Multi-target task scheduling method and device based on cloud edge cooperative computing and medium
By improving the TWO_ARCHIVE_DS algorithm and combining dual search and information compensation mechanisms, the problems of multi-objective conflict and heterogeneous resource distribution in cloud-edge collaborative computing are solved, generating a high-quality Pareto optimal solution set and improving the efficiency and adaptability of task scheduling.
Patent Information
- Application Number
- CN202510974721.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-15
- Publication Date
- 2025-11-11
AI Technical Summary
In cloud-edge collaborative computing environments, existing technologies struggle to effectively address task scheduling issues arising from multi-objective conflicts, heterogeneous resource distribution, and Pareto front complexity. In particular, when edge node resources are limited and unevenly distributed, traditional scheduling algorithms cannot accurately quantify multi-objective interactions and struggle to generate balanced scheduling schemes.
An improved TWO_ARCHIVE_DS algorithm is adopted, which introduces a dual search mechanism and an information compensation mechanism to dynamically update the convergence and diversity archives. By combining two-point mapping cross and simulated binary cross, the task allocation is optimized to generate a Pareto optimal solution set, thereby achieving multi-objective collaboration and rational resource allocation.
It improves the quality of task scheduling solution sets, enhances the adaptability of the algorithm in dynamic edge environments, effectively alleviates node load pressure, and improves resource utilization efficiency and the feasibility of scheduling schemes.
Smart Images

Figure CN120929206A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of multi-objective optimization technology, and in particular to a multi-objective task scheduling method, device and medium based on cloud-edge collaborative computing. Background Technology
[0002] With the rapid development of the Internet of Things (IoT), mobile internet, and network physical systems, various devices and data are growing at an alarming rate. Traditional centralized service architectures based on cloud computing face enormous pressure when processing massive amounts of data, especially due to long transmission distances and network instability, hindering the normal operation of high-demand applications and services. To alleviate this pressure, edge computing has emerged. Compared with cloud computing, edge computing features low latency, high real-time performance, and mobility support, allowing many tasks that were originally executed in the cloud to be offloaded to edge nodes closer to users. A typical edge computing architecture consists of IoT devices, edge nodes, and the cloud. As an intermediate layer, the edge nodes typically have limited and heterogeneous storage and computing resources, meaning there are differences in resource types, capacity, and performance. In this environment, how to optimally allocate limited heterogeneous resources and ensure the effective processing of different application tasks becomes the key to improving the overall performance of edge computing, making the optimization of task scheduling strategies at the edge layer crucial.
[0003] Current research on edge computing task scheduling primarily focuses on objectives such as load balancing and energy optimization on edge servers. However, existing technical solutions have significant shortcomings in addressing the complex needs of real-world scenarios, mainly in the following two stages:
[0004] During the model building phase, the inherent complexity of edge computing and the limited and heterogeneous resources of edge nodes present modeling challenges. When tasks have high computational demands or resource requirements, a single edge node may not be able to handle the load, necessitating collaboration among multiple edge nodes. Simultaneously, the uneven distribution of resources among nodes directly impacts overall computational efficiency. Therefore, effectively improving task execution efficiency hinges on how to rationally offload tasks to appropriate nodes (local, edge, or cloud) and construct models that accurately quantify edge computing scheduling requirements (e.g., considering latency, energy consumption, resource utilization, and other dimensions simultaneously). Existing models struggle to comprehensively characterize these interconnected and potentially conflicting scheduling needs.
[0005] In the task scheduling phase, task scheduling in an edge computing environment is essentially a multi-objective optimization problem, such as minimizing latency, reducing energy consumption, and balancing load, and these objectives often conflict with each other. Furthermore, the geographical and resource distribution of edge nodes is typically uneven. This multi-objective conflict and uneven node distribution require that the solutions generated by the scheduling algorithm, i.e., the scheduling schemes, be able to adaptively adjust according to the actual distribution of nodes and have good distribution on the Pareto optimal front (i.e., provide a set of high-quality solutions that balance multiple objectives). However, traditional task scheduling algorithms, especially the widely used method of transforming multiple objectives into a single objective solution through weighted summation, struggle to effectively address this problem: on the one hand, the weighted summation method cannot accurately quantify the complex interactions of multiple objectives during scheduling, making it difficult to obtain a truly balanced solution; on the other hand, such methods are usually unable to adequately adapt to the dynamic environment of heterogeneous and unevenly distributed nodes, and are also unable to cope with the complex discontinuities that may exist on the Pareto front. Therefore, selecting or designing a scheduling algorithm that can effectively solve such multi-dimensional objective models, alleviate objective conflicts, and ensure the uniformity and adaptability of the solution set distribution is extremely challenging.
