A Multi-Workflow Scheduling Method Based on an Improved Gradient Optimizer

By improving the multi-workflow scheduling method of the gradient optimizer, and utilizing K-means location clustering and adaptive mutation strategy, the multi-workflow scheduling in the cloud environment is optimized, which solves the problems of resource waste and low scheduling efficiency of existing algorithms in the cloud environment, and achieves more efficient multi-objective optimization.

CN116225658BActive Publication Date: 2026-11-10BEIJING INST OF TECH
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202310261140.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-03-17
Publication Date
2026-11-10
Estimated Expiration
2043-03-17

AI Technical Summary

Technical Problem

Existing multi-workflow scheduling algorithms in cloud environments struggle to simultaneously optimize the interests of cloud service providers and consumers. Furthermore, existing multi-objective optimization algorithms find relatively few non-dominated solutions, and their distributions are far from the true Pareto front, leading to wasted computing resources and low scheduling efficiency.

Method used

We adopt a multi-workflow scheduling method based on an improved gradient optimizer, which divides the population into multiple clustered populations through K-means location clustering. By combining evolutionary direction awareness, adaptive mutation and elite crossover strategies, we optimize the multi-objective multi-population evolutionary framework, improve search diversity and computational resource utilization.

Benefits of technology

It improves the search diversity and computing resource utilization of multi-workflow scheduling in the cloud environment, finds more non-dominated solutions, optimizes the total completion time and execution cost of workflows, and enhances global and local search capabilities.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116225658B_ABST
    Figure CN116225658B_ABST
Patent Text Reader

Abstract

This invention discloses a multi-workflow scheduling method based on an improved gradient optimizer. Addressing the computational resource waste caused by the equal distribution of particles in existing multi-population multi-objective (MPMO) frameworks, this invention proposes a multi-objective multi-population evolutionary framework based on particle position clustering to solve the multi-workflow scheduling problem in cloud environments. Through K-means-based position clustering, the entire population is divided into multiple cluster populations. Each cluster population selects an optimization objective based on its own fitness level, which not only improves the diversity of the search but also avoids cross-evolution, explores more promising regions, finds more non-dominated solutions, and effectively improves the generation efficiency of workflow scheduling schemes.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of multi-workflow scheduling technology in cloud computing environments, specifically relating to a multi-workflow scheduling method based on an improved gradient optimizer. Background Technology

[0002] Cloud computing is a large-scale, heterogeneous, distributed computing infrastructure in the scientific and commercial fields. It provides users with high-quality, low-cost services on demand, and its elastic scalability and pay-as-you-go features have led to its continuous expansion and deepening application across various industries. Workflow is a common technique for building large-scale computing and data-intensive distributed applications from different research fields. With the continuous development and numerous advantages of cloud computing, more and more enterprises and organizations are migrating or deploying their workflow applications to cloud data centers for execution.

[0003] Multi-workflow scheduling in a cloud environment aims to map a set of workflow tasks to a suitable set of cloud service resources, finding the optimal workflow scheduling method to minimize the execution time and cost of multiple workflow applications while simultaneously meeting the Quality of Service (QoS) requirements of multiple users. A suitable multi-workflow scheduling algorithm should be able to improve the benefits for the cloud service provider while ensuring the QoS requirements of multiple users. Therefore, multi-workflow scheduling in a cloud environment is a constrained multi-objective optimization problem.

[0004] Currently, workflow scheduling problems are mainly solved using heuristic algorithms and metaheuristic algorithms. Heuristic algorithms are optimization algorithms based on rules or expert experience, designed for a specific type of problem, such as the Heterogeneous Earliest Finish Time (HEFT) algorithm, the Budget-Constrained Heterogeneous Earliest Finish Time (BHEFT) algorithm, and the Max-Min algorithm. These algorithms are easy to implement and fast, but they cannot guarantee the quality of the solution and are not suitable for solving complex or large-scale scheduling problems. Metaheuristic algorithms, on the other hand, are based on random search techniques, such as Genetic Algorithms (GA), Particle Swarm Optimization (PSO), and Ant Colony Optimization (ACO). They have better applicability, are generally not dependent on a specific problem, and can find an approximate optimal solution within a finite time, thus solving single-objective optimization problems. However, application scheduling in cloud environments often needs to consider the interests of both cloud service providers and consumers simultaneously, meaning that multiple objectives often need to be optimized concurrently. Existing multi-objective optimization algorithms, such as the Multi-Objective Heterogeneous Earliest Complete Time (MOHEFT) algorithm, the Non-Dominated Sorting Genetic Algorithm (NSGA II), the Multi-Population Co-evolutionary Particle Swarm Optimization (CMPSO) algorithm, and the Multi-Population Ant Colony Optimization (MOACS) algorithm, have improved upon HEFT, GA, PSO, and ACO algorithms, respectively, making them suitable for multi-objective workflow scheduling problems. However, these algorithms find relatively few non-dominated solutions, and their distribution is far from the true Pareto front. Therefore, the performance of existing metaheuristic-based multi-objective workflow scheduling algorithms needs further improvement. Summary of the Invention

[0005] In view of this, the present invention provides a multi-workflow scheduling method based on an improved gradient optimizer, which realizes the scheduling of multi-workflow tasks in a cloud environment.

[0006] This invention provides a multi-workflow scheduling method based on an improved gradient optimizer, comprising the following steps:

[0007] Step 1: Model the input multi-workflow tasks with deadline constraints and the available virtual machine resources; the multi-workflow tasks are represented as follows: For the i-th workflow, Workflow set Total number of workflows included; virtual machine resources are represented as v f For the f-th virtual machine;

[0008] Step 2: Assign sub-deadlines to all subtasks of the workflow, and determine the scheduling sequence by prioritizing the task with the earliest deadline; evaluate the quality of infeasible solutions based on the constraint satisfaction evaluation index Θ, where Θ represents the sum of the percentages of workflows exceeding constraints; the position of each individual in the population is X = (x1, x2, ..., ..., x...). N ) represents a scheduling scheme, where the index of each dimension element represents a workflow task number, and the value of each dimension element represents the virtual machine assigned to the task; nine predefined initial particles are inserted into the population, and the remaining individuals in the population are initialized using a random generation method; individuals that satisfy the constraints among the nine predefined initial particles are inserted into the global external archive set to complete the initialization of the global external archive set.

[0009] Step 3: If the current iteration number n satisfies in To determine the iteration threshold, the K-means location clustering algorithm is used to divide the data into K cluster subpopulations P1,...,P1. m ,...,P K Otherwise, proceed to step 5;

[0010] Step 4: Based on the population P1,...,P m ,...,P K The total fitness (makespan) and cost of all particles are used to determine the most suitable evolutionary direction ψ1,...,ψ m ,...,ψ K ;

[0011] Step 5: Store the global non-dominated solutions of the population obtained during the evolution process in the global external archive set, and then guide the update of each cluster population through the elite individuals in the global external archive set.

