Task scheduling method under heaven and earth edge computing network
Through improved multi-objective evolution algorithm and probability field search strategy, a task scheduling problem model is built, which solves the problem of insufficient task scheduling performance in the world edge computing network, minimizes delay and energy consumption, and improves scheduling quality.
Patent Information
- Application Number
- CN202411067081.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-06
- Publication Date
- 2024-12-06
AI Technical Summary
The prior art is difficult to minimize the completion delay of dependent tasks and the energy consumption of drones in the world edge computing network, and it is easy to fall into local optimization, resulting in poor scheduling performance.
Using an improved multi-objective evolution algorithm based on decomposition, the task-dependent scheduling problem model is constructed, and the task-dependent scheduling order vector and position vector are used as optimization variables, combined with the probability field search strategy, the optimal scheduling scheme is solved to ensure the parallelism of the task. and balance resources.
It effectively avoids premature convergence, realizes the delay of tasks and minimizes energy consumption, improves the task scheduling quality of the world edge computing network, and significantly optimizes the scheduling performance.
Smart Images

Figure CN119095116A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the fields of network communication technology and virtualization technology, and in particular to a task scheduling method in a space-ground edge computing network. Background Art
[0002] The widespread use of ground-based smart devices has promoted the rapid growth of a variety of mobile applications, such as face recognition and augmented reality. These applications are usually computationally intensive and latency-sensitive, which places more stringent requirements on the computing resources of processing devices, while ordinary mobile terminals are unable to cope with them, resulting in an increasingly prominent contradiction between computationally intensive mobile applications and resource-constrained terminal devices, making it a major challenge to achieve the quality of experience expected by end users. Thanks to Mobile Edge Computing (MEC), smart devices can dispatch their applications to edge servers near base stations for offloading execution, thereby reducing energy consumption and improving response speed. However, the communication coverage of base stations is inherently limited, which hinders the connection of smart devices outside the coverage of base stations. In addition, ground base stations are vulnerable to military attacks or natural disasters, resulting in scarce computing resources and degraded service performance.
[0003] To solve the above problems, the integrated space-ground MEC network deploys computing and communication resources on aerial mobile platforms, such as unmanned aerial vehicles (UAVs), to provide convenient and on-demand computing services for smart devices. In recent years, this network has become an important technical support for the 6G network architecture and has been widely used in emergency communications and mission computing scenarios. In particular, drones can be regarded as aerial edge computing platforms that collect and process computing-intensive applications from smart devices. However, due to the limitations of battery technology, size, and payload, the battery capacity and computing power of drones are limited, resulting in limited flight time and application scenarios. Some researchers have integrated high-altitude platforms into the integrated space-ground MEC system to achieve collaborative execution between high-altitude platforms and drones. To improve parallelism and ensure good quality of experience, applications can be divided into multiple interdependent subtasks, which are also called dependent tasks. For example, the face recognition application can be decomposed into five interdependent subtasks: image collection, target detection, preprocessing, feature extraction, and classification, where the target detection subtask can only be executed after the image collection subtask is completed. This division allows some subtasks to be scheduled to the high-altitude platform for remote execution, while the rest are processed locally on the drone. Therefore, the scheduling algorithm must not only determine where the tasks in the application are executed (local drone or high-altitude platform), but also the order in which the tasks are executed. This is because the results of executing some tasks are prerequisites for the execution of other tasks. If the order of task execution is not considered, the application may not successfully complete the execution.
[0004] Optimization objectives such as latency and energy consumption can reflect task scheduling performance from different dimensions. These objectives are often conflicting, which means that improving one objective will deteriorate another. Currently, some studies have used the second-generation non-dominated sorting genetic algorithm (NSGA2) to solve multi-objective dependent task scheduling problems, but the non-dominated sorting genetic algorithm is prone to premature convergence, resulting in poor scheduling performance, such as longer task latency or higher equipment energy consumption. Summary of the invention
[0005] The purpose of this application is to provide a task scheduling method under the space-ground edge computing network, which can simultaneously minimize the task completion delay and drone energy consumption, avoid falling into the problem of local optimality, and improve the task scheduling quality under the space-ground edge computing network.
[0006] To achieve the above objectives, this application provides the following solutions:
[0007] The present application provides a task scheduling method in a space-ground edge computing network, including:
[0008] Taking the scheduling order vectors and scheduling position vectors of all subtasks in the dependent task as optimization variables, and minimizing the average execution delay of the dependent task and the average energy consumption of the drone as the optimization goal, a dependent task scheduling problem model of the target system model is constructed; the dependent task scheduling problem model includes an objective function and constraints; the target system model is a model constructed using a space-ground integrated mobile edge computing network; the target system model consists of a high-altitude platform, several drones and multiple ground equipment; each drone collects a dependent task within a time interval, and the dependent tasks collected by the drone are decomposed into multiple subtasks with dependencies;
[0009] The improved decomposition-based multi-objective evolutionary algorithm is used to solve the dependent task scheduling problem model and obtain an optimal scheduling solution; the optimal scheduling solution includes a scheduling order vector and a scheduling position vector of all subtasks in each dependent task within a time interval.