[0006] Chinese patent application CN116719612A discloses a cloud resource scheduling optimization method, device, and storage medium. It optimizes the system energy consumption, resource loss, load balancing, and turnaround time of the cloud resource scheduling system as optimization objectives. It mathematically models the virtual machine placement problem, defines resource constraint formulas, and establishes a multi-objective constraint optimization model for resource scheduling in a cloud computing environment. This addresses the problem of deteriorated Pareto front differentiation capability caused by a high proportion of non-dominated solutions in high-dimensional objectives and irregular Pareto fronts. However, this application does not address the heterogeneity of edge nodes, task priority, and dynamic adaptability in cloud-edge collaborative scenarios. Its dual-archive algorithm lacks adaptability in handling multi-objective conflicts and complex Pareto fronts in edge computing, failing to adapt to the limited and unevenly distributed resources of edge nodes. Therefore, in cloud-edge computing environments, overcoming the obstacles of model construction accuracy and task scheduling algorithm adaptability, effectively solving the problems caused by multi-objective conflicts, heterogeneous resource distribution, and the complexity of Pareto fronts, and improving task scheduling efficiency are technical problems that need to be solved. Summary of the Invention
[0007] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a multi-objective task scheduling method, device and medium based on cloud-edge collaborative computing. It adopts an improved TWO_ARCHIVE_DS algorithm, which adaptively updates the convergence archive and diversity archive according to the indicators, and migrates boundary solution individuals between the convergence archive and diversity archive through an information compensation mechanism, thereby achieving the effect of multi-objective collaboration and improving the quality of allocation scheme.
[0008] The objective of this invention can be achieved through the following technical solutions:
[0009] According to one aspect of the present invention, a multi-objective task scheduling method based on cloud-edge collaborative computing is provided, the specific steps of which include:
[0010] S1. Collect the pending task flow from the local device; calculate the urgency value of each task based on the task's waiting time, remaining execution time limit, and generation timestamp; sort the tasks according to the urgency value and generate a task queue with priority labels.
[0011] S2. Read the task mapping constraints and set the optimization objectives as minimizing task transmission time, task processing time, total system energy consumption, and maximizing load balancing between nodes. For each task in the task queue, based on its allocatable nodes (including edge nodes and cloud nodes), use the encoding formula to convert it into the corresponding gene fragment to form the genotype representation of the initial population.
[0012] S3. Initialize the node population, determine the mapping relationship from tasks to edge nodes or cloud nodes, and generate an initial solution; use the improved TWO_ARCHIVE_DS algorithm with the initial solution, task mapping constraints, and optimization objective input to calculate the objective function value, and perform fitness evaluation, selection, crossover, and mutation operations based on the objective function value to generate a new offspring population until the termination condition is reached; the improvement includes the introduction of a dual search mechanism and an information compensation mechanism, and the use of a strategy combining two-point mapping crossover and simulated binary crossover as the crossover operator;
[0013] S4. Receive the Pareto optimal solution set output by the improved TWO_ARCHIVE_DS algorithm. The Pareto optimal solution set includes a task-node mapping relationship list and a corresponding target value vector. According to the preset decision preference rules, select the optimal task-node mapping relationship list from the solution set as the final task scheduling scheme for output and execution.
[0014] Furthermore, in S3, the specific steps of the TWO_ARCHIVE_DS algorithm for iterative evolution of the population include:
[0015] Generate a uniformly distributed set of weight vectors, randomly initialize the population, and obtain the genotypes of the initial task allocation scheme; create empty convergence archives and diversity archives;
[0016] Select parent task allocation schemes from convergent and diverse archives, apply simulated binary crossover operator to modify the gene segments of their task allocation nodes, apply polynomial mutation operation to randomly change some gene segments, and generate new offspring task allocation schemes.
[0017] For each child scheme, perform a decoding operation to obtain the corresponding task-node mapping relationship list and calculate its multi-objective function value vector; based on the cosine of the angle between the multi-objective function value vector calculated for each child scheme and each vector in the weight vector set, divide each child into the subpopulation corresponding to the weight vector with the smallest angle.
[0018] Calculate the first distance of each offspring scheme to its respective subpopulation weight vector and the second distance to the hyperplane constructed from the objective function values of all schemes in the current population. Obtain the allocation of each offspring scheme in the convergence archive and diversity archive through the dual search mechanism.
[0019] Using an information compensation mechanism, each optimization objective dimension is compared between the convergent archive and the diversity archive, and individuals in the convergent archive and the diversity archive are migrated and replaced. When the termination condition is met, the convergent archive and the diversity archive are merged as the Pareto optimal solution set and output.
