mobile task generated by it, the energy consumption of each mobile terminal processing the mobile task generated by it, and the energy consumption of the virtual machine processing the task when the mobile task generated by each mobile device is offloaded to the virtual machine on the cloud server are obtained; and the weighted sum operation is performed in combination with the corresponding weight to obtain the energy consumption of all devices to complete all tasks;

[0022] Accordingly, the system cost is obtained according to the time and energy consumption of all devices to complete all tasks, in combination with the total time and total energy consumption of all tasks calculated locally, and the proportion of the calculation time delay and the calculation energy consumption in the system cost;

[0023] Based on the minimum system cost, the calculation offloading decision problem is converted into a minimum system cost problem with constraints.

[0024] The beneficial effects of the present application include: in the cloud edge end environment, aiming at the problems of limited processing capacity of mobile devices, limited resources of edge servers and difficult satisfaction of task service quality requirements, a mobile task offloading method for minimizing resource cost meeting the deadline requirement is proposed, first, the resource task cost function is given according to the time and energy consumption of mobile tasks processed on different device resources, in combination with the target planning algorithm, the cost of resource task processing is calculated, the multiple service quality satisfaction problem is converted into the task offloading problem of minimum resource cost, then the appropriate resources are matched for the tasks by using the Hungarian matching algorithm and the conflict of task resource demand is solved constantly, so that the task offloading scheme of the whole task meets the minimum system total cost, the system total cost is minimized under the premise of meeting the task processing deadline, and good processing performance is shown. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 is a system architecture diagram of the cloud edge end collaborative offloading of multiple tasks provided by the present application;

[0026] Figure 2 is a task processing process schematic diagram;

[0027] Figure 3 is a whole flow schematic diagram of a task offloading method with minimum resource cost meeting the deadline requirement of the cloud edge end provided by the present application;

[0028] Figure 4 is a specific cloud edge end computing environment example diagram;

[0029] Figure 5 is a processing process diagram of a target tracking task;

[0030] Figure 6 is a schematic diagram of the relationship between omega1 and time;

[0031] Figure 7 is a schematic diagram of the relationship between omega1 and energy consumption;

[0032] Figure 8 is a schematic diagram of the relationship between omega1 and cost;

[0033] Figure 9 is a system total time delay curve diagram under different numbers of mobile terminals;

[0034] Figure 10is a system total energy consumption curve diagram under different mobile terminal quantities;

[0035] Figure 11 is a computing time curve diagram under different virtual machine quantities;

[0036] Figure 12 is a computing energy consumption curve diagram under different virtual machine quantities;

[0037] Figure 13 is a system total time delay curve diagram under different network bandwidths;

[0038] Figure 14 is a system total energy consumption curve diagram under different network bandwidths. DETAILED DESCRIPTION

[0039] The embodiment of the application provides a task offloading method of a cloud edge end with minimized resource cost meeting as-of-time requirement, research on minimized resource processing cost meeting task processing time delay requirement is carried out in the scene of limited mobile device processing capacity, limited edge server resource and difficult satisfaction of task service quality requirement, and the specific work is as follows: (1) modeling of an edge computing framework of cloud edge end cooperation, a mobile task model, a computing time model and a computing energy consumption model of local computing, edge computing and cloud computing, and a resource minimum cost model in an edge server is proposed. (2) based on a target programming algorithm and a Hungarian algorithm, a mobile task offloading algorithm with minimized resource cost that satisfies the as-of-time requirement (MCDROA) is proposed to solve an offloading strategy of minimized resource cost meeting task time delay requirement. According to time and energy consumption spent in processing tasks, the algorithm firstly calculates the cost spent in processing tasks by each resource through the target programming algorithm, and then continuously optimizes and finds out a resource minimum cost scheme based on the Hungarian algorithm, so that all tasks are completed by the system with total minimum cost.

[0040] Firstly, a cloud edge end system and mobile task modeling are introduced. The cloud edge end system modeling is specifically as follows:

[0041] Figure 1is a system architecture for task offloading in cooperation with cloud computing, edge computing and mobile terminals (referred to as cloud-edge-terminal for short), mainly including three physical devices of cloud server, edge server and mobile terminal, and using virtual machines (VM) virtually generated by the cloud server and the edge server to provide services to the mobile terminal task. Each mobile terminal can establish a cellular connection with the edge server, or establish a link with the cloud server. The embodiment studies the different requirements of different mobile terminals for task delay and energy consumption, dynamically adjusts the task execution strategy and virtual machine resource configuration strategy to optimize the service performance of the whole system, and reduces the system operation cost.

[0042] For mobile terminals: mobile terminals are the environment for generating tasks, such as face recognition, target tracking, online games, etc. It is assumed in the embodiment that each mobile terminal runs only one mobile application. The mobile application can generate multiple tasks with dependency relationship, and each task can choose mobile device computing, virtual machine computing in the edge server or virtual machine processing in the cloud server.