[0012] Step 6: After a set number of iterations, the number of mutation dimensions η and the mutation rate ρ are adaptively updated according to the population evolution speed. In each iteration, a mutation operation is performed on the individuals in the global external archive set according to the current number of mutation dimensions η and the mutation rate ρ, and the global external archive set is updated using the mutation results.

[0013] Step 7: Use the selected set of excellent parent generations to perform cross operations, and copy the mapping sequence of the better task to the resource corresponding to each sub-workflow to the child generation in turn. Generate child individuals through cross operations based on elite inheritance, and update the global external archive set with the generated new child individuals.

[0014] Step 8: Increment the iteration count n by 1. If n is not less than the maximum iteration count, output the current global external archive set as the result and end this process; otherwise, go to step 3.

[0015] Furthermore, the virtual machines are numbered in order of their computing power.

[0016] Furthermore, in step 2, sub-deadlines are assigned to all tasks in the workflow, and the scheduling sequence is determined by prioritizing the task with the earliest deadline.

[0017] Task Sub-deadline for:

[0018]

[0019] in, For entry task The upward rank value, For workflow Deadline constraints;

[0020] Task Upward rank value for:

[0021]

[0022] Where pm is the resource pool The fastest processing speed of a virtual machine in China. For the task Instruction length, To the task To its sub-task The size of the transmitted data is denoted by b, where b is the network bandwidth between different virtual machines.

[0023] Furthermore, the constraint satisfaction evaluation index Θ is calculated as follows:

[0024]

[0025] Among them, M i and D i The first i Total execution span and deadline constraints for each workflow.

[0026] Furthermore, in step 2, evaluating the merits of infeasible solutions based on the constraint satisfaction evaluation index Θ includes: for two infeasible solutions X1 and X2, the following condition is met: When X1 is better than X2, C(X1) and C(X2) represent the workflow execution costs corresponding to scheduling schemes X1 and X2, respectively, and Θ(X1) and Θ(X2) represent the degree of constraint exceedance of scheduling schemes X1 and X2, respectively.

[0027] Furthermore, the nine predefined initial particles are defined as follows: by randomly assigning all tasks to virtual machines of the same type to generate a scheduling scheme, six predefined scheduling schemes i1 to i6 can be obtained from the six types of virtual machines; then, a predefined scheduling scheme i7 is generated using the HEFT algorithm; first, two tasks scheduled to different types of virtual machines in i7 are randomly selected, and their assigned virtual machines are swapped to generate a new predefined scheduling scheme i8; then, all tasks in i7 that were assigned to the virtual machine with the largest processing capacity are assigned to the virtual machine with the largest processing capacity per unit cost to generate a new predefined scheduling scheme i9. Thus, the predefined scheduling schemes i1 to i9 are the nine predefined initial particles.

[0028] Furthermore, in step 3, the K-means location clustering algorithm is used to divide it into K cluster subpopulations P1,...,P m ,...,P K The following method is used to calculate the distance between individuals in the population:

[0029] The q-th individual in the population With the wth individual The Euclidean distance d between them q,w for in, For the qth individual Values ​​of each dimension For the w-th individual Values ​​of each dimension For workflow The total number of tasks.

[0030] Furthermore, in step 4, based on the population P1,...,P m ,...,P K The total fitness (makespan) and cost of all particles are used to determine the most suitable evolutionary direction ψ1,...,ψ m ,...,ψ K The method is as follows:

[0031] Calculate the rank of each particle in the cluster relative to the makespan of the entire population, and obtain the average rank of all particles in the cluster; based on the K averages of the K clusters, calculate the rank of the clusters P1,...,P... m ,...,P K Sort the data to obtain each cluster population P. m In P1,...,P m ,...,P K Ranking in If P m Ranking Less than the preset Makespan ranking threshold h M Then in P m The set of evolutionary directions ψ m Add makespan to the middle;

[0032] Calculate the rank of each particle in the cluster with respect to cost within the entire population, and obtain the average rank of all particles in that cluster; based on the K averages of the K clusters, perform a multiplication of the clusters P1,...,P... m ,...,P K Sort the data to obtain each cluster population P. m In P1,...,P m ,...,P K Ranking in if Less than the preset cost ranking threshold h C Then in P m The set of evolutionary directions ψ m Add cost to the middle;

[0033] This leads to the various cluster populations P1,...,P m ,...,P K The corresponding evolutionary directions ψ1,...,ψ m ,...,ψ K .

[0034] Furthermore, step 5 involves storing the globally non-dominated solutions obtained during the evolutionary process in a global external archive set, and then guiding the updates of each cluster population through elite individuals in the global external archive set, including:

[0035] Step 5.1: Create K clusters P1,...,P m ,...,P K Define the non-improvement flags ε1,...,ε respectively. m ,...,ε K During the entire iteration process, every Perform position clustering again and assign ε1,...,ε m ,...,ε K Reset to zero; in each cluster interval If the clustered population P m If the fitness value corresponding to the optimization direction ψm decreases, then let P m Non-improved sign ε m Let ε be zero; otherwise, let ε be zero. m =1;

[0036] If ψ m Not an empty set and εm =0, then when ψ m ={makespan} or ψ m Choose P when the cost is equal to {cost}. m The optimal particle in Makespan is taken as the optimal solution. Choose P m The worst particle in Makespan is taken as the worst solution. When ψ m ={makespan,cost} from P m One particle is randomly selected from the particles with the best makespan and the particles with the best cost as P. m optimal solution like For P m The best particle in Makespan is the worst solution. For P m The worst particle in Makespan, if For P m The particle with the optimal cost is the worst solution. For P m The particle with the worst cost;

[0037] If ψ m empty set or ε m =0, then from and A particle is randomly selected from the set as the cluster P. m optimal solution like From Then, the worst individual in the global external archive set is selected as the worst solution. like From Then, the individual with the worst cost from the global external archive is selected as the worst solution. in, and These represent the sets of excellent individuals in the global external archive set, namely makespan and cost, respectively.

[0038] One optimal solution, one worst solution, and four solutions randomly selected from the population are used as guide solutions;

[0039] Step 5.2: Update the population P using the guided solution according to the update formula of the GBO algorithm, calculate the fitness of the new particles, perform fast non-dominated sorting, and then use the non-dominated solutions in P to update the global external archive set.

[0040] Furthermore, in step 6, the method for adaptively updating the number of mutation dimensions η and the mutation rate ρ according to the population evolution speed is as follows:

[0041] If the current iteration number n satisfies For a pre-defined number of iterations, the change in HV value Δ is calculated. hv ,when The number of non-improvement iterations will be used. Increase Let η be the minimum change, and let η = 1; then... Then Clear to zero and set η to zero. in, This is the floor function. For workflow The total number of tasks; and based on the current number of non-improvement iterations. Calculate the mutation probability ρ at this time as follows:

[0042] If the current iteration number n does not satisfy Then the number of mutation dimensions η and the mutation rate ρ are not updated.