[0010] According to the specific embodiments provided in this application, this application discloses the following technical effects:
[0011] The present application provides a task scheduling method under a space-ground edge computing network. After constructing a dependent task scheduling problem model of a target system model, the dependent task scheduling problem model is solved by utilizing an improved decomposition-based multi-objective evolutionary algorithm to obtain an optimal scheduling solution. This solves the problem that the existing algorithm is prone to premature convergence, resulting in poor scheduling performance. It achieves the simultaneous minimization of dependent task completion delays and drone energy consumption, avoids falling into local optimal problems, and improves the task scheduling quality under the space-ground edge computing network. BRIEF DESCRIPTION OF THE DRAWINGS
[0012] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0013] Figure 1 This is an application environment diagram of a task scheduling method in a space-ground edge computing network in one embodiment of the present application;
[0014] Figure 2 A flowchart of a task scheduling method in a space-earth edge computing network provided in one embodiment of the present application;
[0015] Figure 3 A schematic diagram of a flow chart of solving a dependent task scheduling problem model using an improved decomposition-based multi-objective evolutionary algorithm provided in an embodiment of the present application;
[0016] Figure 4 A schematic diagram of the topological structure of the dependent task G1 provided in an embodiment of the present application;
[0017] Figure 5 A schematic diagram of a cross-operation process provided by an embodiment of the present application;
[0018] Figure 6 A schematic diagram of a variation operation process provided in an embodiment of the present application;
[0019] Figure 7 A schematic diagram showing the Pareto front comparison results of the task scheduling method in the space-ground edge computing network provided by the present application and the existing NSGA2 algorithm in the scenario (20, 3) provided in one embodiment of the present application;
[0020] Figure 8 A schematic diagram showing the Pareto front comparison results of the task scheduling method in the space-ground edge computing network provided by the present application and the existing NSGA2 algorithm in the scenario (20, 6) provided in one embodiment of the present application;
[0021] Fig. 9 A schematic diagram showing the Pareto front comparison results of the task scheduling method in the space-ground edge computing network provided by the present application and the existing NSGA2 algorithm in the scenario (40, 3) provided in one embodiment of the present application;
[0022] Fig.10 A schematic diagram showing the comparison results of the Pareto front of the task scheduling method in the space-ground edge computing network provided by the present application and the existing NSGA2 algorithm in the scenario (40, 6) provided in one embodiment of the present application;
[0023] Fig.11 A schematic diagram showing the Pareto front comparison results of the task scheduling method in the space-ground edge computing network provided by the present application and the existing NSGA2 algorithm in the scenario (60, 3) provided in one embodiment of the present application;
[0024] Fig.12 This is a schematic diagram of the Pareto front comparison results of the task scheduling method in the space-ground edge computing network provided by this application and the existing NSGA2 algorithm in the scenario (60, 6) provided in one embodiment of the present application. DETAILED DESCRIPTION
[0025] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0026] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.
[0027] The task scheduling method under the ground-to-earth edge computing network provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown in the figure, UAV U1 collects face recognition dependent tasks, and UAV The augmented reality dependent tasks are collected. The smart device collects the images required for the face recognition dependent tasks. The topological structures of the face recognition dependent tasks and the augmented reality dependent tasks are as follows: Figure 1As shown, the face recognition dependent tasks include subtasks such as copying images, image detection, feature fusion, classification recognition and result display. Augmented reality dependent tasks include subtasks such as target tracking, target recognition, perspective transformation and fusion processing. In terms of topological structure, the subtasks are optimized by the task scheduling method under the edge computing network of the sky and the earth proposed in this application. Specifically: the scheduling order vector and scheduling position vector of all subtasks in the dependent task are used as optimization variables, and the average execution delay of the dependent task and the average energy consumption of the drone are minimized as the optimization goal. The dependent task scheduling problem model of the target system model is constructed, and the dependent task scheduling problem model is solved by using the improved decomposition-based multi-objective evolutionary algorithm to obtain the optimal scheduling solution. The optimal scheduling solution includes the scheduling order vector and scheduling position vector of all subtasks, that is, whether each subtask is executed locally by the drone or unloaded. Unloaded execution means unloading to the high-altitude platform for execution.
[0028] Among them, smart devices can be but are not limited to various desktop computers, laptops, smart phones, monitoring equipment, tablet computers and IoT devices portable wearable devices.