[0020] Furthermore, the specific steps of the dual search mechanism include:
[0021] The current schemes in the convergence archive are sorted using the first distance; the adaptive coefficient of the convergence archive is calculated based on the sorting, and then the fitness of the schemes in the convergence archive with the current child scheme is obtained, i.e., the first fitness. Based on the fitness, it is determined whether the current child scheme should enter the convergence archive.
[0022] The existing schemes in the diversity archive are sorted using the second distance; the adaptation coefficient of the diversity archive is calculated based on the sorting, and then the fitness of the schemes in the diversity archive with the current offspring scheme is obtained, i.e., the second fitness. The fitness is used to determine whether the current offspring scheme should enter the diversity archive.
[0023] Furthermore, the first fitness (CA) i The expression is obtained by sorting the current schemes in the convergence archive using the quartiles:
[0024]
[0025] Fitness (CA) i )=k c *d1+(1-k c )d2,
[0026] Where, k c For the convergence archive adaptive coefficients, in the order of the convergence archive, the first individual value is Q0, the 0.25Nth individual value is Q1, the 0.5Nth individual value is Q2, the 0.75Nth individual value is Q3, and the last individual value is Q4. N is the population size, d1 is the first distance, and d2 is the second distance.
[0027] The second fitness (DA) i The expression is derived from the quartiles of existing schemes in the diversity archive, after sorting them.
[0028]
[0029] Fitness (DA) i )=(1-k d )d1+k d *d2,
[0030] Where, k d For the diversity archive adaptive coefficient, in the sorting of the diversity archive, the first individual has a value of Q5, the 0.25Nth individual has a value of Q6, the 0.5Nth individual has a value of Q7, the 0.75Nth individual has a value of Q8, and the last individual has a value of Q9. N is the population size, d1 is the first distance, and d2 is the second distance.
[0031] Furthermore, the expressions for the first distance and the second distance are:
[0032] d1=F(x)*W j / ||W j ||,
[0033]
[0034] Where d1 is the first distance, d2 is the second distance, F(x) is the target value vector corresponding to individual x, and W j This is the weight vector corresponding to the subpopulation to which individual x belongs.
[0035] Furthermore, the specific steps of the information compensation mechanism include:
[0036] For the current individual, compare the optimal solution in the convergence archive and the diversity archive to optimize the target dimension. If the convergence of the optimal solution in the convergence archive is better than the corresponding optimal solution in the diversity archive, then the solution is migrated from the convergence archive to the diversity archive. If the convergence of the optimal solution in a certain target dimension in the diversity archive is better than the corresponding optimal solution in the convergence archive, then the solution is migrated from the diversity archive to the convergence archive.
[0037] Furthermore, the crossover in S3 includes task exchange crossover and task allocation crossover. Task exchange crossover involves selecting two parent task allocation schemes, randomly selecting two crossover points, exchanging the task sequence fragments between the crossover points in the two parent schemes, keeping the task ID unchanged, and generating a new offspring scheme. Task allocation crossover involves selecting two parent task allocation schemes, allocating gene fragments to the nodes corresponding to each task in the schemes, and performing arithmetic crossover operations based on the crossover probability and distribution index to generate a new offspring scheme.
[0038] Furthermore, the mutations in S3 include task swapping mutations and task allocation mutations. In a single task allocation scheme, the task swapping mutation randomly selects two gene fragments representing different tasks at different locations and swaps their positions. In a single task allocation scheme, the task allocation mutation randomly selects one or more node allocation gene fragments corresponding to tasks and randomly changes their values to a new node ID that meets the constraints.
[0039] According to a second aspect of the present invention, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the method described thereon.
[0040] According to a third aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method described thereon.
[0041] Compared with the prior art, the present invention has the following beneficial effects:
[0042] (1) Effectively coordinate and optimize multiple objectives and improve the quality of solutions: By constructing a multi-task scheduling model with the optimization objectives of minimizing transmission time, processing time, total system energy consumption and maximizing load balancing between nodes, and using the improved TWO_ARCHIVE_DS algorithm for solution, a dual search mechanism is introduced. The distance from an individual to the weight vector and the distance to the hyperplane are calculated as dual indices, and adaptive coefficients dynamically calculated based on these indices are used to form convergence fitness and diversity fitness, which guide the updating of convergence archive and diversity archive respectively. At the same time, the information compensation mechanism achieves the complementary advantages of convergence and diversity by migrating the boundary solution individuals with the best performance in each objective dimension between the convergence archive and the diversity archive. This mechanism effectively balances the conflicts between multiple objectives such as latency, energy consumption, and load, and overcomes the limitation of traditional weighted sum methods in accurately quantifying the interaction of multiple objectives. It can generate a well-distributed and widely covered Pareto optimal solution set in a cloud-edge collaborative heterogeneous environment, thereby improving the quality of the solution set.