[0043] Beneficial effects:

[0044] 1. This invention addresses the computational resource waste caused by the equal distribution of particles in existing multi-population multi-objective frameworks (MPMO). It proposes a multi-objective multi-population evolutionary framework (PCMO) based on particle position clustering to solve the multi-workflow scheduling problem in cloud environments. By using K-means-based position clustering, the entire population is divided into multiple cluster populations. Each cluster population selects an optimization objective based on its own fitness level, which not only improves the diversity of the search but also avoids cross-evolution, explores more promising regions, and finds more non-dominated solutions.

[0045] 2. This invention proposes a multi-population co-evolutionary mechanism based on dynamic clustering to balance global and local development. First, the evolutionary process is divided into multiple clustering stages. In each clustering stage, the entire population is first clustered into multiple subpopulations based on fitness changes, and then the multiple clustered subpopulations are divided into dominant and subordinate categories based on fitness ranking. Second, elite individuals from the Global External Archive (GEA) are used to guide the update of subordinate populations, accelerating the evolution speed of subordinate particles. Combined with the evolution of dominant populations, the local optimal solution of the clustered populations or the global elite solution of the GEA are selectively used for updates, which not only enhances the global search capability but also improves the utilization rate of computing resources.

[0046] 3. This invention proposes an elite enhancement strategy based on adaptive mutation, which improves the ability to develop elite individuals. First, the evolutionary speed of the population is determined based on the hypervolume change rate of the Pareto front, and then the mutation rate and the number of mutation dimensions are calculated sequentially. Then, individuals to be mutated are selected from the GEA according to the mutation rate, and two methods are adopted: multi-dimensional mutation based on random distribution and single-point mutation based on Gaussian function, to enhance the diversity of elite individual search. In the whole iteration process, by adaptively adjusting the mutation scale and range, the local search capability of the algorithm is greatly increased.

[0047] 4. This invention leverages the advantages of a single workflow scheduling scheme to propose a workflow-aware elite crossover and inheritance strategy. First, it constructs a parent elite population using outstanding individuals from the current population and elite individuals from the GEA. Two parent individuals are selected from this population, and in their corresponding multi-workflow scheduling schemes, the task-resource mapping subsequences corresponding to each workflow are traversed sequentially. The better subsequence is selected and copied to the offspring, realizing a crossover operation based on elite inheritance. Second, new elite individuals are generated by integrating all outstanding subsequences to replace the dominated individuals in the GEA, exploring more promising elite solutions. Attached Figure Description

[0048] Figure 1 The flowchart illustrates a multi-workflow scheduling method based on an improved gradient optimizer, as provided by this invention.

[0049] Figure 2 The nondominated frontier graphs of the multi-workflow scheduling method based on the improved gradient optimizer provided by this invention and existing algorithms are compared on the dataset D-8-218 and under the constraint factor λ=0.25.

[0050] Figure 3 The nondominated frontier graphs of the multi-workflow scheduling method based on the improved gradient optimizer provided by this invention and existing algorithms are compared on the dataset D-8-218 and under the constraint factor λ=0.5.

[0051] Figure 4 The nondominated frontier graphs of the multi-workflow scheduling method based on the improved gradient optimizer provided by this invention and existing algorithms are compared on the dataset D-8-218 and under the constraint factor λ=0.75. Detailed Implementation

[0052] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0053] This invention provides a multi-workflow scheduling method based on an improved gradient optimizer (PCMO-GBO). The basic idea is to introduce population grouping based on K-means location clustering, optimization target allocation based on evolutionary direction awareness, regional co-evolution based on multiple populations, elite enhancement based on adaptive mutation, and elite crossover and inheritance strategies based on workflow awareness on the basis of the original gradient optimizer (GBO). This invention proposes a multi-workflow scheduling method based on an improved gradient optimizer to solve the multi-workflow scheduling problem in the cloud computing environment. Under the condition of meeting the workflow deadline constraints, it optimizes the total completion time and execution cost of the workflow as much as possible.

[0054] This invention provides a multi-workflow scheduling method based on an improved gradient optimizer, the process of which is as follows: Figure 1 As shown, the main steps include:

[0055] Step 1: Problem modeling.

[0056] Step 1.1 Modeling of multiple workflow tasks.

[0057] A set of workflows can be represented as a collection Among them, the i-th workflow It can be modeled as a directed acyclic graph (DAG), and Let each be a set consisting of a task and a directed edge in the i-th workflow. Indicates task and its son's mission The dependencies between them, i.e., tasks Only in its parent task Execution begins before the process is finished. It is pre-prepared for the i-th workflow Defined deadline constraints. In this invention, the j-th task of the i-th workflow... The parent task set and child task set are defined as follows: and use These represent the starting and ending tasks of the i-th workflow, respectively.

[0058] Most existing workflow scheduling algorithms focus on single-DAG scheduling and cannot be directly applied to multi-DAG scheduling problems. Therefore, this invention employs a synchronization-based workflow composition (SWC) method to merge multiple workflows from different requests into a single "large" workflow. A metaheuristic algorithm is then used to map all workflow tasks onto a set of available resources, enabling parallel application scheduling of multiple workflows. First, two "virtual" tasks with zero processing time and zero communication time with other tasks are created: a "virtual" entry task and a "virtual" exit task. Then, the starting tasks of all workflows are connected to the "virtual" entry task, and the ending tasks of all workflows are connected to the "virtual" exit task, thus forming a combined workflow.

[0059] Step 1.2 Cloud resource modeling.

[0060] This invention defines all resources provided by a cloud data center that are available for executing workflow applications as a resource set. Each virtual machine v f It has two properties (σ) f ,u f ), σ f u f They represent virtual machines v f The processing speed and rental unit price.

[0061] It should be noted that the original virtual machine numbering method randomly assigns numbers to virtual machines, meaning that the virtual machine number is unrelated to its performance: even if several virtual machines have very close numbers, their speeds may differ significantly. To ensure that virtual machines with similar numbers in the virtual machine encoding sequence have the same or similar processing capabilities (speeds), this invention introduces a virtual machine numbering reordering method based on processing speed. That is, virtual machines are numbered in order of their computing power, so that virtual machines with similar numbers have similar processing speeds, laying the foundation for the design of subsequent clustering algorithms.

[0062] Step 1.3 Modeling the cloud workflow scheduling problem.

[0063] This invention aims to optimize a set of workflows The two objectives are makespan and cost, and the deadline constraints for each workflow must be satisfied. To calculate... The first optimization target, makespan, first needs to calculate each workflow Makespan M i .

[0064] Task In virtual machine v f Execution time depending on instruction length and v f Processing capacity σ f The specific calculation is as shown in equation (1):

[0065]

[0066] in, For the task Instruction length, σ f For virtual machine v f Its processing capacity.

[0067] Task Transmission time It is the sum of the time it takes to transmit data to all subtasks, which can be calculated according to equations (2) and (3):

[0068]

[0069]

[0070] in, for The set of subtasks Representative task yes subtasks, v f and v m They are execution and The virtual machine's ID index, and They represent the tasks respectively. To the mission The data size and transmission time are given by , and b is the network bandwidth between different virtual machines. When the parent and child tasks are scheduled to execute on the same virtual machine, the transmission time can be considered zero.