[0043] Let the set of mobile terminals be represented as MB={mb1,mb2,…,mb K}, and each mobile terminal can be represented as a three-tuple mb k ={a k ,o k ,memory k}, where k is the index of the mobile terminal, a k represents the total computing power of the mobile terminal (i.e. the number of CPU cycles per unit time), o k represents the current processing power ratio of the mobile terminal, and memory k represents the available memory resources of the mobile terminal.

[0044] For edge servers: edge servers are deployed at the network edge close to mobile terminals to provide computing, network and storage services for tasks submitted by mobile terminals.

[0045] The set of edge servers is represented as ES={es1,es2,…,es M}, and each edge server can be represented as a five-tuple es m ={a m ,o m ,memory m ,er m ,ep m}, where m is the index of the edge server, a m represents the total computing power of the edge server es m , o m represents the current processing power ratio of the edge server es m , memorym available memory resource of the edge server es m available memory resource of the edge server es m data transmission rate between the edge server and the mobile terminal ep m available memory resource of the edge server es m transmit power of the edge server es

[0046] For cloud servers: the set of cloud servers is denoted as CS = {cs1, cs2,..., cs L}, and each cloud server can be denoted as a three tuple cs l = {f l , cr l , cp l}. Wherein, / is the index of the cloud server, f l represents the total computing power of the cloud server cs l , cr l represents the data transmission speed between the cloud server cs l and the mobile terminal, and cp l represents the transmit power of the cloud server cs l .

[0047] For virtual machines: a virtual machine refers to a computer that is virtually created on a server (cloud server or edge server) according to the requirements of a task by using virtualization technology, and is used to execute a certain type of task. All virtual machines form a virtual machine pool, and the virtual machines are allocated to tasks by a scheduling center according to a certain strategy. A virtual machine vm ij can be denoted by a four tuple Wherein, cpu i is the available processing capacity of the server where the virtual machine is located, memory i is the available memory size of the server where the virtual machine is located, is the CPU percentage of the virtual machine vm ij allocated from the cpu i of the server, is the memory percentage of the virtual machine vm ij allocated from the memory i of the server. When a task is offloaded to an edge server or a cloud server for data processing, the actual processing is performed on the virtual machine virtually created on the server.

[0048] Mobile task modeling is as follows: A mobile task is a task submitted by a mobile application running on a mobile terminal, such as a face recognition task or a root-finding task. Tasks are generally divided into two types: independent tasks and dependent tasks. Independent tasks are those that are indivisible, such as finding the root of an equation; dependent tasks are those that can be automatically divided into multiple subtasks with sequential dependencies. Subtasks may interact with each other, meaning that a subtask cannot start execution before obtaining all the information of its parent task, such as a face recognition task.

[0049] This embodiment assumes that the mobile terminal is MB. k The resulting mobile application contains N subtasks with dependencies. Each subtask can be represented as a quintuple. Among them, w j Indicates task The computational load is measured in the number of CPU cycles required; f min.j Indicates processing task T j k Minimum computing power required; memory j Indicates task The minimum memory resources required for execution; Indicates mobile terminal mb k The generated task The amount of data to be transferred during unloading, t max.j Indicates task Maximum allowable delay time.

[0050] When the task is unloaded to the edge node es m Above, if the edge nodes satisfy the task If the required computing and memory resources are met, the task will be processed immediately upon arrival at the resource node. If the edge node has sufficient storage resources but insufficient CPU computing resources, the task will be stored in memory upon arrival at the computing node. The CPU resources will be released after the computing resources have completed several tasks to meet the task requirements before data processing begins. The specific processing procedure is as follows: Figure 2 As shown.

[0051] Figure 2 In the diagram, a1, a2, and a3 represent the arrival of tasks 1, 2, and 3 at the edge server es, respectively. m The timeframe is as follows: Assume that when Task 1 and Task 3 arrive at the node, the node's resources are greater than the resources required by the tasks, and the tasks will execute immediately upon arrival. If, when Task 2 arrives at the node, the node's remaining resources are insufficient to meet Task 2's requirements, then Task 2 cannot execute immediately. After Task 1 completes, it immediately releases its occupied resources. If the total available resources at this point can meet the computational resource requirements of Task 2, then Task 2 will begin execution.

[0052] The mobile task processing model and system processing cost are formalized as follows: The scheduling center allocates physical resources to mobile tasks in the form of virtual machines based on the processing time cost and processing energy consumption cost of the task.

[0053] For the mobile task processing time model:

[0054] (1) Processing time of mobile tasks on local devices: When the computing and memory resources of the mobile terminal meet the requirements of the mobile task... If the resource requirements are insufficient, the task can be processed locally. In this case, the processing time on the mobile terminal depends on the computational load of the task and the processing power of the mobile terminal. The current processing power f of the mobile terminal... k for:

[0055] f k =(1-o k )a k (1)

[0056] Among them, a k Indicates the total processing power of the mobile terminal, o k This indicates the percentage of currently available processing power on the mobile terminal.

[0057] Mobile terminal mb k Process the movement tasks it generates Time required for:

[0058]

[0059] Among them, w j Indicates a movement task The computational load.

