Efficient multi-target unloading and scheduling optimization method for large-scale DAG tasks

Abstract

The invention discloses an efficient multi-target unloading and scheduling optimization method for a large-scale directed acyclic graph (DAG) task, and belongs to the technical field of mobile edge computing (MEC) and industrial internet of things (IIOT). The method comprises the following steps: constructing a system model containing DAG task topology, delay energy consumption dual-objective optimization and resource competition constraint; a comprehensive multi-objective optimization solution scheme is adopted for solving, a dynamic probability coding mechanism is introduced to replace traditional individual representation, and the high-dimensional decision space search efficiency is improved; designing a cooperative target domain-based decomposition strategy to enhance convergence and diversity; and in combination with a serial active scheduling mechanism based on random weight, generating a task sequence meeting priority constraints. According to the method, the collaborative optimization problem of large-scale dependent tasks under delay, energy consumption and resource competition in an industrial Internet of Things environment is effectively solved, and the task execution efficiency and energy efficiency are remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of industrial Internet of Things (IIoT) and mobile edge computing technology, and in particular to an optimization method and system for efficient and energy-saving computational offloading and resource scheduling of large-scale DAG tasks with complex dependencies in resource-constrained IIoT environments. Background Technology

[0002] The deepening advancement of Industry 4.0 and smart manufacturing has led to the deployment of a large number of sensing and computing devices in the IIoT environment. The computing tasks generated by these devices typically consist of multiple subtasks with a strict execution order, and their dependencies can be modeled as a Directed Acyclic Graph (DAG). Such tasks are common in scenarios such as visual quality inspection and predictive maintenance, and have extremely stringent requirements on task completion latency and device energy consumption.

[0003] Mobile edge computing provides a feasible technical path to meet the aforementioned performance requirements by offloading computing tasks from resource-constrained terminal devices to servers at the network edge. However, achieving efficient DAG task offloading and scheduling in the IIoT environment still faces three core challenges: 1. Complex task topology: Predecessor and successor constraints between subtasks require scheduling to follow a strict execution sequence, resulting in a discrete and complex decision space.

[0004] 2. Multi-objective conflict: Minimizing task completion time and minimizing total system energy consumption are two competing objectives that require effective trade-off optimization.

[0005] 3. Intense resource competition: Edge servers have limited computing resources, and when a large number of terminal devices simultaneously initiate unloading requests, it will trigger intense resource competition.

[0006] To address these challenges, existing technologies primarily employ mathematical programming, heuristic rules, or multi-objective evolutionary algorithms. However, these methods have significant limitations: mathematical programming methods have high computational complexity when dealing with large-scale problems, making them unsuitable for industrial scenarios with high real-time requirements; traditional multi-objective evolutionary algorithms (such as NSGA-II) suffer from difficulties in maintaining solution set diversity, slow convergence speed, and susceptibility to local optima when handling large-scale DAG tasks with hundreds of decision variables, and in particular, lack the ability to efficiently explore high-dimensional discrete decision spaces.

[0007] Therefore, there is an urgent need in this field for a novel optimization scheduling framework that can effectively coordinate task dependencies, multi-objective optimization and resource competition constraints, and has high search efficiency, strong convergence and good scalability. Summary of the Invention

[0008] To overcome the aforementioned deficiencies of the prior art, this invention provides an efficient multi-objective offloading and scheduling optimization method for large-scale DAG tasks, thereby solving the problems existing in the background art.

[0009] This invention provides the following technical solution: an efficient multi-objective offloading and scheduling optimization method for large-scale DAG tasks, comprising the following steps: S1. Construct a large-scale DAG task offloading and resource scheduling model in an industrial mobile edge computing environment. The model aims to minimize the total completion time of all tasks and the total energy consumption of all mobile devices, and includes task priority constraints, edge server capacity constraints, and task soft deadline constraints. S2. A comprehensive multi-objective optimization solution scheme is used to solve the model described in S1 to obtain an approximate Pareto optimal solution set; wherein, a dynamic probability coding mechanism is used to encode individual solutions, and a cooperative objective domain decomposition strategy is adopted. S3. A serial scheduling mechanism based on random weight rules is adopted to decode the unloading decision obtained in S2 into a scheduling table of task start time, transmission time and end time that satisfies the DAG task priority constraints.