[0029] In an exemplary embodiment, Figure 2 As shown, a task scheduling method in a space-ground edge computing network is provided, including the following steps 201 to 202. Among them:
[0030] Step 201: Taking the scheduling order vector and scheduling position vector of all subtasks in the dependent task as optimization variables, and minimizing the average execution delay of the dependent task and the average energy consumption of the UAV as the optimization goal, a dependent task scheduling problem model of the target system model is constructed; the dependent task scheduling problem model includes an objective function and constraints; the target system model is a model constructed using a space-ground integrated mobile edge computing network; the target system model consists of a high-altitude platform, several UAVs and multiple ground equipment; each UAV collects a dependent task within a time interval, and the dependent tasks collected by the UAV are decomposed into multiple subtasks with dependencies.
[0031] Step 202: using an improved decomposition-based multi-objective evolutionary algorithm, the dependent task scheduling problem model is solved to obtain an optimal scheduling solution; the optimal scheduling solution includes a scheduling order vector and a scheduling position vector of all subtasks in each dependent task within a time interval.
[0032] Implement the above steps 201 to 202, and solve the dependent task scheduling problem model by using the improved decomposition-based multi-objective evolutionary algorithm to obtain the optimal scheduling solution, which solves the problem that the existing algorithm is prone to premature convergence, resulting in poor scheduling performance, and achieves the simultaneous minimization of dependent task completion delay and drone energy consumption, avoiding the problem of falling into local optimality, and improving the task scheduling quality under the edge computing network of the sky and the earth. In addition, the present application also introduces the probabilistic field search strategy into the decomposition-based multi-objective evolutionary algorithm to improve its multi-objective optimization performance. Under the premise of satisfying the task dependency constraints, it can better balance the task execution delay and drone energy consumption, and obtain a high-quality non-dominated solution set (optimal scheduling solution), providing decision support for decision makers in the context of multi-party games. Simulation experiments and data analysis show that compared with the second-generation non-dominated genetic algorithm, the method disclosed in the present application has significant optimization in indicators such as anti-generation distance and generation distance, and can obtain better solutions. The Multi-objective Evolutionary Algorithm Based on Decomposition (MOEA / D) decomposes the multi-objective optimization problem into a series of single-objective optimization problems, so that the algorithm can use the structural information of the problem, thereby improving the search efficiency and the quality of the solution. MOEA / D can also cooperate and compete between the various sub-problems, which can better maintain the diversity of the population and avoid falling into the local optimum. Therefore, this application adopts MOEA / D to solve the multi-objective dependent task scheduling problem.
[0033] The following introduces the problem of dependent task scheduling in the ground-to-earth edge computing network. The target system model consists of a high-altitude platform, N uav The system consists of a drone and multiple ground devices, and its characteristics are described as follows: the high-altitude platform floats stably at a fixed height and has sufficient computing resources to provide computing offloading services for drones. Each drone collects a complex dependent task that requires a lot of computing resources within a certain time interval. The dependent tasks collected by the drone can be decomposed into multiple subtasks with dependencies. The subsequent subtasks can only start executing if they depend on the output data of the previous subtasks. A directed acyclic graph is used to represent the dependencies between subtasks. For the high-altitude platform, as long as there is no dependency between the subtasks, multiple subtasks unloaded from the drone can be processed in parallel. i The collected dependent tasks can be expressed as G i =(V i ,ε i ), where V i and ε i Represent the subtask set and edge set respectively, G i Include subtasks; e(v i,l ,vi,k )∈ε i Indicates v i,l and v i,k The dependency between the two subtasks, namely v i,l After execution is completed i,k Each subtask v i,j can be represented as a triple i,j ,O i,j ,B i,j >, where I i,j and O i,j is the input data size and output data size of the task, B i,j The number of CPU cycles required to execute the task, i = 1, ..., N uav ; N uav is the number of drones, For dependent task G i The total number of neutron tasks.
[0034] The calculation model of the drone includes the local calculation model and the scheduling calculation model, as follows:
[0035] 1) Local computing model:
[0036] Assuming that the drone U i Can be represented as a tuple in and For the The maximum computing frequency and maximum power consumption of the heterogeneous cores, and is the transmission power and the receiving power. Assuming that each heterogeneous core has F different computing frequencies, that is, each heterogeneous core corresponds to F different voltage levels, Represents the calculation frequency set. i The current voltage level of the kth heterogeneous core is δ τ (τ=1,...,F), then the actual calculation frequency The corresponding actual power consumption α i,k is a coefficient that depends on the chip structure, and λ is a constant between [2,3].