[0071] Assuming a task Scheduled to virtual machine v f If executed above, then start time The calculation can be performed according to formula (4):

[0072]

[0073] in, Representative task The parent task, v l Is execution The virtual machine's ID. It is a task In virtual machine vl The completion time during execution, a f Indicates virtual machine v f Available time. If v f If it has not been rented, then a f For v f Booting time It is a constant independent of the task and the virtual machine; otherwise, a f For v f The completion time of the last task.

[0074] Receive mission start time Then, you can calculate according to (5). Completion time

[0075]

[0076] in, Representative task In v f Execution time on Representative task Transmission time, Is execution Required software loading time It is a constant independent of the task and virtual machine. (Boolean variable) Indicates task The required software has been loaded into v f Above, "0" indicates that the required software has not yet been successfully loaded.

[0077] Based on equations (1) to (5), each workflow can be derived. makespanMi and composite workflow The total makespan, M, is calculated as shown in equations (6) and (7):

[0078]

[0079]

[0080] For combined workflows The execution cost C is estimated using Amazon Web Services (Amazon EC2)'s pay-as-you-go model, where each task is billed based on usage (or execution) time from the user's perspective. Amazon Web Services specifies a minimum rental period of Δ for virtual machines. min When the task execution time is less than Δ min At that time, press Δ min Billing will be charged when the execution time is greater than Δ.min When the time is specified, the billing unit is seconds. Therefore, in v f Execute the task Cost The calculation is shown in equation (8):

[0081]

[0082] in, This represents the floor function, Δ is the rental time interval, and u f Indicates v f The rental cost per unit time. Therefore, based on formula (8), the combined workflow can be calculated. The total execution cost C is as shown in equation (9):

[0083]

[0084] The scheduling problem studied in this invention is to combine workflows. All tasks are assigned to the most suitable set of virtual machines to satisfy each workflow. Deadline constraints At the same time, optimize The total execution cost C and the total execution span M can be expressed mathematically as equation (10):

[0085]

[0086] Step 2, encoding and initialization.

[0087] Step 2.1: Assign sub-deadlines to each workflow task.

[0088] In this invention, each workflow scheduling scheme consists of two parts: a task resource mapping substring and a task execution order substring. Each element index of the task resource mapping substring represents the task number of the combined workflow, and each element value represents the virtual machine number executing this task. The position of each individual in the population is X = (x1, x2, ..., ..., x...). N () represents a scheduling scheme, where the minimum and maximum values ​​of each element in each dimension of the individual are 1 and 1, respectively. (Total number of available virtual machines)

[0089] After determining the task resource mapping substring, the execution order of the tasks needs to be determined to complete the DAG scheduling. This invention first assigns sub-deadlines to the tasks in each workflow based on the dependencies between tasks and the deadline constraints of the workflow, and then prioritizes the tasks with the smaller sub-deadlines to determine the task execution order substring.

[0090] When determining task priorities, this invention introduces a heuristic upward ranking method. This method calculates the upward rank value based on a pre-transmission time calculation model, i.e., the task... Transmission time for The sum of data transfer times between the task and all its parent tasks, but this invention uses a post-transfer time calculation model, i.e., task Transmission time for The sum of data transmission times between all its subtasks. Therefore, this invention modifies the formula for calculating the upward rank value based on the post-transmission time calculation model, and the modified task... The upward rank value is calculated as shown in equation (11):

[0091]

[0092] Where pm represents the resource pool The fastest virtual machine speed, Indicates task The upward rank value, Indicates task Instruction length, Indicates from the task To its sub-task The size of the transmitted data is denoted by b, where b is the network bandwidth between different virtual machines.

[0093] Similarly, considering the post-transmission time calculation model of this invention, the method for calculating task sub-deadlines in the original upward ranking method cannot be directly used, because the sub-deadlines obtained using the original method will be smaller than the actual ones. This leads to incorrect sub-deadlines causing incorrect task priority execution order, resulting in tasks that should be executed first having to wait a long time before execution. Therefore, this invention corrects the task... Sub-deadline The calculation method is shown in equation (12):

[0094]

[0095] in, Indicates entry task The upward rank value, Represents workflow Deadline constraints.

[0096] Step 2.2: Evaluation of the satisfaction of multi-workflow constraints based on workflow synthesis.

[0097] This invention uses the SWC workflow merging method to combine multiple workflows. Integrate into a "large" workflow To construct a multi-workflow model, particle encoding is performed according to the single-workflow scheduling concept. During algorithm iteration and non-dominated particle sorting, the merits of two individuals, X1 and X2, need to be evaluated. Depending on whether an individual satisfies the constraints, an individual may correspond to a feasible scheduling scheme or an infeasible one. Therefore, the comparison of individuals X1 and X2 can be divided into the following three cases:

[0098] (1) When both X1 and X2 are feasible solutions, the order of X1 and X2 is used to determine the quality, and the scheduling scheme corresponding to the individual ranked first is better.

[0099] (2) When there is one feasible solution and one infeasible solution among X1 and X2, the scheduling scheme corresponding to the feasible solution is better;

[0100] (3) When both X1 and X2 are infeasible solutions, the constraint exceedance evaluation index Θ designed in this invention is used for evaluation, as follows:

[0101] For scheduling problems with deadline constraints, when comparing the merits of two infeasible solutions X1 and X2, the existing method is to determine which solution has a smaller makespan for its corresponding "large" workflow. However, since each workflow has its own deadline constraint D... i Only when an individual is on all All satisfy M i ≤D i Only then is the solution feasible for that individual. Clearly, a single individual corresponds to a "large" workflow. A small makespan does not indicate the specific constraint satisfaction of an individual for each workflow, nor how many sub-workflow deadline constraints it simultaneously satisfies. This method of comparing infeasible solutions cannot accurately determine the quality of two infeasible solutions, thus lacking a clear direction for population evolution, ultimately reducing search efficiency and affecting the quality of finding near-optimal solutions.

[0102] To this end, the present invention proposes a constraint exceedance evaluation index Θ, which evaluates the constraint exceedance of the solution from the perspective of satisfying each workflow constraint. That is, it is the sum of the percentages of the total completion time of all workflows in the entire "large" workflow that exceed their corresponding deadline constraints. The specific calculation is shown in Equation (13):

[0103]

[0104] Among them, M i Di and Di represent the total execution span and deadline constraints of the i-th workflow, respectively. Representative workflow set Total number of workflows included.

[0105] For two infeasible solutions X1 and X2, X1 is better than X2 if and only if:

[0106]

[0107] Where C(X1) and C(X2) represent the execution costs of the “large” workflows corresponding to scheduling schemes X1 and X2, respectively, and Θ(X1) and Θ(X2) represent the degree of constraint exceedance of scheduling schemes X1 and X2, respectively.