[0010] Preferred: The specific steps of step S1 are as follows: (1.1) Define each mobile terminal as a quintuple. ,in express Location-based computing power Indicates in Location Subtask The degree of data correlation It is the device's computing power consumption. Indicates the transmission power used for offloading. Indicates idle power consumption; mobile devices The graphics computation task is represented as, , Indicates task Soft deadline, It is a collection of subtasks. This represents the output data of the task; (1.2) Graph tasks are subject to priority constraints, defined as follows: For subtasks and The collection of all precursor tasks; subtasks The start and end times are respectively used as and express; The completion time of a subtask is defined as: in Indicates the duration of execution; Queuing delay caused by dependency constraints is defined as: It is a congestion scaling factor, subtask The start time is determined in the following way: in express Whether to execute in the same location as the direct predecessor, with a transmission time of , For random network jitter, It is by Dependency delay loss, It is the dependency density coefficient. This represents the earliest available processing execution location; The execution duration of a subtask is defined as: in, Indicates the required CPU cycles. It is a graph complexity factor, through Calculated; Task The completion time is defined as: in, It covers the total propagation delay caused by subtask dependencies; (1.3) Define the device energy consumption model: mobile device Total energy consumption Including energy consumption calculation Transmission energy consumption and idle energy consumption : The energy consumption is calculated using the following formula: in, Subtasks Is it executed locally or uninstalled? The transmission energy consumption of the unloading task is: The idle energy consumption of the edge server during the waiting period when executing tasks is: The total energy consumption of all devices is: (1.4) A bi-objective optimization problem is constructed to minimize the overall task completion time and total energy consumption. The problem is defined as follows: in, Represents the decision vector consisting of the unloading and location decisions of all subtasks; objective function This indicates the total completion time, while This represents the total energy consumption of the mobile device.

[0011] Preferred: The specific steps of step S2 are as follows: (2.1) A comprehensive multi-objective optimization solution scheme (MOQEA / D) is proposed. First, by generating uniformly distributed weight vectors, the multi-objective optimization problem is decomposed into multiple scalar single-objective subproblems, and each individual in the solution set is assigned a subproblem and a neighborhood of a fixed size; (2.2) A dynamic probability coding mechanism is used to encode the weight information, and the selection tendency weight of each decision variable is initialized. The encoded solution information is then used. Represented as: Represents weight information The Middle The preference weights for each decision variable, with values ​​ranging from [value range missing]. ; The length of the weight information is determined through a decoding operation. The solution can be obtained for individual cases. The decoding rules are defined as follows: Indicates in Initialize with uniformly distributed random numbers. and The probability of taking either 0 or 1 is 1 / 2; Choose the tendency weight The update process is as follows:

[0012] It is the rotation angle for selecting the bias weights, and its value range is... ; Defined as: Representative by Encoded solution individuals, Represents the target solution individual; during the selection bias weight update process, when or In some cases, boundary overflow may occur, and boundary handling operations will then be performed: (2.3) A collaborative target domain decomposition strategy is proposed; a set of targets of size is generated. The uniformly weighted vector divides the multi-objective optimization problem into... There are three subproblems with different fitness functions, and Chebyshev aggregation functions are used as the fitness functions. satisfy ,but ; Divide the initial solution set evenly into The system divides the system into subgroups and assigns subproblems with different fitness functions to different subgroups. Finally, it calculates the Euclidean distances between weight vectors and between weight vectors of different classes, and selects the subgroup with the smallest Euclidean distance. The set of indices corresponding to each weight vector, forming subgroups. of Neighborhood set subgroup The corresponding neighborhood is .