[0043] (2) Enhancing the dynamic adaptability and rationality of resource utilization of the algorithm: In view of the characteristics of limited, heterogeneous and unevenly distributed edge node resources, the TWO_ARCHIVE_DS algorithm incorporates an information compensation mechanism. By comparing the optimal boundary solutions of the convergence archive and the diversity archive in each optimization objective dimension, individuals are dynamically migrated between the two archives according to their relative advantages in convergence or diversity. This migration utilizes the guiding characteristics of the boundary solutions, enabling the algorithm to adaptively adjust the search direction according to the population evolution state and the uneven distribution of node resources, thereby enhancing the algorithm's adaptability in dynamic edge environments. Combined with the priority ranking based on task waiting time, remaining time and generation time, as well as the explicit load balancing requirements in the optimization objective, the algorithm ensures that system resources can be allocated more rationally and efficiently according to the urgency of tasks and the node status, effectively alleviating the problem of uneven node load pressure.
[0044] (3) Improve the feasibility and evolution efficiency of the scheduling scheme: The genetic operation design is closely integrated with the task scheduling application scenario. Two-point mapping crossover is used to handle the task execution order, and simulated binary crossover is used to handle the task-to-node allocation. The mutation operation is also designed for task order and node allocation respectively. This special operator strategy effectively balances the feasibility of the solution and the search efficiency. The task queue is transformed into gene fragments through encoding operation, and the gene fragments are restored into a specific task-node mapping relationship list through decoding operation. This enables the genetic algorithm to directly operate on the core elements of the scheduling scheme and accurately calculate the objective function value during evaluation. The combination of this encoding and decoding mechanism and special operators significantly improves the efficiency of generating feasible and efficient scheduling schemes during the evolution process. Attached Figure Description
[0045] Figure 1 This is a flowchart of a multi-objective task scheduling method based on cloud-edge collaborative computing;
[0046] Figure 2 This is the population representation of the multi-objective task scheduling method based on cloud-edge collaborative computing in this embodiment;
[0047] Figure 3 This is a schematic diagram illustrating task switching at edge nodes;
[0048] Figure 4 This is a schematic diagram illustrating task allocation on edge nodes.
[0049] Figure 5 This is a schematic diagram of the dual search mechanism of the TWO_ARCHIVE_DS algorithm in the convergence direction.
[0050] Figure 6 This is a schematic diagram of the dual search mechanism of the TWO_ARCHIVE_DS algorithm in the direction of diversity.
[0051] Figure 7 This is a schematic diagram of individual migration in the information compensation mechanism of the TWO_ARCHIVE_DS algorithm. Detailed Implementation
[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0053] In cloud-edge collaborative computing task scheduling, a cloud-edge collaborative computing model is used as a premise. This model consists of local devices, device agents, a fog layer, and a cloud layer. Local devices are responsible for continuously transmitting the sensed data, and each group of devices connects to a device agent. The device agent is responsible for communication between layers, processing sensor signals requested by user devices, and its main function is to determine the priority of input tasks by identifying critical and urgent tasks. Non-urgent or non-critical tasks are transmitted to the cloud for processing, while urgent tasks are immediately transmitted to the nearest fog node for processing. Edge nodes are responsible for executing tasks transmitted to them. After a task is transmitted to an edge node, the edge node receives a set of ordered tuples.<L,T,R> Here, L represents local device nodes and global (edge and cloud) nodes, T is the task flow, and R is resource availability. Based on this tuple, edge nodes can quickly determine whether a task should be processed on the edge node or transmitted to the cloud. Cloud nodes are responsible for receiving and processing tasks uploaded to the cloud. The method in this embodiment can effectively solve multi-objective problems with uneven shape distribution in the real world, and can solve the task scheduling problem in edge computing, achieving multi-objective optimization.
[0054] like Figure 1 The diagram illustrates a multi-objective task scheduling method based on cloud-edge collaborative computing. The specific steps include:
[0055] S1. Collect the pending task flow from the local device; calculate the urgency value of each task based on the task's waiting time, remaining execution time limit, and generation timestamp; sort the tasks according to the urgency value and generate a task queue with priority labels.
[0056] S2. Read the task mapping constraints and set the optimization objectives as minimizing task transmission time, task processing time, total system energy consumption, and maximizing load balancing between nodes. For each task in the task queue, based on its allocatable nodes (including edge nodes and cloud nodes), use the encoding formula to convert it into the corresponding gene fragment to form the genotype representation of the initial population.
