A cloud platform task scheduling method based on dimension learning strategy and grey wolf optimization

Abstract

The application discloses a cloud platform task scheduling method based on a dimension learning strategy and grey wolf optimization, which comprises the following steps: 1, initializing parameters of a DLH-GWO algorithm; 2, selecting an execution cost, a maximum completion time, a waiting time and a resource utilization to construct a multi-objective optimization model, and calculating the fitness of a grey wolf individual to obtain the positions of the top three wolves; 3, combining a search strategy of the dimension learning DLH and the positions of the top three wolves, and calculating two candidate update positions of the remaining wolves in the grey wolf population N; 4, comparing the fitness values of the two candidate update positions and the current position of the grey wolf individual, and selecting a better position to update the position of the grey wolf individual; 5, performing the same strategy position adjustment, selection and update operation on all the grey wolf individuals; and 6, judging whether the maximum iteration number is reached in the tth iteration, if not, continuing the next search, otherwise outputting an optimal solution, balancing the global and local search capabilities of the algorithm and improving the availability of the cloud platform.

Description

Technical Field

[0001] The present invention belongs to the field of cloud computing, and specifically is a cloud platform task scheduling method based on dimensional learning strategy and gray wolf optimization. Background Art

[0002] Cloud computing is a new network application technology that provides various solutions to user needs through the "cloud." Its core concept is to use the Internet as the core, cloud services as the link, and users as the ultimate goal. In recent years, with the expansion of computing infrastructure, resource management in large, heterogeneous, and distributed environments has become a challenging task. In cloud environments, due to the decentralized and uncertain nature of resources, people often encounter resource allocation problems caused by factors such as resource heterogeneity and failures. To ensure efficient workload and application performance, research on cloud task resource scheduling is of great significance.

[0003] A task is the basic computing unit that runs on a resource node. Multiple tasks together constitute a job, so job scheduling requires different types of resources and processing capabilities. Scheduling also helps users manage and use a large number of resources. As an NP-class problem, scheduling refers to a method to achieve optimal resource allocation among all given tasks within a limited time to obtain the required quality of service. Task scheduling assigns a set of workflows to multiple appropriate processors in a way that minimizes execution and communication costs, that is, by mapping tasks to appropriate resources to provide an optimized solution. The efficient operation of the scheduling process benefits from an appropriate scheduling strategy, which assigns certain tasks to limited resources to optimize one or more objectives. Scheduling strategies are generally used to find the optimal solution for resource-constrained target tasks, that is, to clarify which tasks are executed on which resources.

[0004] The design and implementation of scheduling algorithms must consider multiple factors, including cost, latency, priority, and energy efficiency. Researchers have proposed a variety of scheduling strategies to improve the performance of cloud computing. Common scheduling methods include exhaustive and deterministic algorithms. Deterministic algorithms (such as the shortest job first algorithm, multi-level feedback queue algorithm, and round-robin algorithm) offer superior solutions to scheduling problems compared to exhaustive methods. However, deterministic algorithms are only suitable for situations with small data volumes and are less effective when dealing with large-scale distributed data. Another scheduling strategy, metaheuristic algorithms, can effectively address this problem. Metaheuristic algorithms typically use iterative strategies to optimize tasks within a reasonable timeframe and are highly efficient for solving relatively complex problems. Extensive research has been conducted on metaheuristic algorithms, such as the gray wolf optimization algorithm, the ant colony optimization algorithm, the whale optimization algorithm, and the sea lion optimization algorithm. These algorithms have all improved the efficiency and accuracy of cloud task scheduling to a certain extent. However, when faced with more complex scheduling optimization problems, they are prone to problems such as a lack of population diversity and stagnation in the search process due to local optima. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of the present invention is to provide a cloud platform task scheduling method based on dimensional learning strategy and gray wolf optimization, which can well balance the global and local search capabilities of the algorithm and improve the availability of the cloud platform.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A cloud platform task scheduling method based on dimensional learning strategy and gray wolf optimization includes the following steps:

[0008] Step 1: Initialize the parameters of the DLH-GWO algorithm so that the position information of the gray wolf population N is stored in a matrix Position with N rows and D columns in a random distribution state;