[0013] Preferred: The specific steps of step S3 are as follows: (3.1) Generate a scheduling sequence. Based on the predecessor and successor relationships of tasks, generate a random priority matrix using a task sequence generation method based on random priority numbers; for A task, by generating a scale of random priority matrix Each element represents the priority level of the corresponding task, and then Algorithm 1 is used to generate the scheduling sequence. Algorithm 1 initializes the set of executable tasks, iteratively selects the highest priority task, and updates the successor task count and the current set of executable tasks, ultimately generating a task scheduling sequence that satisfies the DAG priority constraint. (3.2) Perform serial scheduling. Based on the scheduling sequence, task data volume, transmission speed, local and edge computing time, and task computing flag, sequentially calculate the transmission time, start time, and end time of each task; wherein, the task start time... Due to the end time of its precursor mission With its own transmission time Decision, end time From start time and local or edge server The calculation time is determined by this.

[0014] The technical effects and advantages of this invention are as follows: Compared with the prior art, the present invention has the following significant advantages: (1) High global search efficiency, suitable for large-scale problems: Due to the adoption of a dynamic probabilistic encoding mechanism, compared with traditional binary encoding, it can represent an exponential solution space with the same solution set size. In the early stage of the search, the algorithm can represent an exponential solution space with linear length weight information, thus achieving effective exploration of high-dimensional decision space with a smaller solution set size, resulting in a better convergence speed. Figure 4 As shown. This is particularly suitable for large-scale DAG tasks with hundreds of variables, such as... Figure 5 As shown, in large-scale task instances, the Pareto front obtained by the multi-objective optimization solution scheme of this invention is significantly better than that of the comparison algorithm.

[0015] (2) High-quality solution set with a balance between convergence and diversity: The introduction of a collaborative target domain decomposition strategy effectively avoids premature convergence by uniformly partitioning the weight vector and allocating subgroups for parallel optimization. Simultaneously, the information exchange mechanism between subgroups ensures global exploration capability, resulting in an approximate Pareto optimal solution set that is not only closer to the true frontier but also more uniformly distributed. The obtained Pareto solution set significantly outperforms algorithms such as MOEA / D in terms of distribution. Figure 5 and Figure 7 As shown in the (HV index).

[0016] (3) Highly practical and capable of handling complex industrial constraints: The designed serial active scheduling mechanism based on random weights can strictly follow the inherent priority constraints of DAG tasks and transform the unloading decision (0 / 1 variable) generated by the optimization algorithm into a specific scheduling scheme (start, transmission, and end time) that can be executed on the time axis, thus meeting the strict requirements of industrial scenarios for the feasibility and executability of scheduling schemes.

[0017] (4) Significant improvement in overall performance. Experimental results show that in multiple IIoT task offloading test instances of different scales, the objective optimization solution proposed in this invention outperforms mainstream comparative algorithms such as NSGA-II and MOEA / D in terms of the mean and variance of the two core performance indicators of multi-objective optimization algorithms, namely inverted generational distance (IGD) and hypervolume (HV). Figure 6 , Figure 7 This demonstrates its significant advantages in comprehensive performance in optimizing both delay and energy consumption. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the DAG task unloading and scheduling system architecture in the IIoT environment described in this embodiment of the invention.

[0020] Figure 2 This is a schematic diagram of the collaborative target domain decomposition strategy described in an embodiment of the present invention.

[0021] Figure 3 This is a flowchart of the overall process of task unloading and scheduling scheme based on the objective optimization solution described in the embodiments of the present invention.

[0022] Figure 4 This is a graph showing the impact of different numbers of sub-problems (N) on algorithm performance (IGD index) as described in the embodiments of the present invention.

[0023] Figure 5 These are comparison diagrams of approximate Pareto fronts obtained by different algorithms under different task scales according to embodiments of the present invention.

[0024] Figure 6 These are the mean and variance of the IGD for solving multi-objective problems as described in this invention example.

[0025] Figure 7 These are the mean and variance of HV for solving multi-objective problems as described in this invention example.