[0057] S3. Initialize the node population, determine the mapping relationship from tasks to edge nodes or cloud nodes, and generate an initial solution; use the improved TWO_ARCHIVE_DS algorithm with the initial solution, task mapping constraints, and optimization objective input to calculate the objective function value, and perform fitness evaluation, selection, crossover, and mutation operations based on the objective function value to generate a new offspring population until the termination condition is reached; improvements include the introduction of a dual search mechanism and an information compensation mechanism, and the use of a strategy combining two-point mapping crossover and simulated binary crossover as the crossover operator;
[0058] S4. Receive the Pareto optimal solution set output by the improved TWO_ARCHIVE_DS algorithm. The Pareto optimal solution set includes a list of task-node mapping relationships and the corresponding target value vector. According to the preset decision preference rules, select the optimal task-node mapping relationship list from the solution set as the final task scheduling scheme for output and execution.
[0059] The specific steps of the TWO_ARCHIVE_DS algorithm for iterative evolution of the population include:
[0060] Generate a uniformly distributed set of weight vectors, randomly initialize the population, and obtain the genotypes of the initial task allocation scheme; create empty convergence archives and diversity archives;
[0061] Select parent task allocation schemes from convergent and diverse archives, apply simulated binary crossover operator to modify the gene segments of their task allocation nodes, apply polynomial mutation operation to randomly change some gene segments, and generate new offspring task allocation schemes.
[0062] For each child scheme, perform a decoding operation to obtain the corresponding task-node mapping relationship list and calculate its multi-objective function value vector; based on the cosine of the angle between the multi-objective function value vector calculated for each child scheme and each vector in the weight vector set, divide each child into the subpopulation corresponding to the weight vector with the smallest angle.
[0063] Calculate the first distance of each offspring scheme to its respective subpopulation weight vector and the second distance to the hyperplane constructed from the objective function values of all schemes in the current population. Obtain the allocation of each offspring scheme in the convergence archive and diversity archive through the dual search mechanism.
[0064] Using an information compensation mechanism, each optimization objective dimension is compared between the convergent archive and the diversity archive, and individuals in the convergent archive and the diversity archive are migrated and replaced. When the termination condition is met, the convergent archive and the diversity archive are merged as the Pareto optimal solution set and output.
[0065] In the TWO_ARCHIVE_DS algorithm, the initialization of the node population involves generating uniformly distributed weight vectors. Based on the current number of targets M, the sampling quantity H in the target dimension is determined, with a sampling step size of 1 / H, uniformly sampling across the M-dimensional space. Newly generated offspring require generation into a subpopulation. For each solution in the subpopulation, the nearest weight vector is found based on the angle between the solution and the weight vector. Individuals belonging to the same weight vector are considered to be in the same subpopulation. The formula for calculating the angle cosθ is as follows:
[0066] cosθ=F(x)*W j / |F(x)|*|W j |i=1,2,3…N,
[0067] Where x represents an individual in the population, F(x) is the target value vector corresponding to that individual, and W... j Let be the weight vector in the target space. cosθ is the cosine of the angle between the individual and the weight vector; the larger the value, the smaller the angle between the individual and the weight.
[0068] The specific steps of the dual search mechanism include:
[0069] The current schemes in the convergence archive are sorted using the first distance; the adaptive coefficient of the convergence archive is calculated based on the sorting, and then the fitness of the schemes in the convergence archive with the current child scheme is obtained, i.e., the first fitness. Based on the fitness, it is determined whether the current child scheme should enter the convergence archive.
[0070] The existing schemes in the diversity archive are sorted using the second distance; the adaptation coefficient of the diversity archive is calculated based on the sorting, and then the fitness of the schemes in the diversity archive with the current offspring scheme is obtained, i.e., the second fitness. The fitness is used to determine whether the current offspring scheme should enter the diversity archive.
[0071] First fitness (CA) i The expression is obtained by sorting the current schemes in the convergence archive using the quartiles:
[0072]
[0073] Fitness (CA) i )=k c *d1+(1-k c )d2,
[0074] Where, k c For the convergence archive adaptive coefficients, in the order of the convergence archive, the first individual value is Q0, the 0.25Nth individual value is Q1, the 0.5Nth individual value is Q2, the 0.75Nth individual value is Q3, and the last individual value is Q4. N is the population size, d1 is the first distance, and d2 is the second distance.
[0075] Second fitness (DA) i The expression is derived from the quartiles of existing schemes in the diversity archive, after sorting them.
[0076]
[0077] Fitness (DA) i )=(1-k d )d1+k d *d2,
[0078] Where, k d For the diversity archive adaptive coefficient, in the sorting of the diversity archive, the first individual has a value of Q5, the 0.25Nth individual has a value of Q6, the 0.5Nth individual has a value of Q7, the 0.75Nth individual has a value of Q8, and the last individual has a value of Q9. N is the population size, d1 is the first distance, and d2 is the second distance.