[0060] (2) Processing time of mobile tasks on the edge server: The mobile terminal unloads the mobile task to the virtual machine of the edge server for processing through the cellular network. The processing time of the mobile task includes the transmission time of the mobile task being uploaded to the virtual machine, the waiting time of the mobile task in the virtual machine, the processing time of the mobile task in the virtual machine, and the transmission time of returning the task processing result to the mobile terminal.

[0061] Based on whether the resources of the edge server meet the resource requirements of the virtual machine, the mobile task... On the edge server m Processing time There are two main scenarios:

[0062] When the mobile task arrives at the edge server, if the current available computing resources and storage resources of the server meet the resource requirements of the task, a virtual machine (denoted as vm m ) is generated to execute the task immediately. The current processing capacity f m of the edge server is:

[0063] f m =(1-o m )a m (3)

[0064] wherein a m represents the total computing capacity of the edge server es m , and o m represents the current load of the edge server es m (i.e., the proportion of the current processing capacity).

[0065] The time required for the mobile task generated by the mobile terminal mb k to be offloaded to the virtual machine vm m for processing is:

[0066]

[0067] wherein t represents the transmission time of the mobile task offloaded to the virtual machine vm m , t represents the computing time of the mobile task on the virtual machine vm m , and t represents the time for the virtual machine vm m to return the result to the mobile terminal.

[0068] The mobile task generated by the mobile terminal adopts an orthogonal frequency division multiple access scheme for data transmission. To ensure the orthogonality of the user uplink transmission channel, it is assumed that only one device is allowed to occupy each link between the mobile terminal and the edge server. According to the Shannon formula, the uplink transmission rate er k of the mobile terminal mb m and the virtual machine vm m is:

[0069]

[0070] wherein W is the bandwidth between the channels, p k represents the transmission power of the mobile terminal, N o represents the Gaussian noise power inside the channel, and h k.m represents the mobile terminal mb k and the virtual machine vm m ​​​​The channel gain between the mobile terminal and the edge server.

[0071] The mobile terminal mb k generates a mobile task and uploads it to the virtual machine vm m . The transmission time is:

[0072]

[0073] wherein, Tm represents the size of the mobile task generated by the mobile terminal. and The values of Tm and Tvm can be solved by equations (7) and (8).

[0074]

[0075]

[0076] wherein, Tvm represents the percentage of computing capacity allocated to the virtual machine from the server, Tvm represents the size of the computing result data obtained by the virtual machine processing the task .

[0077] When the task arrives at the edge server es m , if the current remaining resources of the edge server are less than the required resources of the virtual machine required by the mobile task, the mobile task needs to wait until the virtual machine on the server finishes its task and releases enough resources, and then a virtual machine (denoted as vm m ) is generated to execute the task , and the completion time is equation (9):

[0078]

[0079] wherein, Tvm represents the time for the task to arrive at the edge server es m to the virtual machine vm m to start processing the task .

[0080]

[0081] (3) The processing time of the mobile task in the cloud center: when the mobile device mb k generates a mobile task T j k is offloaded to the virtual machine vm l on the cloud server cs l for processing, the vml Processing task T j k Required time is:

[0082]

[0083] wherein, denotes the task transmitted by the mobile terminal to the virtual machine vm l time, CPT l j denotes the task computing time on the virtual machine vm l denotes the task the time of the result data returned by the virtual machine vm l to the mobile terminal.

[0084]

[0085]

[0086]

[0087] wherein, cr l denotes the data transmission speed between the cloud server cs l and the mobile terminal mb k f l denotes the processing capacity of the cloud server cs l is the CPU percentage allocated to the virtual machine vm l from the CPU of the server, l denotes the result data set size obtained by the virtual machine vm l processing task

[0088] Mobile task processing energy consumption model:

[0089] (1) When the task is selected to be calculated locally, the mobile terminal mb k processing mobile task produces energy consumption as shown in formula (15):

[0090]

[0091] wherein, denotes the energy consumption produced by the mobile terminal 1 CPU cycle, σ1 denotes the switching capacitor of the mobile terminal (switching capacitor determined by the chip structure).

[0092] (2) When the mobile terminal mb​​​​k generated mobile task offloaded to edge server es m generated virtual machine vm m When performing data processing, assuming that the energy consumption in the parallel computing process is not considered in the model, the vm m The energy consumption generated by processing the mobile task is shown in equation (16):

[0093]

[0094]

[0095]

[0096]

[0097] wherein, represents the mobile task offloaded to the vm m generated transmission energy consumption, represents the vm m processing the mobile task generated energy consumption, represents the vm m the processing task transmission energy consumption of the result obtained by returning to the mobile terminal, σ2×f i 2 represents the es m 1 CPU cycle generated energy consumption, ep m represents the edge server es m transmit power.

[0098] (3) When the mobile terminal mb k generated task offloaded to the cloud server cs l generated virtual machine vm l performing data processing, the virtual machine vm l The energy consumption generated by processing the mobile task is shown in equation (20):

[0099]

[0100]

[0101]

[0102]

[0103] wherein, represents the mobile terminal offloading the task to the virtual machine vml Transmission energy consumption during the process Represents virtual machine (vm) l Handling mobile tasks The energy consumption generated Represents virtual machine (vm) l Task processing The result is returned to the mobile terminal, resulting in transmission power consumption, σ3×f l 2 This represents the energy consumption of a cloud server in one CPU cycle, cp l Indicates cloud server cs l The transmission power.