[0037] If the drone U i Use the maximum computation frequency of the kth heterogeneous core to execute subtask v i,j , which can achieve the minimum execution delay The actual execution delay of the subtask is Then the drone U i Execute v i,j The energy consumption is:
[0038]
[0039] remember and Represents subtask v i,j The completion times in the local calculation queue, the uplink transmission queue, the high-altitude platform queue, and the downlink receiving queue are calculated. If v i,j When executed locally, otherwise UAV i Start executing v i,j Before that, all direct predecessor tasks of the subtask must be completed, then v i,j The preparation time for local execution is:
[0040]
[0041] Among them, pre(v i,j ) is v i,j The set of all direct predecessor tasks.
[0042] 2) Scheduling calculation model:
[0043] UAV i Subtask v needs to be transmitted through the wireless uplink channel i,j The input data is sent to the high-altitude platform, and the preparation time for transmitting the input data can be calculated as
[0044]
[0045] The delay in transmitting the input data is Among them I i,j For subtask v i,j The input data size of R i For UAV i The wireless uplink transmission rate between the high-altitude platform can be calculated by the formula:
[0046]
[0047] Where Ψ represents the bandwidth of the MEC system; σ 2 is the noise power; For UAV i and the high-altitude platform, where g0 represents the channel power gain when the reference distance is 1m, For U i The distance to the high-altitude platform can be obtained based on the coordinates of the drone and the coordinates of the high-altitude platform. i,j The energy consumed by the input data
[0048] When the high-altitude platform receives the input data of the subtask, it can enter the preparation execution phase and execute the subtask v i,j The preparation time can be calculated as
[0049]
[0050] High altitude platform execution subtask v i,j The resulting delay is f hap The computing power of high altitude platforms.
[0051] High altitude platform complete sub-task v i,j The output result of the task can be immediately returned to the drone, and the delay Among them, i,j v i,j The output data size. The corresponding energy consumption Therefore, the drone offloads subtask v i,j Energy consumption for execution on high altitude platform
[0052] use represents the i-th dependent task G i The scheduling order vector of all subtasks in , where s i,t ∈v i is the tth executed subtask; represents the i-th dependent task G i The scheduling position vector of all subtasks in like Then v i,j Execute on UAV i The first i,j Otherwise, it will be dispatched to a high altitude platform for execution.
[0053] Dependent Task G i The completion time of depends on the completion time of all the end tasks in the subtask, which is defined as
[0054]
[0055] Where end(G i ) represents the dependent task G i The end task collection in Indicates the end of the task v i,j The completion time of G i The execution delay is equal to the maximum completion time of all the finished tasks.
[0056] UAV i Complete dependent task Gi The total energy consumption is the sum of the energy consumption generated by executing all subtasks, calculated as
[0057]
[0058] Where E i,j To execute v i,j Energy consumption generated; 1 (Δ) is an indicator function, which is 1 if Δ is true and 0 otherwise.
[0059] In the ground-to-ground edge computing network, two important indicators for measuring task scheduling performance are the execution delay of dependent tasks and the energy consumption of drones. This is because the delay directly determines the completion time of dependent tasks, and energy consumption is particularly important for drones with limited battery power. This application defines the average execution delay of all dependent tasks D avg and the average energy consumption of the drone E avg for
[0060]
[0061] On this basis, S i and L i To optimize the variables, to minimize D avg and E avg To optimize the goal, the objective function of the dependent task scheduling problem model is constructed as follows:
[0062]
[0063] The constraints are as follows:
[0064] st
[0065]
[0066] Among them, constraint C1 represents the subtask v i,j The execution location constraint of the subtask v i,j It can be executed locally on any core of the drone or offloaded to a high-altitude platform for execution. i,j For dependent task G i The jth subtask in . Constraint C2 means that subtask v i,j In UAV i Before starting execution, subtask v i,j All direct predecessor tasks must be on the UAV U i Constraint C3 indicates that subtask v i,j In UAV i Before starting the operation, the UAV i Subtask v must be completely received from the high altitude platform i,jThe output data generated after all direct predecessor tasks are executed. Constraint C4 represents the output data of subtask v i,j Before executing on the high-altitude platform, the high-altitude platform must be i Completely received v i,j Constraint C5 indicates that subtask v i,j Before starting execution on the high-altitude platform, subtask v i,j All direct predecessor tasks must be completed on the high altitude platform; Indicates any.
[0067] In another exemplary embodiment of the present application, Figure 3 As shown, the above step 202 is replaced by the following steps 301 to 303:
[0068] Step 301: Initialize algorithm parameters and set the initial non-dominated solution set to an empty set; the algorithm parameters include a weight vector, a neighbor vector set of each weight vector, the number of individuals in the population, and a reference point; the weight vector represents a vector uniformly distributed in the target domain; the neighbor vector set includes several weight vectors closest to the weight vector; each individual in the population is represented as a scheduling scheme.
