Resource scheduling optimization method in cloud computing based on intelligent algorithm

Abstract

The invention discloses a resource scheduling optimization method in cloud computing based on an intelligent algorithm. The method comprises the following steps: data acquisition, single resource scheduling task feature correction, resource task and system load state alignment, integration of multi-task scheduling requirements and random generation of different scene tasks as a training data set; constructing an intelligent optimization model: inputting the data set into the constructed intelligent optimization model, and training the model to obtain an optimal resource scheduling strategy; and inputting a real-time cloud computing task demand into the optimized model, and outputting an optimal resource allocation and scheduling scheme. According to the method, the resource scheduling efficiency can be effectively improved, the calculation delay and the resource waste are reduced, and the method can be applied to an actual cloud calculation scene.

Description

Technical Field

[0001] The present invention relates to a cloud computing resource scheduling optimization method based on an intelligent algorithm, and belongs to the technical field of resource scheduling optimization. Background Art

[0002] The core mechanism of cloud computing is to distribute computing tasks across a resource pool consisting of a large number of heterogeneous virtualized servers or virtual machines (VMs), thereby efficiently utilizing resources. As a market-oriented utility, cloud computing requires cloud service providers and users to maximize profits and return on investment. Therefore, advanced resource scheduling strategies are crucial to support a variety of application scenarios, including software execution, user applications, tasks, and workflows. The effectiveness of scheduling strategies directly impacts system performance, such as resource utilization efficiency and operating costs, making them a critical factor in cloud computing systems.

[0003] In cloud computing, due to the dynamic provisioning, allocation, and management of virtual machines, scheduling problems are typically divided into two main layers: the first layer involves scheduling user-submitted tasks and mapping them to a set of available virtual machine resources; the second layer involves mapping virtual machines to hosts, allocating virtual machines to suitable physical hosts for creation or migration. This mapping relationship directly impacts the processing power of cloud computing systems. Optimizing task scheduling can not only significantly improve system resource utilization efficiency but also achieve better results in terms of time and price costs. However, the optimization complexity of scheduling problems is NP-hard, meaning that the solution time increases exponentially with the problem size. Traditional algorithms face the "curse of dimensionality" when faced with high-dimensional problems.

[0004] To solve such complex optimization problems within a reasonable timeframe, metaheuristic algorithms have gained increasing attention in recent years. Compared to traditional methods, metaheuristic algorithms can achieve near-optimal solutions in polynomial time, demonstrating their effectiveness in solving complex optimization problems. In fact, various metaheuristic algorithms and their variants have been widely applied to scheduling problems in various fields, including cloud computing. Improved solution methods based on intelligent optimization algorithms can effectively address the high complexity of insufficient resource utilization and time costs in cloud computing, significantly improving overall system efficiency. Summary of the Invention

[0005] The purpose of the present invention is to overcome the above-mentioned NP-hard problem and provide an improved intelligent algorithm for resource scheduling based on cloud computing that can better solve the problem of high time cost complexity and can be applied to real scenarios.

[0006] The technical solution of the present invention is:

[0007] An improved intelligent algorithm for resource scheduling in cloud computing according to the present invention comprises the following steps:

[0008] Step 1. Data preprocessing: Based on the detailed information of the input task and the underlying available computing resources (e.g., VM), tasks and resources will be mapped according to a certain strategy.

[0009] Step 2. Detailed optimization process of the scheduler, using the following model in a cloud computing setting:

[0010]

[0011] Where M is the number of compute nodes (VMs) {N1, N2, ..., N m} and a set of N computing tasks {T1, T2, ..., T n} and N>M. a ij is the decision variable, a ij =1 means that the i-th task is executed on the j-th VM, otherwise a ij =0, for each i∈[1,N]

[0012] Step 3. In order to characterize the overall processing capacity and resource consumption of the cloud computing system in the task scheduling scenario, three attributes (E n (E t ), S n (S t ), C n (C t )) to represent each resource node. The first two are processing power and load capacity, which can be represented by the node's CPU computing power and memory size, respectively. The concept of resource bandwidth is used as the third attribute to abstract the general resources that a node can provide.

[0013] Step 4. To solve the MOO problem, the matrix is first normalized using the min-max normalization method, and then the three objective functions are expressed as F1, F2, and F3, as shown below. The reason for this normalization is that E n 、S n and C n (and E t 、S t and C t ) are at different scales, and in this case the search path for the optimal solution will be tilted, i.e., f i Large values of will dominate the optimization process and small problems will be completely ignored.