[0009] Step 2: Select execution cost, maximum completion time, wait time, and resource utilization to build a multi-objective optimization model. The custom multi-objective fitness function is:

[0010] F=min{Execution Cost, Makespan, WaitTime}+max{Utilization} (1)

[0011] According to formula (1), the fitness of the gray wolf individuals is calculated, and the position information X of the top three wolves with the best performance is obtained. α 、X βand X δ ;

[0012] Step 3: Combine the search strategy based on dimensional learning DLH and the position information X of the first three wolves α 、X β and X δ , respectively calculate the two candidate update positions of the remaining wolves in the gray wolf population N. In the DLH search strategy, the gray wolf X i Each dimension of the new position of (t) is expressed as:

[0013] X i-DLH (t+1)=X i，d (t)+rand×(X n，d (t)-X r，d (t)) (2)

[0014] Where, X n，d (t) represents X i (t) The d-th dimension of the adjacent gray wolf, X r，d (t) represents the dth dimension of the gray wolf randomly selected from the matrix Position, X i-DLH (t+1) indicates that the DLH search strategy is gray wolf X i (t) The dth dimension of a new candidate position generated;

[0015] X n，d (t) represents the i The neighborhood N of (t) i (t) is randomly selected, the neighborhood N i (t) is expressed as:

[0016] N i (t) = {X j (t)|D i (X i (t), X j (t))≤R i (t), X j (t)∈Q} (3)

[0017] Where D i For X i (t) and X j (t), the Euclidean distance between i (t) is the neighborhood radius, and its expression is:

[0018] R i (t)=||X i (t)-X i-GWO (t+1)|| (4)

[0019] Where, X i-GWOA location search is defined as:

[0020]

[0021]

[0022]

[0023] Where D α 、D β and D δ Represent the distances between ω wolf and α wolf, β wolf and δ wolf respectively, and is a coefficient vector used to control the degree of divergence when the gray wolf searches for prey. Represents the location information of w wolf, and Represents the location information of α wolf, β wolf and δ wolf respectively, and It represents the position update of the gray wolf individual caused by the influence of α wolf, β wolf and δ wolf respectively;

[0024] Step 4: Compare the two candidate update positions X i-GWO (t+1) and X i-DLH (t+1) and the fitness value of the current position of the gray wolf individual, select a better position to update the position of the gray wolf individual, expressed as:

[0025]

[0026]

[0027] Step 5: Repeat steps 2 to 4 to perform the same strategy of position adjustment, selection, and update operations on all gray wolf individuals;

[0028] Step 6: Determine whether the tth iteration has reached the maximum number of iterations Maxiter. If t < Maxiter, return to step 2 to continue the next search process; otherwise, output the optimal solution.

[0029] Furthermore, the parameters of step 1 include the location information of the gray wolf individuals in the gray wolf population N, the maximum number of iterations Maxiter, the coefficient vector and and the convergence factor a, where:

[0030] The gray wolf population N is randomly distributed in the given search space matrix Position, and the position distribution of gray wolf individuals follows:

[0031] X ij =l j+rand j [0, 1]×(u j -l j ), i∈[1, N], j∈[1, D] (10)

[0032] The position of the i-th wolf in the t-th iteration is expressed as a substantial vector:

[0033] X i (t) = {x i1 , x i2 ,...,x iD} (11)

[0034] In formulas (10) and (11), u j and l j Indicates the upper and lower limits of the search range of each dimension of the gray wolf individual; rand j [0, 1] represents a random number between 0 and 1; D represents the number of tasks, that is, the dimension of the problem. After each iteration, the information of the entire matrix Position will be updated and indicate a better solution state;

[0035] The coefficient vector and It is used to simulate the divergence phenomenon of wolves searching for prey separately and then surrounding and attacking. When |C|>1 or |A|>1, the global search space is explored; when |C|<1 or |A|<1, the local search space between the gray wolf and the prey is explored. The coefficient vector and They are defined as:

[0036]

[0037]

[0038] In formulas (12) and (13), and is a random number between 0 and 1, During the iteration process, the nonlinear convergence factor a of its internal elements is expressed as:

[0039]

[0040] Furthermore, the maximum completion time MakeSpan in step 2 is defined as:

[0041]

[0042] Where, ETC ij Defined as:

[0043]

[0044] In formulas (15) and (16), D represents the number of virtual machines, L represents the number of tasks assigned to the corresponding virtual machine, and ETC ij represents the time to run task i on virtual machine j, MI represents the cloud task length, and MIPS represents the average execution speed of single-word fixed-point instructions;

[0045] The waiting time WaitTime is defined as:

[0046] WaitTime=Max(WaitTime, WaitTime+ETC ij +commTime) (17)

[0047] Execution Cost is defined as:

[0048] Execution Cost=ε*Cost1+μ*Cost2+ρ*Cost3 (18)

[0049] Cost1=ETC ij ETC ij Cost (19)

[0050] Cost2=commTime·commCost (20)

[0051] Cost3=waitTime·waitCost (21)

[0052]

[0053] In formula (18), ε+μ+ρ=1; in formula (19), ETC ij Cost represents the custom running task cost; commCost in formula (20) represents the custom communication time cost; waitCost in formula (21) represents the custom waiting cost; in formula (22), i represents the cloud task number, outputFile represents the output file size, and dataTransfer represents the cloud task transmission speed.

[0054] Furthermore, the step 3 also includes the behavior of the gray wolf surrounding the prey, which is defined as:

[0055]

[0056]

[0057] Where, represents the position vector of the prey, represents the position vector of the gray wolf individual, and t represents the current iteration number; represents the distance between the gray wolf and its prey, Indicates the position of the gray wolf individual updated in each iteration.

[0058] Compared with the prior art, the present invention has the following technical effects:

[0059] The present invention introduces a hunting strategy based on dimensional learning DLH to give full play to the hunting ability of individual gray wolves. Among them, the DLH strategy has a faster convergence speed than other algorithms by increasing the dimension, and enhances its local search capability by using a scheme based on random position selection in the neighborhood. At the same time, based on dimensional learning of the neighborhood, a balance is achieved between global exploration and local optimization, which can avoid local optimality and thus explore the search space extensively. It can be seen that the cloud platform task scheduling method designed by the present invention based on the search strategy of GWO and the DLH search strategy further enhances the balance between exploration and development, avoids local optimality, reduces task completion time and cost, and improves the resource utilization of virtual machines. It can well balance the global and local search capabilities of the algorithm and improve the availability of the cloud platform.

[0060] Random initialization is used in the initialization process to increase the diversity of the population; in the iterative process, a nonlinearly changing convergence factor is used to fully fit the actual search and development conditions, thereby obtaining better optimization results. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a flow chart of the present invention;

[0062] Figure 2 This is a cloud task scheduling framework diagram of the present invention;

[0063] Figure 3 It is the nonlinear variation diagram of the convergence factor of the present invention; DETAILED DESCRIPTION

[0064] The specific contents of the present invention are further explained in detail below with reference to the embodiments.

[0065] The experimental environment of this embodiment is Windows 10 64-bit operating system. Core TMAn i7-6700 CPU at 3.40GHz and 3.41GHz, with 8GB of RAM, was used in the cloud computing environment. CloudSim, a cloud computing simulation software released by the Grid Lab and Gridbus project at the University of Melbourne, Australia, provides cloud computing features, supports cloud computing resource management and scheduling simulation, and is widely used in cloud platform simulation experiments. This paper used CloudSim 5.0 simulation software for testing experiments and compared the proposed method with the baseline Grey Wolf Optimization (GWO) algorithm and the improved Mean Grey Wolf Optimization (MGWO) algorithm under the same conditions. The performance of the proposed method was evaluated using three evaluation metrics: MakeSpan, Utilization, and Execution Cost.