[0104] The minimum total cost of the system can be formally described as: the total time T required for the system to complete all tasks. total Total energy consumption E total As shown in equations (24) and (25). Where α k +α m +α l =1, α k α m and α l The value is determined by the execution location of the task.

[0105]

[0106]

[0107] To evaluate the merits of the current computational offloading strategy, this embodiment defines the system cost F as:

[0108]

[0109] in ω1 and ω2 represent the total time and total energy consumption for all tasks to be computed locally. ω1 and ω2 represent the proportions of computation latency and computation energy consumption in the system cost, respectively.

[0110] From equation (26), it can be seen that when T total E total The smaller the value, the smaller the F value. Therefore, to minimize the total system cost, the computational unloading decision problem can be transformed into a problem of minimizing F under constraints, i.e.:

[0111]

[0112] Resource cost minimum mobile task offloading strategy: in the embodiment model, the mobile terminal generates a mobile task which can be processed in the mobile terminal, edge server or cloud server (the three are collectively referred to as system device resources). If the terminal task selects to be unloaded to the edge server or the cloud server for processing, the server processes according to the quality of service of the mobile task. The embodiment assumes that the virtual machine releases the resources immediately after processing the task. No matter which device resource processes the mobile task, it needs to spend a certain amount of time and energy. According to the time and energy consumed by the task in each device resource, the target planning algorithm is used to obtain the resource cost of each device resource processing each task, and a resource task cost matrix is generated. Then, according to the cost matrix, the Hungarian algorithm is used to match the appropriate system device resources for the task, so as to complete all tasks with the minimum total cost.

[0113] Multi-criteria target planning algorithm evaluates the cost of resource processing task: the target planning algorithm is mainly used to solve multi-objective decision-making problems. Through multiple criteria, the advantages and disadvantages of each decision are evaluated, and the optimal decision is selected. Due to the diversity of objectives and criteria, multi-objective decision-making is relatively complex, and it is difficult to find a solution that makes all objectives optimal. The target planning algorithm converts many contradictory objectives (such as ensuring large production and high product yield, low cost, etc. in production) into one objective to solve multi-objective decision-making. In this embodiment, according to the time and energy consumed by the mobile task in different device resources, the target planning algorithm is used to evaluate the resource cost of each device resource processing the current task, and the multi-objective offloading problem is converted into a minimum cost offloading problem. Based on this, the embodiment provides a cloud edge end resource cost minimum task offloading method that meets the deadline requirement, as shown in Figure 3 , which includes:

[0114] Step S1: calculate the time and energy consumed by all devices to complete all tasks, and store them in the time matrix and energy consumption matrix respectively:

[0115] Calculate the time and energy consumed by all devices to complete all tasks, and store them in the time matrix and energy consumption matrix respectively. If the resources in the current running environment cannot meet the resource requirements of a task, find the maximum value max d in the two matrices, and set the value at the corresponding position of the task to max d +1.

[0116] Wherein, the calculation process of the time and energy consumed by all devices to complete all tasks is as follows:

[0117] The time required for each mobile terminal to process the mobile task generated thereby, the time required for the mobile task generated by each mobile terminal to be processed by the corresponding virtual machine, and the time required for the virtual machine to process the task when the mobile task generated by each mobile device is offloaded to the virtual machine on the cloud server are obtained, and a weighted summation operation is performed in combination with the corresponding weights to obtain the time required for all devices to complete all tasks. The specific process can be seen from formula (24) above.

[0118] The energy consumption generated by each mobile terminal in processing the mobile task generated thereby, the energy consumption generated by the mobile task generated by each mobile terminal in being processed by the corresponding virtual machine, and the energy consumption generated by the virtual machine in processing the task when the mobile task generated by each mobile device is offloaded to the virtual machine on the cloud server are obtained, and a weighted summation operation is performed in combination with the corresponding weights to obtain the energy consumption required for all devices to complete all tasks. The specific process can be seen from formula (25) above.

[0119] Accordingly, the system cost is obtained in combination with the total time and total energy consumption of the task calculated locally and the proportion of the calculation time delay and the calculation energy consumption in the system cost according to the time and energy consumption required for all devices to complete all tasks. The specific process can be seen from formula (26) above.

[0120] Based on the minimum system cost, the calculation offloading decision problem is converted into a minimum system cost problem with constraints. The specific process can be seen from formula (27) above.

[0121] Step S2: all elements in the time matrix and the energy consumption matrix are normalized respectively:

[0122] Generally, the less time and energy consumption required for the system device to complete the mobile task, the lower the cost paid by the system. All elements in the time matrix and the energy consumption matrix are normalized respectively by formula (28).

[0123]

[0124] wherein, d ij represents an element in the matrix, d max represents the maximum value of all elements in the matrix.