[0069] Step 302: Determine whether the current number of iterations is greater than the maximum number of iterations. If the current number of iterations is greater than the maximum number of iterations, the iteration ends and a non-dominated solution set is output. If the current number of iterations is not greater than the maximum number of iterations, execute step 303.
[0070] Step 303: Determine whether the index of the current individual is greater than the number of individuals in the population. If the index of the current individual is greater than the number of individuals in the population, the current number of iterations is increased by 1, the index of the current individual is set to 1, and the process returns to step 302. If the index of the current individual is not greater than the number of individuals in the population, a new individual corresponding to the current individual is generated using a probabilistic field search strategy. Based on the new individual corresponding to the current individual, the reference point and all neighbor individuals of the current individual are updated. The individuals dominated by the new individual corresponding to the current individual are deleted from the current non-dominated solution set. If no individual in the current non-dominated solution set dominates the new individual corresponding to the current individual, the new individual corresponding to the current individual is added to the current non-dominated solution set, the index of the current individual is increased by 1, and the process returns to the step of "determining whether the index of the current individual is greater than the number of individuals in the population".
[0071] The specific process of step 301 is: generate N pop uniformly distributed weight vector in and M is the number of targets. Based on the Euclidean distance, each weight vector w is calculated n xW neighbor vector set Initialize Npop Population of individuals in A i =(S i ,L i ) is the dependent task G i Initialization reference point in Set non-dominated solution set Set the current iteration number g=1 and the index of the current individual n=1.
[0072] If the current number of iterations g≤G max , G max is the maximum number of iterations, execute steps 401 to 404; otherwise, output the current non-dominated solution set
[0073] The above step 303 is replaced by the following steps 401 to 404:
[0074] Step 401: If n≤N pop , execute step 402 to step 404; otherwise set g=g+1 and n=1, and return to step 302.
[0075] Step 402: Generate a new individual y using the probabilistic domain search strategy new , that is, get the new individual corresponding to the current individual.
[0076] Step 403: Based on y new , update the reference point Z ref And all neighbor individuals of subproblem n.
[0077] Step 404: From Delete the y new The dominant individual, if There is no individual dominating y new , then add y new arrive In. Set n=n+1.
[0078] In step 303, a new individual corresponding to the current individual is generated using a probabilistic domain search strategy, which specifically includes steps 501 to 504:
[0079] Step 501: For each weight vector, initialize three individuals y new ,y k and l , generate random numbers.
[0080] Step 502: If the random number is not greater than the set threshold, two indexes k and l are randomly selected from the neighbor vector set corresponding to the weight vector, and y is setk =x k and l =x l .
[0081] Step 503: If the random number is greater than the set threshold, randomly select an index k from the neighbor vector set corresponding to the weight vector and set y k =x k . Calculate the current non-dominated solution set The crowding distance of each individual in is calculated by using the objective function of the task scheduling problem model. Specifically: The number of individuals in is I, and the initialization The crowding distance of the i-th individual in is use express The jth target value of the i-th individual in , i = 1, ..., I, j = 1, ..., M. According to the j-th target Sort the individuals in ascending order and set set up The crowding distance of the i-th individual in Where i = 2, ..., I-1, j = 1, ..., M, and They are The maximum and minimum values of the jth target of all individuals in the solution. Sort all individuals in the current non-dominated solution set in descending order according to the crowding distance of all individuals, and randomly select an individual from the first set number of individuals in the current non-dominated solution set, denoted as z l , and set y l =z l .
[0082] Step 504: y k and l Perform crossover and mutation operations to generate two new offspring individuals and Offspring and are the current individual y k and l The corresponding new individual.
[0083] Step 402 applies the probability domain search strategy to generate a new individual y new , the specific implementation is as follows.
[0084] First, initialize three individuals y new ,y k and l, and then randomly generate a random number rand(0,1) between (0,1). If rand(0,1)≤P sel , then from Randomly select two indexes k and l in , set y k =x k and l =x l .
[0085] If rand(0,1)>P sel , first from Randomly select an index k in y and set y k =x k Then calculate The crowding distance of each individual in Sort all individuals in descending order, and finally Randomly select an individual from the first 20% of individuals in l , and set y l =z l .
[0086] For k and l Perform genetic operations to generate two new offspring individuals and The genetic operation includes a crossover operation and a mutation operation. According to step 301, an individual is composed of all the scheduling schemes of dependent tasks, and a scheduling scheme of dependent tasks includes a scheduling order vector and a scheduling position vector.