[0026] Figure 8 This is the serial sequence generation process of Algorithm 1 described in the example of this invention.

[0027] Figure 9 This is the serial sequence execution flow of Algorithm 2 as described in the example of this invention. Detailed Implementation

[0028] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these examples are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0029] This invention provides an innovative multi-objective optimization method that solves the latency and energy consumption optimization problems in large-scale DAG task offloading by combining a multi-objective collaborative optimization algorithm, a collaborative optimization strategy, and a serial task scheduling mechanism. Experimental results demonstrate that this invention can significantly improve task offloading efficiency and effectively enhance resource utilization in the Industrial Internet of Things (IIoT), providing a highly efficient solution for intelligent manufacturing, real-time monitoring, and large-scale data processing.

[0030] An efficient multi-objective offloading and scheduling optimization method for large-scale DAG tasks includes the following steps: S1, constructing a large-scale DAG task offloading and resource scheduling model in an industrial mobile edge computing environment. The model takes minimizing the total completion time of all tasks and minimizing the total energy consumption of all mobile devices as optimization objectives, and includes task priority constraints, edge server capacity constraints and task soft deadline constraints. S2. A comprehensive multi-objective optimization solution scheme is used to solve the model described in S1 to obtain an approximate Pareto optimal solution set; wherein, a dynamic probability coding mechanism is used to encode individual solutions, and a cooperative objective domain decomposition strategy is adopted. S3. A serial scheduling mechanism based on random weight rules is adopted to decode the unloading decision obtained in S2 into a scheduling table of task start time, transmission time and end time that satisfies the DAG task priority constraints.

[0031] As a further aspect of the present invention, the specific steps of step S1 are as follows: (1.1) Define each mobile terminal as a quintuple. ,in express Location-based computing power Indicates in Location Subtask The degree of data correlation It is the device's computing power consumption. Indicates the transmission power used for offloading. Indicates idle power consumption; mobile devices The graphics computation task is represented as, , Indicates task Soft deadline, It is a collection of subtasks. This represents the output data of the task; (1.2) Graph tasks are subject to priority constraints, defined as follows: For subtasks and The collection of all precursor tasks; subtasks The start and end times are respectively used as and express; The completion time of a subtask is defined as: in Indicates the duration of execution; Queuing delay caused by dependency constraints is defined as: It is a congestion scaling factor, subtask The start time is determined in the following way: in express Whether to execute in the same location as the direct predecessor, with a transmission time of , For random network jitter, It is by Dependency delay loss, It is the dependency density coefficient. This represents the earliest available processing execution location; The execution duration of a subtask is defined as: in, Indicates the required CPU cycles. It is a graph complexity factor, through Calculated; Task The completion time is defined as: in, It covers the total propagation delay caused by subtask dependencies; (1.3) Define the device energy consumption model: mobile device Total energy consumption Including energy consumption calculation Transmission energy consumption and idle energy consumption : The energy consumption is calculated using the following formula: in, Subtasks Is it executed locally or uninstalled? The transmission energy consumption of the unloading task is:

[0032] The idle energy consumption of the edge server during the waiting period when executing tasks is:

[0033] The total energy consumption of all devices is:

[0034] (1.4) A bi-objective optimization problem is constructed to minimize the overall task completion time and total energy consumption. The problem is defined as follows: in, Represents the decision vector consisting of the unloading and location decisions of all subtasks; objective function This indicates the total completion time, while This represents the total energy consumption of the mobile device.