[0108] This method allows for an accurate evaluation of the overall satisfaction of multiple workflow deadline constraints by the scheduling schemes corresponding to two individuals. This enables the selection of individuals with a high degree of overall constraint satisfaction, guiding the population towards truly meeting the deadlines and accelerating algorithm convergence.

[0109] Step 2.3: Population and external archive set initialization.

[0110] Step 2.3.1, Population initialization.

[0111] To better guide the population, nine predefined scheduling schemes are selected as initial particles. First, all tasks are randomly assigned to virtual machines of the same type, generating a scheduling scheme. This results in six predefined scheduling schemes i1-i6 for the six types of virtual machines. Second, the HEFT algorithm is used to generate a predefined scheduling scheme i7. Two tasks from i7 scheduled to different types of virtual machines are randomly selected, and their assigned virtual machines are swapped, generating a new predefined scheduling scheme i8. Then, all tasks from i7 assigned to the virtual machine with the highest processing capacity are reassigned to the virtual machine with the highest processing capacity per unit cost, generating a new predefined scheduling scheme i9. Finally, these nine predefined initial particles are inserted into the initial population, and the remaining individuals are generated through random initialization, completing the initialization of the entire population. Simultaneously, the iteration count n is initialized to 0 (n = 0), and the maximum number of iterations is defined as N.

[0112] Step 2.3.2: Initialize the external archive set.

[0113] The external archive set GEA is used to store the set of all non-dominated solutions obtained in the history of algorithm iterations. This invention initializes the GEA by placing the predefined scheduling schemes i1-i9 that satisfy the deadline constraints into the GEA.

[0114] Step 3: Population grouping based on K-means location clustering.

[0115] After population initialization, the population is grouped using the K-means positional clustering algorithm. After grouping, particles within the same cluster have similar task-virtual machine mappings, and due to the virtual machine number re-clustering operation, they often have similar fitness levels, which is beneficial for assigning appropriate evolutionary targets to the clusters later.

[0116] Step 3.1: Determine whether to perform re-clustering.

[0117] Determine if the current iteration count n exists. If it exists, perform re-clustering and proceed to step 3.2; otherwise, proceed to step 5.

[0118] Step 3.2: Divide the population into P1,...,P m ,...,P K Clustering of subpopulations.

[0119] The introduction of the K-means algorithm involves calculating the distance between particles. This invention uses Euclidean distance to model the distance between particles and defines the q-th individual in the population. With the wth individual The Euclidean distance d between them q,w The calculation is as shown in equation (15):

[0120]

[0121] in, Representing the q-th individual Values ​​of each dimension Representing the w-th individual Values ​​of each dimension Representative workflow The total number of tasks.

[0122] During re-clustering, K particles are randomly selected from the population P as the initial cluster centers. The specific value of K is calculated as shown in equation (16):

[0123]

[0124] rand() is a function that randomly selects a value between 0 and 1. It is a floor function.

[0125] The clustering process requires multiple iterations. After obtaining K cluster centers, the Euclidean distance from a particle to the K cluster centers is calculated according to equation (15). By comparing the K distances, the particle is assigned to the cluster with the closest distance. After all particles in the population have completed clustering, the centroids of the current K clusters are calculated, and then the K centroids are taken as new cluster centers. This clustering iteration process is repeated until the number of clustering iterations reaches the maximum number of K-means iterations N.Kmeans Or the change in the position of the population centroid is less than the threshold δ min The K-means clustering iterative process terminates, and the current population grouping results P1,...,P are output. m ,...,P K .

[0126] Step 4: Population optimization target allocation based on evolutionary direction awareness.

[0127] After reordering the virtual machine numbers and grouping them based on Euclidean distance, particles in the same cluster often have similar positions and fitness, which is beneficial for selecting suitable optimization objectives for the cluster population and accelerating its evolution. This invention proposes an evolutionary direction-aware optimization objective allocation strategy, which assigns optimization objectives to multiple clusters P1,...,P... after clustering. m ,...,P K Choose a reasonable evolutionary direction, based on the selected evolutionary direction as P1,...,P m ,...,P K Choose a more reasonable update method to improve algorithm performance.

[0128] This invention specifies four options for the direction of cluster evolution: empty set, makespan alone, cost alone, and both makespan and cost. This is for cluster P. m When selecting the evolutionary direction, first calculate the fitness (makespan) and cost of the population particles according to equations (7) and (9), and then apply these values ​​to each cluster population P1,...,P m ,...,P K Evaluations were performed on makespan and cost respectively to obtain the populations P1,...,P of each cluster. m ,...,P K The evolutionary direction ψ1,...,ψ m ,...,ψ K .

[0129] The specific evaluation method is as follows: Calculate the rank of each particle in a cluster relative to the entire population, thus obtaining the average rank of all particles in that cluster. Based on the fact that there are K average rank values ​​for K cluster centers, the ranking of the cluster population P1,...,P... is calculated. m ,...,P K Sort the data to obtain each cluster population P. m In P1,...,P m ,...,P K Ranking in If P m Ranking Less than the preset Makespan ranking threshold h M Then in P m The set of evolutionary directions ψ m Add makespan to the expression. Similarly, we can obtain the population P for each cluster. m In P1,...,P m ,...,P K Ranking in if Less than the preset cost ranking threshold h C Then in P m The set of evolutionary directions ψ m Add cost to the equation. Based on this, obtain the various cluster populations P1,...,P m ,...,P K The corresponding evolutionary directions ψ1,...,ψ m ,...,ψ K .

[0130] Step 5: Regional co-evolution based on multiple populations.

[0131] After obtaining the cluster populations P1,...,P m ,...,P K The evolutionary direction ψ1,...,ψ m ,...,ψ K Building upon existing methods, this invention introduces a Global External Archive (GEA) to store the globally non-dominated (elite) solutions obtained throughout the evolutionary process. Elite individuals within the GEA guide updates to each cluster population, enabling multi-region parallel search and co-evolution of the solution space. Specifically, if a region (cluster population) lacks potential, the global elite solutions within the GEA guide individuals to escape that region, improving computational resource utilization. If a region (cluster population) has potential, excellent individuals within the cluster population are selected as guiding solutions to promote the development of potential regions, accelerate algorithm convergence, and ultimately enhance the algorithm's global search capability. The specific steps are as follows:

[0132] Step 5.1: Cluster populations P1,...,P in sequence. m ,...,P K Choose a guiding solution.

[0133] The original GBO update model is based on six guiding solutions: one optimal solution and one worst solution (i.e., executing step 5.1.1 or step 5.1.2), and four solutions randomly selected from the population (i.e., executing step 5.1.3). To address the problems of the original GBO's singular evolutionary direction and susceptibility to local optima, this invention is based on the evolutionary direction ψ1,...,ψ of the clustered population. m ,...,ψK This allows for the selection of the most suitable guiding solution for different cluster populations.