[0066] like Figure 1 and Figure 2 As shown in FIG, a cloud platform task scheduling method based on dimensional learning strategy and gray wolf optimization includes the following steps:

[0067] Step 1: Initialize the parameters of the DLH-GWO algorithm, including the location information X of the gray wolf individuals in the population. i , maximum number of iterations Maxiter, coefficient vector and and convergence factor a; where:

[0068] The gray wolf population N is randomly distributed in the given search space matrix Position, and the position distribution of gray wolf individuals follows:

[0069] X ij =l j +rand j [0, 1]×(u j -l j ), i∈[1, N], j∈[1, D] (10)

[0070] The position of the i-th wolf in the t-th iteration is expressed as a substantial vector:

[0071] X i (t) = {x i1 , x i2 ,...,x iD} (11)

[0072] In formulas (10) and (11), u j and l j Indicates the upper and lower limits of the search range of each dimension of the gray wolf individual; rand j[0, 1] represents a random number between 0 and 1; D represents the number of tasks, that is, the dimension of the problem. After each iteration, the information of the entire matrix Position will be updated and indicate a better solution state;

[0073] The coefficient vector and It is used to simulate the divergence phenomenon of wolves searching for prey separately and then surrounding and attacking. When |C|>1 or |A|>1, the global search space is explored; when |C|<1 or |A|<1, the local search space between the gray wolf and the prey is explored. The coefficient vector and They are defined as:

[0074]

[0075]

[0076] In formulas (12) and (13), and is a random number between 0 and 1, During the iteration process, the nonlinear convergence factor a of its internal elements is expressed as:

[0077]

[0078] Step 2: Customize the fitness function and select the execution cost, maximum completion time, wait time, and resource utilization to build a multi-objective optimization model. The custom multi-objective fitness function is:

[0079] F=min{Execution Cost, Makespan, Waiting time}+max{Utilization} (1)

[0080] The maximum completion time (MakeSpan) is defined as:

[0081]

[0082] Where, ETC ij Defined as:

[0083]

[0084] In formulas (15) and (16), D represents the number of virtual machines, L represents the number of tasks assigned to the corresponding virtual machine, and ETC ij represents the time to run task i on virtual machine j, MI represents the cloud task length, and MIPS represents the average execution speed of single-word fixed-point instructions;

[0085] The waiting time WaitTime is defined as:

[0086] WaitTime=Max(WaitTime, waitTime+ETC ij +commTime) (17)

[0087] Execution Cost is defined as:

[0088] Execution Cost=ε*Cost1+μ*Cost2+ρ*Cost3 (18)

[0089] Cost1=ETC ij ETC ij Cost (19)

[0090] Cost2=commTime·commCost (20)

[0091] Cost3=waitTime·waitCost (21)

[0092]

[0093] In formula (18), ε+μ+ρ=1; in formula (19), ETC ij Cost represents the custom running task cost; commCost in formula (20) represents the custom communication time cost; waitCost in formula (21) represents the custom waiting cost; in formula (22), i represents the cloud task number, outputFile represents the output file size, and dataTransfer represents the cloud task transmission speed;

[0094] Step 3: Combine the dimensional DLH-based search strategy with the position information X, X, and X of the first three wolves to calculate two candidate updated positions for the remaining wolves in the gray wolf population N. Specifically, this includes encircling the prey and improving the DLH search position, where:

[0095] Gray wolf swarming behavior is defined as:

[0096]

[0097]

[0098] Where, represents the position vector of the prey, represents the position vector of the gray wolf individual, and t represents the current iteration number; represents the distance between the gray wolf and its prey, Indicates the updated position of the gray wolf individual at each iteration;

[0099] In the DLH search strategy, Gray Wolf X i Each dimension of the new position of (t) is expressed as:

[0100] X i-DLH (t+1)=X i，d (t)+rand×(X n，d (t)-X r，d (t)) (2)

[0101] Where, X n，d (t) represents X i (t) The d-th dimension of the adjacent gray wolf, X r，d (t) represents the dth dimension of the gray wolf randomly selected from the matrix Position, X i-DLH (t+1) indicates that the DLH search strategy is gray wolf X i (t) The dth dimension of a new candidate position generated;

[0102] X n，d (t)bioassay in X i The neighborhood N of (t) i (t) is randomly selected, the neighborhood N i (t) is expressed as:

[0103] N i (t) = {X j (t)|D i (X i (t), X j (t))≤R i (t), X j (t)∈Q} (3)

[0104] Where D i For X i (t) and X j (t), the Euclidean distance between i (t) is the neighborhood radius, and its expression is:

[0105] R i (t)=||X i (t)-X i-GWO (t+1)|| (4)