[0125] Step S3: set the target weight vector of each task, which includes the target weight of the time and energy consumption of the device to the task:

[0126] Set the target weight vector of each task ω1,ω2 wherein ω1,ω2 represent the target weight of the time and energy consumption of the device r i to the task ,

[0127] Step S4: The weighted arithmetic average operator is used to weight and sum the target value of generation after normalization processing of time matrix and energy consumption matrix, and the value paid by the device to complete the task is calculated, and the value obtained by solving is stored in the resource cost matrix:

[0128] The weighted arithmetic average operator is used to weight and sum the target value of generation after normalization processing of time matrix and energy consumption matrix by formula (29), and the value paid by the device r i to complete the task d ij , and the value obtained by solving is stored in the resource cost matrix D T = (d ij ) u×v , where u, v represent the number of devices and the total number of tasks, respectively.

[0129]

[0130] The above steps S1-S4 constitute the multi-standard target planning algorithm (MSTPA) to evaluate the cost algorithm of resource processing task, and the pseudo code is shown in Table 1:

[0131] Table 1

[0132]

[0133]

[0134] Suppose the cloud edge computing environment is composed of four mobile mobile terminals, an edge server and a cloud server. The edge server virtualizes three virtual machines, and the cloud server virtualizes one virtual machine. Each mobile terminal generates a mobile application containing a number of tasks with dependency relationship, as shown in Figure 4 .

[0135] The tasks generated by the mobile mobile terminal can be executed locally, unloaded to the edge server virtual machine for execution, or executed on the cloud server virtual machine. The time cost and energy cost of the task T i are calculated by formula (3)-formula (24). j k

[0136] Suppose in Figure 4 ​The time and energy consumption of tasks generated by mobile terminals processed on each device in the edge computing environment shown in Table 2. This embodiment does not consider that the mobile terminal generates tasks to be unloaded to other mobile terminals for processing, such as tasks T1, T2, T3 generated by mobile device Mb1 cannot be unloaded to mobile devices Mb2, Mb3, Mb4 for data processing. This part of the computing device can be considered as having a higher cost for mobile tasks, so this part of the data is set to a value greater than the maximum value in the table.

[0137] Table 2

[0138]

[0139]

[0140]

[0141]

[0142] The time and energy consumption matrices are normalized by formula (28), assuming ω1 = ω2 = 0.5, to obtain the normalized matrices T, E and the cost matrix D T respectively:

[0143]

[0144]

[0145]

[0146] Step S5: By the Hungarian algorithm, based on the minimum cost device matching for tasks, the best resource is matched for all tasks, and the best unloading scheme for all tasks is obtained:

[0147] This embodiment matches the minimum cost device resource for tasks by the Hungarian algorithm. The Hungarian algorithm is a combinatorial optimization algorithm for solving task allocation problems in polynomial time. In a cloud edge system, in order to achieve the system to complete all tasks with the minimum total cost, each mobile terminal tends to spend the lowest time and energy to unload mobile tasks. This will inevitably lead to a resource competition problem, specifically, it is a non-cooperative competition of multiple tasks for computing resources. To solve the resource competition problem, the Hungarian algorithm is used to match the minimum cost device resource for tasks, to match the best resource for all tasks as a whole, and to obtain the best unloading scheme for all tasks as a whole. The task device resource matching problem can be represented as a triple:

[0148] D T = {T, R, f} (30)

[0149] where T represents the set of tasks participating in the competition, R represents the set of device resources participating in the processing of tasks, and f represents the cost function adopted by all participants, where equation (28) is used as the cost function to calculate the cost value of each device resource processing the current task. Assigning a task to a device resource is referred to as task obtaining a match, and let S = {s1, s2,..., s n} represent the current optimal unloading scheme of the system, s i represents a match of a task to a specific device resource.

[0150] Definition 1: Maximum matching scheme of tasks to device resource set: a scheme in which as many tasks as possible in the task set obtain device resource matching.

[0151] The process of finding the minimum cost maximum matching scheme is to constantly find device resources so that as many tasks as possible in the task set complete unloading and the execution cost is minimized. In the process of making unloading decisions, the system seeks the minimum cost of all tasks and makes a matching scheme through equation (31).

[0152]

[0153] where d i.j represents the value in the comprehensive cost matrix, x i.j represents the decision value, x i.j ∈ {0, 1}, when x i.j is 1, it indicates unloading to this device resource, and C represents the number of device resources of the system. If the current matching scheme S makes:

[0154]

[0155] At this time, there is device resource competition in the matching scheme. The optimal matching for the task set with device resource competition is found through the Hungarian algorithm, and in the matching process, the best scheme for the tasks with conflicts is constantly found, so that as many tasks as possible obtain matching to achieve maximum matching. The conflicting device resources are stored in the set CR, and the unloaded tasks are stored in the set UT. The minimum matching cost value of the unmatching task set is selected through the Hungarian algorithm to solve the conflict , as follows:

[0156]

[0157]

[0158] where represents the D T transformed matrix.