[0134] To determine the search potential of a region (cluster population), this invention uses K cluster populations P1,...,P m ,...,P K Non-improvement flags ε1,...,ε are defined. m ,...,ε K During the entire iteration process, every Perform position clustering again and assign ε1,...,ε m ,...,ε K Reset to zero. In each iteration interval... If the clustered population P m Optimization direction ψ m If the corresponding fitness value decreases, then P will be... m Non-improved sign ε m Set to zero, meaning that the population P is considered to be zero. m It has potential; otherwise, it is considered P. m No potential, and will ε m Let it be 1, i.e., ε m =1.

[0135] Even without considering any other factors, the population particles will follow the probability ρ r Perform a random movement (i.e., execute step 5.1.1), ρ r It can be calculated according to formula (17):

[0136] ρ r =100 / (n+200) (17)

[0137] Where n is the current iteration number.

[0138] If ε m =1, which means that population P m Optimization direction ψ m The corresponding fitness value remains unchanged, and this cluster population P m If evolution is not obvious or has not occurred, i.e. there is no potential, random movement is required (i.e., proceed to step 5.1.1).

[0139] If cluster P m Evolutionary direction ψ m Since it is an empty set, P has no direction of evolution. m The particles in the process also need to move randomly (i.e., perform step 5.1.1).

[0140] If ψ m Not an empty set and ε m =0, indicating that P mThis is a potential area. In order to explore the potential within the target ψ m The more promising areas will be selected according to ψ m For P m Choose its locally optimal individual as its optimal solution. (That is, perform step 5.1.2).

[0141] Step 5.1.1: Select a guided solution using random shift.

[0142] This invention adopts and Let these represent the sets of excellent individuals in the GEA in terms of makespan and cost, respectively. To ensure the diversity of the Pareto front, this invention... It also includes both the intermediate and optimal individuals in GEA sorted by makespan. Similarly, The construction of a set is similar to

[0143] The specific process of random movement: from and A particle is randomly selected from the set as the cluster P. m optimal solution like From Then select the worst individual from GEA as the worst solution. like From Then select the individual with the worst cost from GEA as the worst solution.

[0144] Step 5.1.2: Select a guiding solution from this cluster.

[0145] From the particles in this cluster, P is... m Select the guided ungrouping P m Evolutionary direction ψ m The effect, if ψ m ={makespan}, select this cluster P m The best particle in Makespan is used as the optimal solution. Select this cluster P m The worst particle in Makespan is taken as the worst solution. If ψ m ={cost}, The selection method is similar. If ψ m ={makespan,cost}, from this cluster P m One particle is randomly selected from the individuals with the best makespan and the individuals with the best cost as P. m optimal solution like For P m The best particle in the Makespan algorithm is the worst solution. For P m The worst particle in Makespan, if For P m The particle with the optimal cost in the middle is the worst solution. For P m The particle with the worst cost.

[0146] Step 5.1.3 Randomly select solutions within the group

[0147] The selection of four random solutions from the cluster population is unaffected by the optimal and worst guiding solutions. If cluster P m If it contains more than 4 particles, then in P m Four guiding solutions are randomly selected from the population P; otherwise, four guiding solutions are randomly selected from the population P.

[0148] Step 5.2: Update the cluster population using the original GBO algorithm.

[0149] Based on the above, all cluster populations P1,...,P m ,...,P K The selected guiding solution updates the population P according to the original GBO algorithm update formula, calculates the fitness of the new particles, performs fast non-dominated sorting, and then uses the non-dominated solutions in P to update GEA (adding the non-dominated solutions in P to GEA, while removing infeasible and dominated solutions).

[0150] Step 6: Elite enhancement strategy based on adaptive mutation.

[0151] The Genetic Algorithm (GEA) contains many potential non-dominated solutions. To better develop elite solutions within the GEA and enhance the local search capability for elite solutions to find more non-dominated solutions, this invention proposes an elite enhancement strategy based on adaptive dynamic dimensional mutation. Specifically, small-dimensional mutation is considered a local search mechanism applied to slow evolutionary stages. When the evolutionary speed is fast, the mutation rate is appropriately reduced, the number of mutation dimensions is increased, and the search range of the mutation value distribution function is narrowed. When the evolutionary speed is slow, the mutation rate is appropriately increased, the number of mutation dimensions is decreased, and the search range of the mutation value distribution function is increased. In other words, when the population evolution speed is slow, i.e., the population is close to the optimal Pareto front, the local search is strengthened to explore the potential of the GEA; when the population evolution speed is fast, i.e., the update effect of the original GBO algorithm is good, the local search of elite individuals is reduced. Compared with existing elite enhancement strategies that only use a fixed processing mode, the adaptive mutation method of this invention is more flexible and efficient, improving the quality of finding non-dominated solutions while avoiding a significant increase in computational cost. The specific steps are as follows:

[0152] Step 6.1: Determine whether to update the number of mutation dimensions and the mutation rate.

[0153] HV is a performance metric based on the closed area to measure the quality of an algorithm. This invention uses HV to determine the timing and scale of mutation. If the current iteration number n satisfies... Then calculate the change in HV value Δ hv (current HV value minus) The value of HV before the next iteration is divided by the current HV value, and this is used to determine the evolutionary rate of the current population. Then, based on Δ... hv Update the number of mutation dimensions η (i.e., execute step 6.2) and the mutation rate ρ (i.e., execute step 6.3); otherwise, directly perform the mutation operation according to the current number of mutation dimensions η and mutation rate ρ (i.e., execute step 6.4).

[0154] Step 6.2: Update the number of mutation dimensions.

[0155] When the population evolves slowly, that is At that time, the number of non-improvement iterations will be... Increase And it uses single-point mutation, i.e., η=1; when the population evolves rapidly, i.e. At that time, Zero out, and use multi-point mutation, with the number of mutation dimensions η calculated as shown in equation (18):

[0156]

[0157] in, It is the floor function. Representative workflow The total number of tasks, where n is the current iteration number.

[0158] Step 6.3: Update the mutation rate.

[0159] Based on the current number of non-improved iterations Calculate the probability of mutation ρ at this time according to equation (19):

[0160]

[0161] Step 6.4: Perform the mutation operation.

[0162] First, the current GEA is sorted using fast non-dominated sorting and crowding distance sorting, selecting the top-ranked elite individuals. Then, a subset of these top-ranked individuals is randomly selected based on the mutation probability ρ, and mutation is performed using the number of mutation dimensions η. If η = 1, a Gaussian distribution function is used for single-point mutation; if η > 1, a random distribution function is used for multi-dimensional mutation. After mutation, the newly generated individuals are subjected to fast non-dominated sorting, and the non-dominated solutions are used to update the external archive set GEA.

[0163] Step 7: Elite crossover and inheritance strategy for workflow awareness.