[0106] Where, X i-GWO A location search is defined as:

[0107]

[0108]

[0109]

[0110] Where D α 、D β and D δ Represent the distances between ω wolf and α wolf, β wolf and δ wolf respectively; and is a coefficient vector used to control the degree of divergence when gray wolves search for prey; Represents the location information of w wolf, and Represents the location information of α wolf, β wolf and δ wolf respectively; and It represents the position update of the gray wolf individual caused by the influence of α wolf, β wolf and δ wolf respectively;

[0111] Step 4: Compare the two candidate update positions X i-GWO (t+1) and X i-DLH (t+1) and the fitness value of the current position of the gray wolf individual, select a better position to update the position of the gray wolf individual, expressed as:

[0112]

[0113]

[0114] Step 5: Repeat steps 2 to 4 to perform the same strategy of position adjustment, selection, and update operations on all gray wolf individuals;

[0115] Step 6: Determine whether the tth iteration has reached the maximum number of iterations Maxiter. If t < Maxiter, return to step 2 to continue the next search process; otherwise, output the optimal solution.

[0116] This example simulates scenarios with varying numbers of cloud tasks and virtual machines in a real-world cloud platform environment. The experiments tested the algorithm for different numbers of cloud tasks and virtual machines, including 50, 100, 150, 200, 250, and 300. The results were averaged ten times under the same conditions and compared with a baseline algorithm. The results show that the proposed cloud platform task scheduling method, based on dimensional learning and gray wolf optimization, outperforms the baseline algorithm as the number of cloud tasks and virtual machines increases. Tables 1 to 3 show these results.

[0117] Table 1 Comparison results of the maximum completion time between the DLH-GWO algorithm and the baseline algorithm

[0118]

[0119] Table 2 Comparison of execution costs between the DLH-GWO algorithm and the benchmark algorithm

[0120]

[0121] Table 3 Comparison results of resource utilization between DLH-GWO algorithm and benchmark algorithm

[0122]

[0123] The results presented in Tables 1-3 for the three evaluation metrics (execution cost, maximum make span, and resource utilization) show that the proposed method exhibits superior scheduling performance compared to other methods in terms of maximum make span, execution cost, and virtual machine resource utilization. While reducing maximum make span, it also reduces costs, effectively conserving cloud resources. Furthermore, the proposed method improves global optimization capabilities, fully utilizing resources during the optimization process to achieve the optimal mapping between cloud tasks and virtual machines. It can also adapt to cloud platforms of varying scales, resulting in better overall performance.

Claims

1. A cloud platform task scheduling method based on dimensional learning strategy and gray wolf optimization, characterized in that: Specifically, it includes the following steps: Step 1: Initialize the parameters of the DLH-GWO algorithm, and store the position information of the gray wolf population N in a matrix Position with N rows and D columns in a random distribution state; Step 2: Select the execution cost Execution Cost, the makespan MakeSpan, the waiting time WaitTime, and the resource utilization Utilization to construct a multi-objective optimization model, and customize the multi-objective fitness function as: F = min{Execution Cost, Makespan, WaitTime} + max{Utilization} (1) According to formula (1), the fitness of the gray wolf individuals is calculated, and the position information X of the top three wolves with the best performance is obtained. α 、X β and X δ ; Step 3: Combine the search strategy based on dimensional learning DLH and the position information X of the first three wolves α 、X β and X δ , respectively calculate the two candidate update positions of the remaining wolves in the gray wolf population N. In the DLH search strategy, the gray wolf X i Each dimension of the new position of (t) is expressed as: X i-DLH (t+1)=X i,d (t)+rand×(X n,d (t)-X r,d (t)) (2) Where, X n,d (t) represents X i (t) The d-th dimension of the adjacent gray wolf, X r,d (t) represents the dth dimension of the gray wolf randomly selected from the matrix Position, X i-DLH (t+1) indicates that the DLH search strategy is gray wolf X i (t) The dth dimension of a new candidate position generated; X n,d (t) represents the i The neighborhood N of (t) i (t) is randomly selected, the neighborhood N i (t) is expressed as: N i (t)={X j (t)|D i (X i (t),X j (t))≤R i (t),X j (t)∈Q} (3) Where D i For X i (t) and X j (t), the Euclidean distance between i (t) is the neighborhood radius, and its expression is: R i (t)=||X i (t)-X i-GWO (t+1)|| (4) Where, X i-GWO A location search is defined as: Where D α 、D β and D δ Represent the distances between ω wolf and α wolf, β wolf and δ wolf respectively, and is a coefficient vector used to control the degree of divergence when the gray wolf searches for prey. Represents the location information of w wolf, and Respectively represent the position information of α wolf, β wolf and δ wolf, and It represents the position update of the gray wolf individual caused by the influence of α wolf, β wolf and δ wolf respectively; Step 4: Compare the two candidate update positions X i-GWO (t+1) and X i-DLH (t+1) and the fitness value of the current position of the gray wolf individual, select a better position to update the position of the gray wolf individual, expressed as: Step 5: Repeat Step 2 to Step 4, and perform position adjustment, selection, and update operations with the same strategy for all gray wolf individuals; Step 6: Determine whether the t-th iteration reaches the maximum number of iterations Maxiter. If t < Maxiter, return to Step 2 to continue the next search process; otherwise, output the optimal solution.