[0159] The found conflict task unloading scheme after solving the conflict is added to S, and other tasks not participating in unloading are recursively searched until the current matching scheme meets:

[0160]

[0161] At this time, the task-to-device resource set matching reaches the maximum matching.

[0162] Determination: The maximum matching of the task-to-device resource set is the optimal matching. The current maximum matching result S is certainly the optimal matching, if device resource competition occurs in the next search process, that is:

[0163]

[0164] At this time, for the matching scheme D T By transformation of formula (33), the transformed result reselects the best unloading scheme to meet:

[0165]

[0166] The best unloading scheme of the current conflict task is added to S. S is updated in real time in the process of continuously searching for new task matching. Therefore, the finally obtained matching is certainly the maximum matching and is optimal.

[0167] Through the above matching theory, the specific steps of finding the best decision S process for all tasks based on the Hungarian algorithm are as follows:

[0168] (1) Initialize the unloading strategy, and randomly select a mobile task from the task node set T Unloading strategy Unloaded task set That is, the unloading strategy is initialized as an empty set, and the unloaded task set is initialized as an empty set.

[0169] (2) Traverse the current task unloadable device resource node set R, and select the minimum cost unloading to update the unloading strategy S; remove the task from the task node set and insert it into the unloaded task set, that is:

[0170] T'←T-{T j}, UT←{T j} (37)

[0171] (3) If unloading to the device resource makes the superposition value of the decision value greater than or equal to 1, that is, it is indicated that there is device resource competition, the Hungarian algorithm is used to solve the resource competition and the optimal unloading scheme S * is selected, and S is updated, that is:

[0172] S←S* = minf(s1, s2,..., s i )(38)

[0173] (4) When the task node set is empty, i.e. , at this time to achieve the maximum number of matches, the best unloading decision S is obtained.

[0174] Table 3 is the pseudo code corresponding to the mobile task unloading algorithm (MCDROA) for minimizing resource cost meeting the deadline requirement.

[0175] Table 3

[0176]

[0177]

[0178] Assuming that the system contains N mobile applications and M computing device resources, the time complexity of the MCDROA algorithm is O(MN+MN 2 ).

[0179] Proof: The MCDROA algorithm mainly consists of two parts: the minimum cost solving part and the task matching device resource problem solving part. First, the MSTPA algorithm is used to calculate the cost of task processing on each resource, and the time complexity is O(MN); then the Hungarian algorithm is used for resource matching, and in each iteration, the matching is increased, so N iterations are needed. In each iteration, when finding a match, each task resource matching is not more than once, so the operation complexity is O(MN), and each time the matching is inserted into the matching scheme, the scheme is updated, and M times of updating relaxation occur in each iteration, and O(N) is needed. The time complexity of this process is O(MN), so the total time complexity of the Hungarian matching process is O(MN 2 ). Thus, the time complexity of the MCDROA algorithm is O(MN+MN 2 ).

[0180] The data generated by simulating the task scene in the evaluation of the resource processing task cost by the multi-criteria objective programming algorithm has been solved by the multi-objective allocation algorithm to obtain the device resource comprehensive cost matrix D T At this time, the process of solving the unloading decision matrix by the Hungarian algorithm is as follows:

[0181]

[0182]

[0183]

[0184] The optimal offloading scheme in the simulation task scenario can be obtained from the finally obtained matrix:

[0185] T1→Ed1, T2→Mb1, T3→Cl1, T4→Mb2, T5→Ed3, T6→Mb3, T7→Ed2, T8→Mb4

[0186] The simulation experiment verification process is as follows:

[0187] (1) Simulation experiment environment:

[0188] 1) Mobile terminal: The mobile terminals in the experiment are A, B, C and D, and the detailed parameter configurations are shown in Table 4. The specific power values are obtained through the power-profile.xml file of the mobile terminal.

[0189] Table 4

[0190]

[0191]

[0192] 2) Edge server: The OpenStack cluster environment with Juno version is built by two servers in the laboratory, and the operating system of the two servers is Ubuntu18.04(64bit) Linux system, the CPU is Intel(R) Xeon(R) Gold5117, the frequency is 2.8GHz, the memory size is 64G, and the hard disk capacity is 2T.

[0193] 3) Network environment: The internal nodes of the edge server are connected by an infinite bandwidth (IB) high-speed switch with a bandwidth of 100Gbps, and the model of the switch is Mellanox QSFP28; the edge server and the external device are connected by an ASUS gigabit Ethernet switch, and the model of the switch is GigaX024X.

[0194] 4) Mobile task program: In this embodiment experiment, target tracking is selected as the mobile task, and the processing process of the task is as shown in Figure 5 , which specifically includes input motion model, feature extraction, observation model, model update, post-processing, result output 6 processing processes. Among them, the input motion model and the result output respectively represent the beginning and the end of the target tracking program, and can only be processed in the mobile terminal.