[0164] In multi-workflow scheduling schemes based on workflow synthesis, when some sub-workflows exhibit excellent fitness while the scheduling performance of other sub-workflows is mediocre, the scheduling scheme corresponding to the excellent sub-workflow still has good reference value. However, this excellent sub-scheduling scheme may be eliminated due to the poor scheduling performance of the entire "large" workflow. To address this, this invention proposes a workflow-aware elite crossover and inheritance strategy. First, in each iteration, the better individuals are selected to form a parent set. Second, through crossover operations, the excellent scheduling schemes (task-resource mapping sub-sequences) corresponding to each sub-workflow are copied to the offspring, inheriting the excellent scheduling knowledge from each sub-workflow to generate new offspring individuals. Finally, these generated excellent offspring individuals are used to update the GEA (Generation-Oriented Allocation) to optimize the execution of each workflow.

[0165] It should be noted that: when it is a parent set When selecting outstanding individuals, although individuals with better fitness are more likely to have excellent sub-workflow scheduling schemes, the set of crossover offspring individuals from the previous iteration is also important. There may be individuals with poor fitness but good local fitness in a certain sub-workflow.

[0166] Step 7.1: Construct a set of parent individuals with elite crossover.

[0167] At each iteration, the parent set used for elite crossover It consists of three parts: 1) the set of offspring obtained from the crossover in the previous iteration. 2) Sort the updated and adaptively mutated GEAs by crowding distance and take the top-ranked ones. 3) After performing fast dominance sorting and crowding distance sorting on the population P, select the top-ranked individuals. One particle.

[0168] Step 7.2, Elite crossover and inheritance of workflow awareness.

[0169] When performing crossover operations, the binary tournament method is first used to... Select two particles Act as the parent generation, then iterate through the data in sequence. Workflow within a given workflow. For a specific workflow... from Choose the better scheduling scheme as the offspring particle The middle corresponds to scheduling scheme Repeat this operation until all elements have been traversed. All workflows within this process constitute the offspring particles. in, Two particles Corresponding to workflow (with constraints) The scheduling scheme with the target being makespan.

[0170] from and Methods for selecting the best individual: When neither scheduling scheme satisfies every workflow Deadline constraints When both scheduling schemes satisfy the condition, the one with the smaller makespan value is preferred. Deadline constraints If both scheduling schemes satisfy the constraints If the solutions do not dominate each other, then an individual is randomly selected; if one solution dominates another, then the undominated solution is selected; if the scheduling scheme corresponding to one of the individuals satisfies the constraints... If the other solution does not satisfy the constraints, then the solution that satisfies the constraints is selected.

[0171] Perform the above cross operation Next, generated in total Each child individual constitutes the child set of this iteration. Then Perform a fast non-dominated sort on the individuals in the dataset, and use the non-dominated solutions to update the external archive set GEA.

[0172] Step 8: Determine whether the number of iterations has reached the preset upper limit.

[0173] Increment the iteration count n by 1. If n ≥ the maximum iteration count N, output the current entire external archive set GEA as the result; otherwise, go to step 3.

[0174] Example:

[0175] To verify the performance of the workflow scheduling method proposed in this invention, experiments were conducted using the cloud workflow scheduling simulation tool WorkflowSim. PCMO-GBO was compared with four popular algorithms: NSGA-II, MOACS, Multi-Objective Gradient Optimization Algorithm (MO-GBO), and Multi-Swarm Hybrid Intelligent Optimization Algorithm (MCHO). Among them, MO-GBO is a dual-objective optimization algorithm implemented on the basis of single-objective GBO.

[0176] This invention selects workflows of different scales and types for experiments. Based on four types of workflows—Montage, Inspiral, CyberShake, and Epigenomics—a dataset D-8-218 (containing 8 workflows and 218 tasks) was constructed, namely:

[0177] Data set D-8-218 = 2*Montage_25 + 2*Inspiral_30 + 2*CyberShake_30 + 2*Epigenomics_24, along with a resource set containing 20 virtual machines, was used for scheduling simulation experiments under three different constraint factors (λ∈{0.25,0.5,0.75}). To avoid randomness in the scheduling results, 20 repeated experiments were conducted, and two popular evaluation metrics, namely the non-dominated front distribution and hypervolume (HV), were used to evaluate the experimental results. The non-dominated front distributions of the experimental results on dataset D-8-218 are as follows: Figure 2 , 3 As shown in Figure 4.

[0178] Figure 2 , 3Figures 4 and 4 show the non-dominated front plots of dataset D-8-218 under three constraint factors. Under the three constraint factors, the Pareto front obtained by PCMO-GBO has a lower objective value and a more uniform distribution, indicating that PCMO-GBO is significantly better than other algorithms in terms of diversity and optimization performance. Under the strict constraint λ = 0.25, PCMO-GBO's advantage over other methods is most obvious. This is because PCMO-GBO can find solutions satisfying all constraints faster in handling complex multi-constraint problems, quickly achieving the transition from the infeasible region to the feasible region. In all cases, MOACS and MCHO can also find some uniformly distributed non-dominated solutions, with MOACS performing better in terms of solution performance, while MCHO has more non-dominated solutions and a wider distribution. MO-GBO and NSGA-II perform the worst. Although NSGA-II has relatively better optimization performance, it obtains fewer non-dominated solutions with a more concentrated distribution. MO-GBO has the worst optimization performance, but it obtains more non-dominated solutions than NSGA-II.

[0179] To facilitate a direct comparison of the HV values ​​of the PCMO-GBO algorithm with other algorithms, this invention uses the Related Improvement Percentage (RIP) to represent the HV results. The formula for calculating RIP is as follows:

[0180]

[0181] in, Table 1 shows the RIP comparison of the experimental results on dataset D-8-218:

[0182] Table 1. RIP values ​​of PCMO-GBO compared to other algorithms

[0183]

[0184] Table 1 shows the HV comparison between PCMO-GBO and the algorithms MOACS, MCHO, NSGA-II, and MO-GBO. The results are represented by RIP in Equation (20), which is the percentage improvement in RIP of PCMO-GBO compared to the other four comparison algorithms. A positive RIP indicates that PCMO-GBO performs better. As can be seen from Table 1, under the dataset D-8-218 and all constraints, the performance of PCMO-GBO is far superior to MOACS, MCHO, NSGA-II, and MO-GBO.

[0185] As can be seen, in all cases, PCMO-GBO outperforms MCHO, MO-GBO, NSGA-II, and MOACS, thus demonstrating a clear advantage for PCMO-GBO.