[0035] As a further aspect of the present invention, the specific steps of step S2 are as follows: (2.1) A comprehensive multi-objective optimization solution scheme (MOQEA / D) is proposed. First, by generating uniformly distributed weight vectors, the multi-objective optimization problem is decomposed into multiple scalar single-objective subproblems, and each individual in the solution set is assigned a subproblem and a neighborhood of a fixed size; (2.2) A dynamic probability coding mechanism is used to encode the weight information, and the selection tendency weight of each decision variable is initialized. The encoded solution information is then used. Represented as:

[0036] Represents weight information The Middle The preference weights for each decision variable, with values ​​ranging from [value range missing]. ; The length of the weight information is determined through a decoding operation. The solution can be obtained for individual cases. The decoding rules are defined as follows:

[0037] Indicates in Initialize with uniformly distributed random numbers. and The probability of taking either 0 or 1 is 1 / 2; Choose the tendency weight The update process is as follows:

[0038] It is the rotation angle for selecting the bias weights, and its value range is... ; Defined as:

[0039] Representative by Encoded solution individuals, Represents the target solution individual; during the selection bias weight update process, when or In some cases, boundary overflow may occur, and boundary handling operations will then be performed:

[0040] (2.3) A collaborative target domain decomposition strategy is proposed; a set of targets of size is generated. The uniformly weighted vector divides the multi-objective optimization problem into... There are three subproblems with different fitness functions, and Chebyshev aggregation functions are used as the fitness functions. satisfy ,but ; Divide the initial solution set evenly into The system divides the system into subgroups and assigns subproblems with different fitness functions to different subgroups. Finally, it calculates the Euclidean distances between weight vectors and between weight vectors of different classes, and selects the subgroup with the smallest Euclidean distance. The set of indices corresponding to each weight vector, forming subgroups. of Neighborhood set subgroup The corresponding neighborhood is .

[0041] As a further aspect of the present invention, the specific steps of step S3 are as follows: (3.1) Generate a scheduling sequence. Based on the predecessor and successor relationships of tasks, generate a random priority matrix using a task sequence generation method based on random priority numbers; for A task, by generating a scale of random priority matrix Each element represents the priority level of the corresponding task, and then Algorithm 1 is used to generate the scheduling sequence. Algorithm 1 initializes the set of executable tasks, iteratively selects the highest priority task, and updates the successor task count and the current set of executable tasks, ultimately generating a task scheduling sequence that satisfies the DAG priority constraint. (3.2) Perform serial scheduling. Based on the scheduling sequence, task data volume, transmission speed, local and edge computing time, and task computing flag, sequentially calculate the transmission time, start time, and end time of each task; wherein, the task start time... Due to the end time of its precursor mission With its own transmission time Decision, end time From start time and local or edge server The calculation time is determined by this.

[0042] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An efficient multi-objective offloading and scheduling optimization method for large-scale DAG tasks, characterized in that, Includes the following steps: S1. Construct a large-scale DAG task offloading and resource scheduling model in an industrial mobile edge computing environment. The model aims to minimize the total completion time of all tasks and the total energy consumption of all mobile devices, and includes task priority constraints, edge server capacity constraints, and task soft deadline constraints. S2. A comprehensive multi-objective optimization solution scheme is used to solve the model described in S1 to obtain an approximate Pareto optimal solution set; wherein, a dynamic probability coding mechanism is used to encode individual solutions, and a cooperative objective domain decomposition strategy is adopted. S3. A serial scheduling mechanism based on random weight rules is adopted to decode the unloading decision obtained in S2 into a scheduling table of task start time, transmission time and end time that satisfies the DAG task priority constraints.