[0087] For k and l Perform genetic operations to generate two new offspring individuals and The genetic operation includes crossover operation and mutation operation. This application uses the i-th dependent task G i Take the example to illustrate the process of genetic operation, i = 1, ..., N pop The crossover operation includes the scheduling order vector crossover and the scheduling position vector crossover. The scheduling order vector crossover must take into account the dependencies between subtasks to prevent illegal sub-generation scheduling order vectors from being generated after the crossover. and Represents G i The two parent scheduling order vectors are then selected from the set A random intersection point is generated in . and Respectively and The gene fragment before the intersection point in Delete the All subtasks with the same subtask, and Delete the All subtasks in the same subtask will then have the fragment and snippets Connect to and The front end, that is, the connected child scheduling order vector is and where operator Indicates connecting two vectors. In the crossover operation of scheduling position vectors, a crossover point is randomly generated and the gene fragments before the crossover point are exchanged.
[0088] To more clearly demonstrate the genetic operation process of two individuals, this application assumes that there is only one drone U1 in the target system model, and U1 has three heterogeneous cores. This application needs to schedule the dependent task G1. Figure 4 The topology of G1 is shown. The dependent task contains 10 subtasks. Subtask v1 is the start task, and subtasks v9 and v 10 It is the end of the mission.
[0089] The operation process of scheduling sequence vector crossover and scheduling position vector crossover is as follows: Figure 5 The intersection of the scheduling order vector must take into account the dependencies between the subtasks to prevent illegal sub-generation scheduling order vectors from being generated after the intersection. First, use and Denote the two parent scheduling order vectors of G1, and then randomly generate a crossover point CP1=4 from the set {1,…,10}. and Respectively and The gene fragment before the intersection CP1 in the middle, according to Figure 5 It can be seen that and First, from Delete subtasks v1, v2, v5 and v4 from Delete subtasks v1, v3, v4 and v2, and then add the fragment and snippets Connect to and The front end, that is, the connected child scheduling order vector is and where operator Indicates connecting two vectors. In the crossover operation of the scheduling position vector, a crossover point CP2 is randomly generated, and the gene fragments before the crossover point are exchanged.
[0090] The mutation operation includes scheduling order vector mutation and scheduling position vector mutation. Variation, this application randomly from Select a subtask to be mutated, such as v m . It is worth noting that starting tasks and ending tasks are The position in is fixed, so the subtask to be mutated cannot be the start task or the end task. Search forward, if you encounter a subtask that makes v m All direct predecessor tasks of are contained in the vector , the search ends. Similarly, Search in reverse order, and if a subtask is encountered, v m All direct successor tasks of are contained in the vector In the search, the search is terminated. m , randomly select a position from the remaining subtask position set, and then move v m To this position, the mutation is completed. This application uses a simple bitwise mutation to perform mutation operations on the scheduling position vector. Specifically, starting from the first gene of the execution position vector, a random number between (0, 1) is first randomly generated. If the random number is less than the mutation probability, a number is randomly selected from {0, 1, ..., K} to replace the value of the gene bit.
[0091] The operation process of scheduling order vector mutation and scheduling position vector mutation is as follows: Figure 6 As shown. For the scheduling order vector Variation, this application randomly from Select a subtask to be mutated, such as v2. It is worth noting that the start task and the end task are in The position in is fixed, so the subtask to be mutated cannot be the start task or the end task. Search forward, if a subtask is encountered, all direct predecessor subtasks of v2 are included in the vector , the search ends. Similarly, If a subtask is encountered, all the direct successor subtasks of v2 are included in the vector In addition to the position of v2, a position is randomly selected from the position set {2,3,5} of the remaining subtasks (in Figure 6Select position 3), then move v2 to position 3, and the mutation of v2 is completed. This application uses a simple bitwise mutation to perform mutation operations on the scheduling position vector. Specifically, it starts from the first gene of the execution position vector. First, a random number between (0,1) is randomly generated. If the random number is less than the mutation probability, a number is randomly selected from {0,1,2,3} to replace the value of the gene position. Figure 6 In the example, the genes at position 2 and position 6 (i.e., 2 and 3) need to be mutated. After the mutation operation, the values of the two gene positions become 0 and 2 respectively.
[0092] In step 403, based on y new To update the reference point Z ref and all neighbor individuals of subproblem n, as follows. First, calculate individual y new The target vector F(y new )=(f 1 (y new ),…,f M (y new )),if Then set In a specific example of this application, the optimization goal M=2 (including the average execution delay of all dependent tasks and the average energy consumption of the drone), so the individual y new The target vector F(y new )=(f 1 (y new ),f 2 (y new ), j = 1, 2. That is, f 1 (y new ) can be the average execution delay of all dependent tasks, f 2 (y new ) can be the average energy consumption of the UAV.