[0079] The expressions for the first distance and the second distance are:
[0080] d1=F(x)*W j / ||W j ||,
[0081]
[0082] Where d1 is the first distance, d2 is the second distance, F(x) is the target value vector corresponding to individual x, and W j This is the weight vector corresponding to the subpopulation to which individual x belongs.
[0083] The specific steps of the information compensation mechanism include:
[0084] For the current individual, compare the optimal solution in the convergence archive and the diversity archive to optimize the target dimension. If the convergence of the optimal solution in the convergence archive is better than the corresponding optimal solution in the diversity archive, then the solution is migrated from the convergence archive to the diversity archive. If the convergence of the optimal solution in a certain target dimension in the diversity archive is better than the corresponding optimal solution in the convergence archive, then the solution is migrated from the diversity archive to the convergence archive.
[0085] Crossover in S3 includes task exchange crossover and task allocation crossover. Task exchange crossover involves selecting two parent task allocation schemes, randomly selecting two crossover points, exchanging the task sequence fragments between the crossover points in the two parent schemes, keeping the task IDs unchanged, and generating a new offspring scheme. Task allocation crossover involves selecting two parent task allocation schemes, allocating gene fragments to the nodes corresponding to each task in the schemes, and performing arithmetic crossover operations based on the crossover probability and distribution index to generate a new offspring scheme.
[0086] Mutations include task swapping mutations and task assignment mutations. In a single task assignment scheme, task swapping mutations randomly select two gene fragments representing different tasks at different locations and swap their positions. Task assignment mutations, in a single task assignment scheme, randomly select gene fragments corresponding to one or more tasks and randomly change their values to a new node ID that meets the constraints.
[0087] The mutation process of task exchange and task allocation first ensures that each gene segment appears only once in the permutation sequence, and the permutation order of the gene segment is random. Then, two genes are randomly selected and exchanged. After the exchange is completed, a gene is randomly selected from the genes that were not exchanged and mutated. The result of the mutated gene may be the original gene or a new gene segment.
[0088] The method in this embodiment uses latency and execution time, SLA violation rate, and resource load as evaluation metrics, specifically including:
[0089] Execution time: The time it takes for a task to complete, which most directly reflects the speed at which the algorithm executes.
[0090] Delay time: Delay time is the difference between the time a task is completed and its expected completion time. The larger the delay time, the worse the performance.
[0091] SLA violation rate: This represents the proportion of tasks that violated SLAs during the scheduling process out of all tasks submitted by the user.
[0092] Service cost: The resources consumed by the algorithm during operation, mainly including the resources consumed in transmission, latency, processing, etc., as well as the energy consumption caused by violating SLO.
[0093] Resource load: Defined as the percentage of load on physical hosts, virtual machines, containers, etc. It is a measure of the number of tasks waiting to be executed on the CPU queue and the number of tasks currently running.
[0094] Population representation of a multi-objective task scheduling method based on cloud-edge collaborative computing, such as Figure 2As shown. Since task scheduling in edge computing can be considered an optimization problem, a directed graph is used to represent the search space. Assume a job (consisting of a set of tasks) is represented by a graph that starts with a root node and ends with a leaf node. Each task in the job is represented by a set of intermediate nodes, with the root node representing the start of the task and the leaf node representing the end of the task. Therefore, in the directed graph, a path from the root to a leaf forms a solution, represented by a string. For example, given a job i, which consists of tasks... i1 , task i2 , task i3 Given the composition, there are multiple possibilities for scheduling these tasks across m edge nodes. For example, one solution configuration is that the job is scheduled as follows: on FN3 On FN7 On FN1 like Figure 2 As shown. Each group has N individuals, which is the solution. Each individual represents a set of edge nodes, and the definition of each set of edge nodes FNTasks is as follows:
[0095] FNTasks={FN1Tasks,FN2Tasks,…,FN m Tasks}
[0096] Each edge node FN j Tasks are responsible for executing the corresponding tasks. The edge nodes are defined as follows:
[0097]
[0098] The TWO_ARCHIVE_DS algorithm randomly selects N individuals to form a population. The N selected individuals are defined as follows:
[0099]
[0100] Figure 3 To illustrate task exchange using TMX on edge nodes, Parent1 and Parent2 are selected as parent individuals and TMX operations are performed. The specific process is as follows: two gene fragments are randomly selected to obtain the crossover subsequence in Parent1; the reference sequence of the crossover gene is found in Parent2 and mapped to the new subsequence; the offspring Offspring1 is constructed, whose subsequence outside the crossover gene fragment of Parent1 and inside the crossover gene fragment of Parent2 remains unchanged; the same process is performed to generate the offspring Offspring2.