[0195] (2) Experiment and analysis:

[0196] The proposed MCDROA method is compared with the non-delayed incentive service offloading algorithm (LO-ISO), the Local computing offloading method, the Edge computing offloading method and the Random computing offloading method. The LO-ISO method first estimates the total service delay of the mobile device, and then maximizes the user profit by the incentive service offloading scheme under the consideration of the unique priority of different services. The Edge method processes all tasks on the edge server. The Local method processes all tasks locally. The Random method randomly offloads tasks to any device for processing. In this section, first, the experiment of the influence of the weight factor on the performance of the MCDROA method is performed, and then the experiment of the influence of the number of terminal users, edge server devices and network bandwidth on the performance of the offloading method is performed. The related parameters used in the experiment are set according to the literature, and the specific data are shown in Table 5.

[0197] Table 5

[0198]

[0199]

[0200] 1) Weight factor influence experiment: Figure 6 , Figure 7 and Figure 8 are the processing time, processing energy consumption and total cost of the system for processing 40 random tasks when ω1 takes different values.

[0201] From Figure 6 , Figure 7 and Figure 8 , it can be seen that with the increase of ω1, the computing time, computing energy consumption and system cost of the system are first reduced and then increased, wherein when ω1 takes 0.6, the computing time of the system takes the minimum value; when ω1 takes 0.5, the computing energy consumption and the system cost take the minimum value. From the change trend of the three experimental results, it can be seen that the value of ω1 has a greater influence on the computing time, computing energy consumption and total cost of the system when it is near 0.5.

[0202] 2) Mobile terminal number influence experiment: In this experiment, 8 virtual machines are virtually output from each edge server resource, and the network bandwidth is set to 8 Mbps. The specific experimental results are shown in Figure 9 and Figure 10 .

[0203] From Figure 9 and Figure 10It can be seen that the five methods are basically linearly increased. This is because when the number of mobile terminals increases, the amount of computing resources required by the generated tasks also increases, at which time the competition for channel resources and computing resources by the offloaded tasks is more intense, so that the tasks are not executed in a reasonable offloading order during the offloading process, resulting in excessive waiting time during the transmission and computing processes of the tasks, and thus the delay is increasingly high and increases increasingly fast. Since the energy consumed by the mobile terminal during the waiting process of the tasks is negligible, the total energy consumption of the system of all offloading methods is linearly increased with the increase of the number of mobile terminals. From Figure 9 and Figure 10 It can be seen that the MCDROA method proposed in the embodiment is optimal because the MCDROA algorithm offloads the tasks in the task set at the total minimum cost, considers the resource requirements of the tasks and all available resources as a whole, allocates the most suitable resources to the tasks as a whole, and reduces the competition for resources by the tasks as much as possible, thereby reducing the waiting time caused by the waiting of the tasks.

[0204] 3) The number of edge server virtual machines affects the experiment: 40 mobile terminals are set in the experiment, and the network bandwidth is 8 Mbps. The specific experimental results are shown in Figure 11 and Figure 12

[0205] From Figure 11 and Figure 12 It can be seen that: with the increase of the number of edge server virtual machines, the total delay and total energy consumption of the system of the MCDROA method, the LO-ISO method, the Edge method and the Random method are significantly reduced. Because the number of virtual machines increases, the competition pressure of the tasks for resources decreases, and the edge server can simultaneously satisfy the task requests of multiple terminals. All steps of the Local method are executed locally, and the delay and energy consumption generated by the offloading method are not affected by the edge server virtual machine, but the total delay and energy consumption generated by the offloading scheme are relatively high. It can be seen from the two figures that the time and energy consumption of the MCDROA method proposed in the embodiment are the least, because the LO-ISO non-delay incentive algorithm guarantees the non-delay offloading of the tasks, but because there may be excessive competition for resources among the tasks, the total completion time is relatively long. The Random strategy does not consider the delay and energy consumption requirements of the tasks, and randomly offloads the tasks to the matching resources, and such an offloading scheme has great randomness, and therefore it is difficult to guarantee that the offloading scheme is optimal. When the number of virtual machines increases, the resource competition pressure of the Edge method decreases, but the Edge method does not consider the influence of the channel bandwidth, which may cause the transmission time to be too long. When the number of virtual machines continuously increases, the MCDROA algorithm proposed in the embodiment is more likely to solve the resource competition among the tasks, and can guarantee that the tasks are more reasonably offloaded.

[0206] ​4) Network bandwidth impact experiment: 40 mobile mobile terminals and 16 edge servers are set in this experiment. The experimental results are shown in FIGS. Figure 13 and Figure 14 .