[0093] For the update of neighbor individuals of the nth subproblem, the Chebyshev decomposition method is used to define the nth subproblem as:
[0094]
[0095] For all neighbor individual indices of subproblem n If g te (y new |w q ,Z ref )≤g te (x q |w q ,Z ref ), then set x q =y new , that is, replace the initial individual with ynew , and set F(x q )=F(y new ).
[0096] In order to verify the effectiveness of the task scheduling method in the edge computing network provided by this application, the experimental parameters and test scenarios are configured. Assume that there are N uav = 3 UAVs, the horizontal and vertical coordinates of each UAV are randomly generated from [200m, 200m], and the coordinates of the high-altitude platform are set to (100m, 100m, 20km). Table 1 shows the setting values of other simulation parameters.
[0097] To test the validity of this application, G i The number of subtasks in and the number of heterogeneous cores K are used as two varying system parameters to generate test scenarios. and K∈{3,6}, six test scenarios can be generated, as shown in Table 2. When K=3, set The corresponding maximum power consumption Similarly, when K = 6, specify and And k = 1, ..., 6. Among them, heterogeneous core 1 is a high-performance core, and heterogeneous core K is a low-performance core.
[0098] Table 1 Simulation parameter settings
[0099]
[0100] Table 2 Test scenarios
[0101]
[0102]
[0103] This application compares the following two multi-objective algorithm performance indicators with the NSGA2 algorithm: Inverted Generational Distance (IGD) and Generational Distance (GD). IGD can simultaneously measure the diversity and convergence of a non-dominated solution set obtained by an algorithm, while GD can measure how close the non-dominated solution set obtained by the algorithm is to the real non-dominated solution set. The smaller the value of both IGD and GD indicators, the better the multi-objective performance of the algorithm. The NSGA2 algorithm and the improved decomposition-based multi-objective evolutionary algorithm provided in this application were each run independently 20 times, and then the IGD and GD values of the 20 times were averaged.
[0104] Table 3 shows the IGD and GD values of NSGA2 and the present application in six test scenarios. It can be found that the present application can obtain smaller IGD and GD values in all six test scenarios. This verifies the effectiveness and feasibility of the present application method. Figure 7-Figure 12 As shown, the Pareto fronts obtained by the two algorithms in six test scenarios are plotted. The horizontal axis is the average execution delay, and the vertical axis is the average energy consumption. Obviously, the Pareto fronts obtained by this application in the six test scenarios have good convergence and diversity, indicating that the distribution of the non-dominated solution set obtained by the algorithm is more uniform, the ductility is better, and the solution quality is better.
[0105] Table 3 IGD and GD results (optimal values are in bold)
[0106]
[0107] The present application also provides an application scenario, which applies the above-mentioned task scheduling method under the edge computing network of the sky and the earth. Specifically: the task scheduling method under the edge computing network of the sky and the earth provided in this embodiment can be applied in the task scheduling scenario. The task scheduling scenario includes a system construction link and a task scheduling process; the target system model enters the task scheduling process from the system construction link to obtain the corresponding optimal scheduling method. The task scheduling method under the edge computing network of the sky and the earth provided in this embodiment belongs to the task scheduling process. Specifically, in the task scheduling process for the target system model, the scheduling order vector and scheduling position vector of all subtasks in the dependent task can be used as optimization variables, and the minimization of the average execution delay of the dependent task and the average energy consumption of the drone can be used as the optimization goal to construct a dependent task scheduling problem model of the target system model, and the improved decomposition-based multi-objective evolutionary algorithm is used to solve the dependent task scheduling problem model to obtain the optimal scheduling solution.
[0108] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0109] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.
Claims
1. A task scheduling method in a space-ground edge computing network, characterized in that: The task scheduling method under the ground-earth edge computing network includes: Taking the scheduling order vectors and scheduling position vectors of all subtasks in the dependent task as optimization variables, and minimizing the average execution delay of the dependent task and the average energy consumption of the drone as the optimization goal, a dependent task scheduling problem model of the target system model is constructed; the dependent task scheduling problem model includes an objective function and constraints; the target system model is a model constructed using a space-ground integrated mobile edge computing network; the target system model consists of a high-altitude platform, several drones and multiple ground equipment; each drone collects a dependent task within a time interval, and the dependent tasks collected by the drone are decomposed into multiple subtasks with dependencies; The improved decomposition-based multi-objective evolutionary algorithm is used to solve the dependent task scheduling problem model and obtain an optimal scheduling solution; the optimal scheduling solution includes a scheduling order vector and a scheduling position vector of all subtasks in each dependent task within a time interval.