[0101] Figure 4This diagram illustrates task allocation using SBX on edge nodes. In the TWO_ARCHIVE_DS algorithm, SBX is used as the crossover operator for task allocation on edge nodes, requiring the generation of a random probability value for each task. When the random probability of a task is not less than 0.5, the {2, 1} gene fragment from Parent1 is selected to construct Offspring1 at its original position; otherwise, the {3, 4, 2} gene fragment from Parent1 is selected to construct Offspring2 at its original position, and the process is similar for Parent2.
[0102] Figure 5 and Figure 6 This diagram illustrates the dual search mechanism of the TWO_ARCHIVE_DS algorithm. Since different archives have different search focuses, search weights are introduced to combine the two objectives of diversity and convergence into a trade-off index, designed as an individual fitness function. This is achieved through a linear combination of different weights to emphasize either convergence or diversity. When calculating the dual index, such as... Figure 5 As shown, the convergence direction index is derived from the distance of the individual to the hyperplane, which is constructed by calculating the minimum and maximum points based on the individual distribution; as... Figure 6 As shown, the diversity orientation index is derived from the distance from the individual to the weight vector.
[0103] Figure 7 This diagram illustrates the individual migration mechanism of the TWO_ARCHIVE_DS algorithm's information compensation mechanism. After each generation of evolution, boundary solutions are migrated between two archives. The optimal solutions x and y for each target dimension are determined in CA and DA, respectively. In the current target dimension, the two individuals with the best performance are compared. If an individual shows a more significant advantage, individual migration is required.
[0104] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the described module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.
[0105] The electronic device of this invention includes a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in read-only memory (ROM) or loaded from a storage unit into random access memory (RAM). The RAM may also store various programs and data required for device operation. The CPU, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.
[0106] Multiple components in the device are connected to an I / O interface, including: input units such as a keyboard, mouse, etc.; output units such as various types of displays, speakers, etc.; storage units such as disks, optical disks, etc.; and communication units such as network interface cards, modems, wireless transceivers, etc. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks. The processing unit performs the various methods and processes described above, such as the method of the present invention. For example, in some embodiments, the method of the present invention may be implemented as a computer software program tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program may be loaded and / or installed on the device via ROM and / or the communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more steps of the method of the present invention described above may be performed. Alternatively, in other embodiments, the CPU may be configured to execute the method of the present invention by any other suitable means (e.g., by means of firmware).
[0107] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SoCs), Complex Programmable Logic Devices (CPLDs), and so on.
[0108] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.
[0109] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.
[0110] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A multi-objective task scheduling method based on cloud-edge collaborative computing, characterized in that, The specific steps include: S1. Collect the pending task flow from the local device; calculate the urgency value of each task based on the task's waiting time, remaining execution time limit, and generation timestamp; sort the tasks according to the urgency value and generate a task queue with priority labels. S2. Read the task mapping constraints and set the optimization objectives as minimizing task transmission time, task processing time, total system energy consumption, and maximizing load balancing between nodes. For each task in the task queue, based on its allocatable nodes (including edge nodes and cloud nodes), use the encoding formula to convert it into the corresponding gene fragment to form the genotype representation of the initial population. S3. Initialize the node population, determine the mapping relationship from tasks to edge nodes or cloud nodes, and generate an initial solution; use the improved TWO_ARCHIVE_DS algorithm with the initial solution, task mapping constraints, and optimization objective input to calculate the objective function value, and perform fitness evaluation, selection, crossover, and mutation operations based on the objective function value to generate a new offspring population until the termination condition is reached; the improvement includes the introduction of a dual search mechanism and an information compensation mechanism, and the use of a strategy combining two-point mapping crossover and simulated binary crossover as the crossover operator; S4. Receive the Pareto optimal solution set output by the improved TWO_ARCHIVE_DS algorithm. The Pareto optimal solution set includes a task-node mapping relationship list and a corresponding target value vector. According to the preset decision preference rules, select the optimal task-node mapping relationship list from the solution set as the final task scheduling scheme for output and execution.