2. The cloud platform task scheduling method based on dimensional learning strategy and gray wolf optimization according to claim 1 is characterized in that: The parameters of step 1 include the location information of the gray wolf individuals in the gray wolf population N, the maximum number of iterations Maxiter, the coefficient vector and and the convergence factor a, where: The gray wolf population N is randomly distributed in the given search space matrix Position, and the position distribution of gray wolf individuals follows: X ij =l j +rand j [0,1]×(u j -l j ),i∈[1,N],j∈[1,D] (10) In the t-th iteration, the position of the i-th wolf is represented by a real vector as: X i (t)={x i1 ,x i2 ,…,x iD } (11) In formulas (10) and (11), u j and l j Indicates the upper and lower limits of the search range of each dimension of the gray wolf individual; rand j [0,1] represents a random number between 0 and 1; D represents the number of tasks, that is, the dimension of the problem. After each iteration, the information of the entire matrix Position will be updated and indicate a better solution state; The coefficient vector and It is used to simulate the divergence phenomenon of wolves searching for prey separately and then surrounding and attacking. When |C|>1 or |A|>1, the global search space is explored; when |C|<1 or |A|<1, the local search space between the gray wolf and the prey is explored. The coefficient vector and They are defined as: In formulas (12) and (13), and is a random number between 0 and 1, During the iteration process, the nonlinear convergence factor a of its internal elements is expressed as: 。 3. The cloud platform task scheduling method based on dimensional learning strategy and gray wolf optimization according to claim 2 is characterized in that: The makespan MakeSpan in Step 2 is defined as: Where, ETC ij Defined as: In formulas (15) and (16), D represents the number of virtual machines, L represents the number of tasks assigned to the corresponding virtual machine, and ETC ij represents the time to run task i on virtual machine j, MI represents the cloud task length, and MIPS represents the average execution speed of single-word fixed-point instructions; The waiting time WaitTime is defined as: WaitTime=Max(WaitTime,waitTime+ETC ij +commTime) (17) The execution cost Execution Cost is defined as: Execution Cost = ε * Cost1 + μ * Cost2 + ρ * Cost3 (18) Cost1=ETC ij ·ETC ij Cost (19) Cost2 = commTime · commCost (20) Cost3 = waitTime · waitCost (21) In formula (18), ε+μ+ρ=1; in formula (19), ETC ij Cost represents the custom running task cost; commCost in formula (20) represents the custom communication time cost; waitCost in formula (21) represents the custom waiting cost; in formula (22), i represents the cloud task number, outputFile represents the output file size, and dataTransfer represents the cloud task transmission speed.

4. The cloud platform task scheduling method based on dimensional learning strategy and gray wolf optimization according to claim 3 is characterized in that: Step 3 further includes the behavior of gray wolves surrounding prey, which is defined as: Where, represents the position vector of the prey, represents the position vector of the gray wolf individual, and t represents the current iteration number; represents the distance between the gray wolf and its prey, Indicates the position of the gray wolf individual updated in each iteration.

Images