2. The efficient multi-objective offloading and scheduling optimization method for large-scale DAG tasks according to claim 1, characterized in that, The specific steps of step S1 are as follows: (1.1) Define each mobile terminal as a quintuple. ,in express Location-based computing power Indicates in Location Subtask The degree of data correlation It is the device's computing power consumption. Indicates the transmission power used for offloading. Indicates idle power consumption; mobile devices The graphics computation task is represented as, , Indicates task Soft deadline, It is a collection of subtasks. This represents the output data of the task; (1.2) Graph tasks are subject to priority constraints, defined as follows: For subtasks and The set of all predecessor tasks; subtasks The start and end times are respectively used as and express; The completion time of a subtask is defined as: in Indicates the duration of execution; Queuing delay caused by dependency constraints is defined as: It is a congestion scaling factor, subtask The start time is determined in the following way: in express Whether to execute in the same location as the direct predecessor, with a transmission time of , For random network jitter, It is by Dependency delay loss, It is the dependency density coefficient. This represents the earliest available processing execution location; The execution duration of a subtask is defined as: in, Indicates the required CPU cycles. It is a graph complexity factor, through Calculated; Task The completion time is defined as: in, It covers the total propagation delay caused by subtask dependencies; (1.3) Define the device energy consumption model: mobile device Total energy consumption Including energy consumption calculation Transmission energy consumption and idle energy consumption : The energy consumption is calculated using the following formula: in, Subtasks Is it executed locally or uninstalled? The transmission energy consumption of the unloading task is: The idle energy consumption of the edge server during the waiting period when executing tasks is: The total energy consumption of all devices is: (1.4) A bi-objective optimization problem is constructed to minimize the overall task completion time and total energy consumption. The problem is defined as follows: in, Represents the decision vector consisting of the unloading and location decisions of all subtasks; objective function This indicates the total completion time, while This represents the total energy consumption of the mobile device.

3. The efficient multi-objective offloading and scheduling optimization method for large-scale DAG tasks according to claim 1, characterized in that, The specific steps of step S2 are as follows: (2.1) A comprehensive multi-objective optimization solution scheme (MOQEA / D) is proposed. First, by generating uniformly distributed weight vectors, the multi-objective optimization problem is decomposed into multiple scalar single-objective subproblems, and each individual in the solution set is assigned a subproblem and a neighborhood of a fixed size; (2.2) A dynamic probability coding mechanism is used to encode the weight information, and the selection tendency weight of each decision variable is initialized. The encoded solution information is then used. Represented as: Represents weight information The Middle The preference weights for each decision variable, with values ​​ranging from [value range missing]. ; The length of the weight information is determined through a decoding operation. The solution can be obtained for individual cases. The decoding rules are defined as follows: Indicates in Initialize with uniformly distributed random numbers. and The probability of taking either 0 or 1 is 1 / 2; Choose the tendency weight The update process is as follows: It is the rotation angle for selecting the bias weights, and its value range is... ; Defined as: Representative by Encoded solution individuals, Represents the target solution individual; during the selection bias weight update process, when or In some cases, boundary overflow may occur, and boundary handling operations will then be performed: (2.3) A collaborative target domain decomposition strategy is proposed; a set of targets of size is generated. The uniformly weighted vector divides the multi-objective optimization problem into... There are three subproblems with different fitness functions, and Chebyshev aggregation functions are used as the fitness functions. satisfy ,but ; Divide the initial solution set evenly into The system divides the system into subgroups and assigns subproblems with different fitness functions to different subgroups. Finally, it calculates the Euclidean distances between weight vectors and between weight vectors of different classes, and selects the subgroup with the smallest Euclidean distance. The set of indices corresponding to each weight vector, forming subgroups. of Neighborhood set subgroup The corresponding neighborhood is .

4. The efficient multi-objective offloading and scheduling optimization method for large-scale DAG tasks according to claim 1, characterized in that, The specific steps of step S3 are as follows: (3.1) Generate a scheduling sequence. Based on the predecessor and successor relationships of tasks, generate a random priority matrix using a task sequence generation method based on random priority numbers; for A task, by generating a scale of random priority matrix Each element represents the priority level of the corresponding task, and then Algorithm 1 is used to generate the scheduling sequence. Algorithm 1 initializes the set of executable tasks, iteratively selects the highest priority task, and updates the successor task count and the current set of executable tasks, ultimately generating a task scheduling sequence that satisfies the DAG priority constraint. (3.2) Perform serial scheduling. Based on the scheduling sequence, task data volume, transmission speed, local and edge computing time, and task computing flag, sequentially calculate the transmission time, start time, and end time of each task; wherein, the task start time... Due to the end time of its precursor mission With its own transmission time Decision, end time From start time and local or edge server The calculation time is determined by this.

Images