2. The task scheduling method in the space-ground edge computing network according to claim 1 is characterized in that: The improved decomposition-based multi-objective evolutionary algorithm is used to solve the dependent task scheduling problem model and obtain the optimal scheduling solution, including: Step 301: Initialize algorithm parameters and set the initial non-dominated solution set to an empty set; the algorithm parameters include a weight vector, a neighbor vector set of each weight vector, the number of individuals in the population, and a reference point; the weight vector represents a vector uniformly distributed in the target domain; the neighbor vector set includes several weight vectors closest to the weight vector; each individual in the population is represented as a scheduling scheme; Step 302: determine whether the current number of iterations is greater than the maximum number of iterations. If the current number of iterations is greater than the maximum number of iterations, the iteration ends and a non-dominated solution set is output. If the current number of iterations is not greater than the maximum number of iterations, execute step 303. Step 303: Determine whether the index of the current individual is greater than the number of individuals in the population. If the index of the current individual is greater than the number of individuals in the population, the current number of iterations is increased by 1, the index of the current individual is set to 1, and the process returns to step 302. If the index of the current individual is not greater than the number of individuals in the population, a new individual corresponding to the current individual is generated using a probabilistic field search strategy. Based on the new individual corresponding to the current individual, the reference point and all neighbor individuals of the current individual are updated. The individual dominated by the new individual corresponding to the current individual is deleted from the current non-dominated solution set. If no individual in the current non-dominated solution set dominates the new individual corresponding to the current individual, the new individual corresponding to the current individual is added to the current non-dominated solution set, the index of the current individual is increased by 1, and the process returns to the step of "determining whether the index of the current individual is greater than the number of individuals in the population".
3. The task scheduling method in the sky-ground edge computing network according to claim 2 is characterized in that: The new individual corresponding to the current individual is generated using the probabilistic domain search strategy, including: For each weight vector, initialize three individuals y new ,y k and l , generate random numbers; If the random number is not greater than the set threshold, two indexes k and l are randomly selected from the neighbor vector set corresponding to the weight vector, and y is set k =x k and l =x l ; If the random number is greater than the set threshold, a random index k is selected from the neighbor vector set corresponding to the weight vector, and y is set k =x k , calculate the crowding distance of each individual in the current non-dominated solution set, sort all individuals in the current non-dominated solution set in descending order according to the crowding distance of all individuals, and randomly select an individual from the first set number of individuals in the current non-dominated solution set, denoted as z l , and set y l =z l ;The crowding distance of individuals is calculated using the objective function of the task scheduling problem model; For k and l Perform crossover and mutation operations to generate two new offspring individuals and Offspring and are the current individual y k and l The corresponding new individual.
4. The task scheduling method in the sky-ground edge computing network according to claim 1, characterized in that: The objective function is as follows: Among them, S i represents the i-th dependent task G i The scheduling order vector of all subtasks in, i=1,...,N uav ; N uav is the number of drones; L i represents the i-th dependent task G i The scheduling position vector of all subtasks in D avg is the average execution delay of all dependent tasks; E avg is the average energy consumption of all drones; D(G i ) is the dependent task G i The completion time; end(G i ) indicates the dependent task G i The end task collection in Indicates the end of subtask v i,j The completion time of v i,j For dependent task G i The jth subtask in For dependent task G i The total number of neutron tasks; E(G i ) is the UAV i Complete dependent task G i Total energy consumption; E i,j To execute v i,j The energy consumption generated; For UAV i Use the kth heterogeneous core to execute subtask v i,j The energy consumption generated; Offload subtask v to drone i,j Energy consumption for execution on high altitude platform; 1 (Δ) is an indicator function, which is 1 if Δ is true and 0 otherwise; l i,j is the scheduling position vector of the jth subtask; For UAV i A heterogeneous set of cores on i,j =0 means dispatching to a high altitude platform for execution.
5. The task scheduling method in the space-ground edge computing network according to claim 4 is characterized in that: The energy consumption of the UAV’s heterogeneous cores executing subtasks is calculated as follows: Among them, δ τ For UAV i The current voltage level of the kth heterogeneous core; λ is a constant; is the maximum power consumption of the kth heterogeneous core; For UAV i Use the maximum computation frequency of the kth heterogeneous core to execute subtask v i,j Minimum execution delay.
6. The task scheduling method in the space-ground edge computing network according to claim 4, characterized in that: The energy consumption calculation formula of unloading subtasks from the drone to the high-altitude platform is as follows: in, Transmit subtask v for drone i,j The energy consumed by the input data; The UAV receives the subtask v from the high-altitude platform i,j The energy consumed by the output data.
Citation Information
Cited By
Unmanned aerial vehicle data collection and energy transmission method
CN119556717A
A method for collecting data and transmitting energy from unmanned aerial vehicles
CN119556717B