2. The multi-objective task scheduling method based on cloud-edge collaborative computing according to claim 1, characterized in that, In S3, the specific steps of the TWO_ARCHIVE_DS algorithm for iterative evolution of the population include: Generate a uniformly distributed set of weight vectors, randomly initialize the population, and obtain the genotypes of the initial task allocation scheme; create empty convergence archives and diversity archives; Select parent task allocation schemes from convergent and diverse archives, apply simulated binary crossover operator to modify the gene segments of their task allocation nodes, apply polynomial mutation operation to randomly change some gene segments, and generate new offspring task allocation schemes. For each child scheme, perform a decoding operation to obtain the corresponding task-node mapping relationship list and calculate its multi-objective function value vector; based on the cosine of the angle between the multi-objective function value vector calculated for each child scheme and each vector in the weight vector set, divide each child into the subpopulation corresponding to the weight vector with the smallest angle. Calculate the first distance of each offspring scheme to its respective subpopulation weight vector and the second distance to the hyperplane constructed from the objective function values of all schemes in the current population. Obtain the allocation of each offspring scheme in the convergence archive and diversity archive through the dual search mechanism. Using an information compensation mechanism, each optimization objective dimension is compared between the convergent archive and the diversity archive, and individuals in the convergent archive and the diversity archive are migrated and replaced. When the termination condition is met, the convergent archive and the diversity archive are merged as the Pareto optimal solution set and output.
3. The multi-objective task scheduling method based on cloud-edge collaborative computing according to claim 2, characterized in that, The specific steps of the dual search mechanism include: The current schemes in the convergence archive are sorted using the first distance; the adaptive coefficient of the convergence archive is calculated based on the sorting, and then the fitness of the schemes in the convergence archive with the current child scheme is obtained, i.e., the first fitness. Based on the fitness, it is determined whether the current child scheme should enter the convergence archive. The existing schemes in the diversity archive are sorted using the second distance; the adaptation coefficient of the diversity archive is calculated based on the sorting, and then the fitness of the schemes in the diversity archive with the current offspring scheme is obtained, i.e., the second fitness. The fitness is used to determine whether the current offspring scheme should enter the diversity archive.
4. The multi-objective task scheduling method based on cloud-edge collaborative computing according to claim 3, characterized in that, The first fitness (CA) i The expression is obtained by sorting the current schemes in the convergence archive using the quartiles: Fitness(CA i )=k c *d1+(1-k c )d2, Where, k c For the convergence archive adaptive coefficients, in the order of the convergence archive, the first individual value is Q0, the 0.25Nth individual value is Q1, the 0.5Nth individual value is Q2, the 0.75Nth individual value is Q3, and the last individual value is Q4. N is the population size, d1 is the first distance, and d2 is the second distance. The second fitness (DA) i The expression is derived from the quartiles of existing schemes in the diversity archive, after sorting them. Fitness(YES i )=(1-k d )d1+k d *d2, Where, k d For the diversity archive adaptive coefficient, in the sorting of the diversity archive, the first individual has a value of Q5, the 0.25Nth individual has a value of Q6, the 0.5Nth individual has a value of Q7, the 0.75Nth individual has a value of Q8, and the last individual has a value of Q9. N is the population size, d1 is the first distance, and d2 is the second distance.
5. A multi-objective task scheduling method based on cloud-edge collaborative computing according to claim 2, characterized in that, The expressions for the first distance and the second distance are: d1=F(x)*W j / ||W j ||, Where d1 is the first distance, d2 is the second distance, F(x) is the target value vector corresponding to individual x, and W j This is the weight vector corresponding to the subpopulation to which individual x belongs.
6. A multi-objective task scheduling method based on cloud-edge collaborative computing according to claim 2, characterized in that, The specific steps of the information compensation mechanism include: For the current individual, compare the optimal solution in the convergence archive and the diversity archive to optimize the target dimension. If the convergence of the optimal solution in the convergence archive is better than the corresponding optimal solution in the diversity archive, then the solution is migrated from the convergence archive to the diversity archive. If the convergence of the optimal solution in a certain target dimension in the diversity archive is better than the corresponding optimal solution in the convergence archive, then the solution is migrated from the diversity archive to the convergence archive.
7. A multi-objective task scheduling method based on cloud-edge collaborative computing according to claim 1, characterized in that, The crossover in S3 includes task exchange crossover and task allocation crossover. Task exchange crossover involves selecting two parent task allocation schemes, randomly selecting two crossover points, exchanging the task sequence fragments between the crossover points in the two parent schemes, keeping the task ID unchanged, and generating a new offspring scheme. Task allocation crossover involves selecting two parent task allocation schemes, allocating gene fragments to the nodes corresponding to each task in the schemes, and performing arithmetic crossover operations based on the crossover probability and distribution index to generate a new offspring scheme.
8. A multi-objective task scheduling method based on cloud-edge collaborative computing according to claim 1, characterized in that, The mutations in S3 include task swapping mutations and task allocation mutations. In a single task allocation scheme, the task swapping mutation randomly selects two gene fragments representing different tasks at different locations and swaps their positions. In a single task allocation scheme, the task allocation mutation randomly selects one or more task-related node allocation gene fragments and randomly changes their values to a new node ID that meets the constraints.
9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 8.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 8.
Citation Information
Patent Citations
Cloud resource scheduling optimization method and device and storage medium
CN116719612A