[0186] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-workflow scheduling method based on an improved gradient optimizer, characterized in that, Includes the following steps: Step 1: Model the input multi-workflow tasks with deadline constraints and the available virtual machine resources for lease; Multi-workflow tasks are represented as , For the i-th workflow, Workflow set Total number of workflows included; virtual machine resources are represented as , For the f-th virtual machine; Step 2: Assign sub-deadlines to all subtasks in the workflow, and determine the scheduling sequence by prioritizing the task with the earliest deadline; then, based on the constraint satisfaction evaluation index... Evaluate the merits of infeasible solutions. This represents the sum of the percentages of workflows exceeding constraints; Location of each individual in the population This represents a scheduling scheme, where the index of each dimension element represents a workflow task number, and the value of each dimension element represents the virtual machine allocated to that task; nine predefined initial particles are inserted into the population, and the remaining individuals in the population are initialized using a random generation method; individuals that satisfy the constraints among the nine predefined initial particles are inserted into the global external archive set to complete the initialization of the global external archive set. Step 3: If the current iteration number n satisfies ,in If the threshold for the number of iterations is used, then the K-means location clustering algorithm is used to divide it into groups. Cluster subpopulation ; Otherwise, proceed to step 5; Step 4: Based on the population The total fitness (makespan) and cost of all particles are used to determine the most suitable evolutionary direction. ; Step 5: Store the global non-dominated solutions of the population obtained during the evolution process in the global external archive set, and then guide the update of each cluster population through the elite individuals in the global external archive set. Step 6: After a set number of iterations, adaptively update the number of mutation dimensions based on the population evolution speed. With the rate of variation And in each iteration, according to the current number of mutation dimensions With the rate of variation Perform a mutation operation on an individual in the global external archive set, and update the global external archive set with the mutation result; Step 7: Use the selected set of excellent parent generations to perform cross operations, and copy the mapping sequence of the better task to the resource corresponding to each sub-workflow to the child generation in turn. Generate child individuals through cross operations based on elite inheritance, and update the global external archive set with the generated new child individuals. Step 8: Increment the iteration count n by 1. If n is not less than the maximum iteration count, output the current global external archive set as the result and end this process. Otherwise, proceed to step 3; The constraint satisfaction evaluation index The calculation method is as follows: , in, and The first Total execution span and deadline constraints for each workflow; In step 2, the constraint satisfaction evaluation index is used. Evaluating the merits of infeasible solutions includes: for two infeasible solutions and If and only if the following conditions are met: At that time, Superior ,in, , They represent the scheduling schemes respectively. and The corresponding workflow execution cost, , They represent the scheduling schemes respectively. and The constraints exceeded the limit.

2. The multi-workflow scheduling method according to claim 1, characterized in that, The virtual machines are numbered in order of their computing power.

3. The multi-workflow scheduling method according to claim 1, characterized in that, In step 2, sub-deadlines are assigned to all tasks in the workflow. The scheduling sequence is determined by prioritizing the task with the earliest deadline. Task Sub-deadline for: , in, For entry task The upward rank value, For workflow Deadline constraints; Task Upward rank value for: , Where pm is the resource pool The fastest processing speed of a virtual machine in China. For the task Instruction length, To the task To its sub-task The size of the transmitted data is denoted by b, where b is the network bandwidth between different virtual machines.

4. The multi-workflow scheduling method according to claim 1, characterized in that, The nine predefined initial particles are defined as follows: by randomly assigning all tasks to virtual machines of the same type to generate a scheduling scheme, six predefined scheduling schemes can be obtained from six types of virtual machines. to Then, the HEFT algorithm is used to generate a predefined scheduling scheme. First, select randomly. Two tasks scheduled to different types of virtual machines are swapped, resulting in a new predefined scheduling scheme. Then All tasks originally allocated to the virtual machine with the highest processing capacity will be reassigned to the virtual machine with the highest processing capacity per unit cost, resulting in a new predefined scheduling scheme. ,but to The predefined scheduling scheme consists of nine predefined initial particles.

5. The multi-workflow scheduling method according to claim 1, characterized in that, In step 3, the K-means location clustering algorithm is used to divide it into... Cluster subpopulation The following method is used to calculate the distance between individuals in the population: The first in the population individual With the wth individual Euclidean distance between for ,in, For the first The first individual Values ​​of each dimension For the first The first individual Values ​​of each dimension For workflow The total number of tasks.

6. The multi-workflow scheduling method according to claim 1, characterized in that, In step 4, based on the population The total fitness (makespan) and cost of all particles are used to determine the most suitable evolutionary direction. The method is as follows: Calculate the rank of each particle in the cluster relative to the makespan in the entire population, and obtain the average rank of all particles in that cluster; based on the K averages of the K clusters, analyze the cluster populations. Sort the data to obtain each cluster population. exist Ranking in ; if Ranking Less than the preset Makespan ranking threshold Then in The set of evolutionary directions Add makespan to the middle; Calculate the rank of each particle in the cluster relative to the cost of the entire population, and obtain the average rank of all particles in that cluster; based on the K averages of the K clusters, analyze the cluster populations. Sort the data to obtain each cluster population. exist Ranking in ,if Less than the preset cost ranking threshold Then in The set of evolutionary directions Add cost to the middle; This yields the various cluster populations. Corresponding evolutionary direction .

7. The multi-workflow scheduling method according to claim 1, characterized in that, Step 5 involves storing the globally non-dominated solutions obtained during the evolution process in a global external archive set, and then guiding the updates of each cluster population through elite individuals in the global external archive set, including: Step 5.1: Create K clusters. Define non-improved symbols separately During the entire iteration process, every Perform a new location clustering and... Reset to zero; in each cluster interval If clustered population Optimization direction If the corresponding fitness value decreases, then let Non-improved mark Set to zero; otherwise, set to zero. =1; like Not an empty set and Then when or Time selection The optimal particle in Makespan is taken as the optimal solution. ,choose The worst particle in Makespan is taken as the worst solution. ;when From One particle is randomly selected from the particles with the best makespan and the particles with the best cost as... optimal solution ,like for The optimal particle in Makespan is the worst solution. for The worst particle in Makespan, if for The particle with the optimal cost is the worst solution. for The particle with the worst cost; like For empty set or Then from and Randomly select a particle from the set as the cluster. optimal solution ;like From Then, the worst individual in the global external archive set is selected as the worst solution. ;like From Then, the individual with the worst cost from the global external archive is selected as the worst solution. ;in, and These represent the sets of excellent individuals in the global external archive set, namely makespan and cost, respectively. One optimal solution, one worst solution, and four solutions randomly selected from the population are used as guide solutions; Step 5.2: Update the population P using the guided solution according to the update formula of the GBO algorithm, calculate the fitness of the new particles, perform fast non-dominated sorting, and then use the non-dominated solutions in P to update the global external archive set.

8. The multi-workflow scheduling method according to claim 1, characterized in that, In step 6, the number of mutation dimensions is adaptively updated according to the population evolution speed. With the rate of variation The method is as follows: If the current iteration number n satisfies , For a pre-defined number of iterations, the change in the HV value is calculated. ,when The number of non-improvement iterations will be used. Increase , Let be the minimum change, and let Immediately Then Clear to zero, and order for ,in, This is the floor function. For workflow The total number of tasks; and based on the current number of non-improvement iterations. Calculate the mutation probability at this time. for ; If the current iteration number n does not satisfy Then the number of mutation dimensions will not be updated. With the rate of variation .

Citation Information

Patent Citations

  • Heuristic type coarse grain parallel grid task scheduling method

    CN102662743A

  • Improved fuzzy neural network bus intelligent scheduling method based on chaos theory

    CN106295886A