[0207] From Figure 13 and Figure 14 , it can be seen that the latency and energy consumption of the Local offloading method are not affected by the virtual machines of the edge servers. The system total latency and terminal total energy consumption of the MCDROA method, the LO-ISO method, the Edge method, and the Random method all decrease first and then tend to be stable with the increase of the network bandwidth. This is because with the increase of the network bandwidth, the data transmission rate between the mobile mobile terminal and the edge server increases, thereby the task transmission time decreases, more tasks are uploaded to the edge server for data processing, and thus the task completion time and the energy consumption of the mobile terminal are reduced. However, when the transmission rate increases to a certain degree, the influence on the task transmission time is less and less, and thus the performance improvement of the offloading method is gradually weakened. From the two figures, it can be seen that the MCDROA algorithm proposed in this embodiment is optimal in terms of system total latency or system total energy consumption, because the LO-ISO method guarantees the task to be offloaded without time delay through the incentive mechanism, without considering the link resource competition and energy consumption problem. The Random method randomly formulates an offloading scheme, when the network bandwidth increases, the data transmission latency decreases, and the probability of unreasonable offloading scheme decreases, and it cannot guarantee that the offloading scheme is optimal. With the increase of the network bandwidth, the network bandwidth competition pressure continuously decreases, although the Edge method offloads all tasks to the edge server for processing, the resource competition pressure of the edge server does not decrease, resulting in too much waiting time of the tasks in the processing process. The MCDROA algorithm proposed in this embodiment comprehensively considers the time and energy consumption of the task, with the increase of the network bandwidth, the time of uploading the task to the server continuously decreases, thereby the task processing time continuously decreases, more tasks are suitable for processing on the server, and thus the pressure of the terminal device is reduced, and the energy consumption of the task processing continuously decreases.

[0208] Aiming at the limited server resources and multi-constrained target problem of edge computing, MCDROA algorithm is proposed. Firstly, the algorithm gives the resource processing task cost function according to the target programming algorithm to evaluate the cost of resource processing task. Then, in the task offloading process, according to the demand of the task for the resources, the resources set that does not meet the condition is preprocessed, and a benefit function with the minimum total cost of the system is constructed to select the resources set that meets the condition for the task from the resources set. Finally, the Hungarian algorithm is used to match the appropriate resources for the task and constantly solve the resource conflict, so that the offloading scheme meets the minimum total cost of the system. Simulation experiments show that the MCDROA algorithm proposed in this embodiment can reduce the time and energy consumption spent to complete the task with the minimum total cost of the system under the condition of limited server resources, and has good system performance. In the experiment, the algorithm in this embodiment is compared with the three classical algorithms of Random, Edge and Local and the LO-ISO algorithm, which proves the effectiveness of the algorithm in this embodiment.

Claims

1. A task offloading method with minimal resource cost that meets deadline requirements at the cloud, edge, and endpoint, characterized in that: include: Calculate the time and energy consumption for all devices to complete all tasks, and store them in the time matrix and energy consumption matrix respectively; Normalize all elements in the time matrix and energy consumption matrix respectively; Set a target weight vector for each task, wherein the target weight vector includes the target weights of the device for the task's time and energy consumption; The normalized time matrix and energy consumption matrix are weighted and summed using a weighted arithmetic mean operator to calculate the cost of the equipment completing the task. The calculated cost is then stored in the resource cost matrix. Using the Hungarian algorithm, based on matching the device with the minimum cost for the task, the best resources are matched for all tasks to obtain the best unloading scheme for all tasks. The optimal unloading scheme for all tasks is obtained by using the Hungarian algorithm, based on matching the device with the minimum cost for each task, and matching the best resources for all tasks, including: Initialize the uninstallation strategy and randomly select a moving task from the task node set. The uninstallation strategy is initialized to an empty set, and the set of uninstalled tasks is initialized to an empty set. Iterate through the set of unloadable device resource nodes for the current task, select the one with the lowest cost to unload, and update the unloading strategy; delete the task from the task node set and insert it into the set of unloaded tasks; If unloading onto the device resource results in a cumulative decision value greater than or equal to 1, it indicates that there is a resource contention. The Hungarian algorithm is used to resolve the resource contention, select the optimal unloading scheme, and update the unloading strategy. The decision value is obtained as follows: the decision value corresponding to unloading onto the device resource is 1. When the set of task nodes is empty, the maximum number of matches is reached, leading to the optimal uninstallation decision.

2. The task offloading method with minimum resource cost that meets deadline requirements at the cloud, edge, and endpoint according to claim 1, characterized in that, The calculation process for the time and energy consumption of all devices to complete all tasks is as follows: The time required for each mobile terminal to process its mobile tasks, the time required for each mobile terminal's mobile tasks to be offloaded to the corresponding virtual machine for processing, and the time required for the virtual machine to process the tasks when each mobile device's mobile tasks are offloaded to the virtual machine on the cloud server. Then, by combining the corresponding weights, a weighted summation operation is performed to obtain the time for all devices to complete all tasks; The energy consumption generated by each mobile terminal in processing its mobile tasks, the energy consumption generated by each mobile terminal in offloading its mobile tasks to the corresponding virtual machine for processing, and the energy consumption generated by the virtual machine in processing the tasks when each mobile device's mobile tasks are offloaded to the virtual machine on the cloud server; and combined with the corresponding weights, a weighted summation operation is performed to obtain the energy consumption of all devices in completing all tasks. Accordingly, based on the time and energy consumption of all devices to complete all tasks, combined with the total time and energy consumption of all tasks being computed locally, and the proportion of computation latency and computation energy consumption in the system cost, the system cost is obtained. Based on minimizing system cost, the computational unloading decision problem is transformed into a problem of minimizing system cost under constraints.

Images