[0014]

[0015] The time cost function and load cost function in the target are composed of F1 and F2, and the price cost function F3 can be expressed as follows. Where P is the price unit, a ij represents the assignment of the i-th task to the j-th virtual machine, E t,i is the processing capacity vector in the resource node of the i-th task, E n,j is the processing capacity vector in the j-th VM resource node. t,i is the vector of load capacity in the resource node of the i-th task, S n,j is the load capacity vector in the j-th VM resource node. t,i is the resource bandwidth vector of the i-th task, C n,j is the resource bandwidth vector of the jth VM. opt is the simplified weight ratio of each cost.

[0016] Step 5. Solve step (4) according to the intelligent optimization based on GTO and improved WHATO algorithm;

[0017] Step 6. An intelligent optimization solution method based on GTO and improved WHATO algorithm is characterized by:

[0018] Step 6.1. Initialization: Using the elite initialization (EOBL) technique, we take advantage of the fact that elite individuals contain more effective information than ordinary individuals. We construct a reverse group from the elite individuals in the current group to increase the diversity of the group, and then construct a new group from the current group and the reverse group. We select the best individuals from the population as the new generation of individuals and enter the next iteration. The mathematical formula of EOBL is as follows: Assume that the average individual in the current population corresponds to its own extreme value As an elite individual.

[0019]

[0020] The above formula is used to solve As the reverse solution of the elite individual, K is a D-dimensional vector on (0,1). α j =max(X i,j ),β j =max(X i,j ),α j =max(X i,j ) and β j =max(X i,j ) are dynamic upper bound and dynamic lower bound respectively. Dynamic bound overcomes the disadvantage of fixed bound that it is difficult to save search experience, so that the elite inverse solution can be located in a narrow search space, which is conducive to the convergence of the algorithm. If the dynamic bound operation leads to If the solution crosses the boundary and becomes infeasible, it can be reset using a random generation method. The reset method is as follows:

[0021]

[0022] Step 6.2. Cauchy Inverse Cumulative Distribution: The Cauchy distribution is a continuous probability distribution. It is also a continuous probability distribution with no mathematical expectation. When the random variable X satisfies its probability density function, it is said to obey the Cauchy distribution. Its probability density function can be defined as:

[0023]

[0024] The location parameter a defines the peak location of the distribution, and the scale parameter b defines the half-width at half maximum. As a probability distribution, it is often referred to as the Cauchy distribution. Physicists refer to this as the Breit-Wigner distribution or the Lorentz distribution. When a = 0 and b = 1, it is called the standard Cauchy distribution. Its probability density can be expressed as follows:

[0025]

[0026] The calculation formula of its cumulative distribution function is:

[0027]

[0028] When the cumulative distribution function of the Cauchy distribution can be calculated as an inverse function, random numbers that obey the uniform distribution can be generated through the inverse transformation method. The inverse function of the cumulative distribution function of the Cauchy distribution can be defined as:

[0029]

[0030] p=randn(1,d)

[0031] Where p is a random number uniformly distributed in the range [0, 1]; d is the dimension of the function; the location parameter a is assigned a value of 0, consistent with the standard Cauchy distribution; and the scale parameter b is assigned a value of 0.01. This distribution makes the algorithm's search less aggressive, thus opening the door to new algorithm design.

[0032] Step 6.3. Tangent flight operator: Similar to Cauchy, both are tangent functions, and the calculation formula is:

[0033]

[0034] v=randn(1,d)

[0035] Where v is a uniformly distributed random number in the range [0,1]; d is the dimension of the function.

[0036] Step 6.4. Adversarial Learning Based on a Convex Lens: Assume that in one-dimensional space, there exists an individual P with a height of h on the coordinate axis interval [lb, ub], and its projection on the X-axis is X (where X is the global optimal individual). Place a convex lens with a focal length of F at the base point O (here, the base point is (lb + ub) / 2). The individual is imaged in an inverted state with a certain height through the convex lens. At this point, the first inverted individual is generated on the X-axis through lens imaging.

[0037] Figure 8 In the example, the global optimal individual X takes O as the base point to find its corresponding reverse individual X * According to the imaging principle of convex lens, we can get:

[0038]

[0039] make =η. η represents the proportional relationship between the stretch factor and the object or image, and is called the stretch factor. ub and lb are the upper and lower limits, respectively. Larger values can be obtained in the early stages of the algorithm iteration, allowing the algorithm to conduct a wider search across different dimensional regions, improving the diversity of the population. Smaller values are obtained in the later stages of the algorithm iteration, allowing for a refined search near the optimal individual and improving local optimization capabilities. The formula for calculating the nonlinear dynamic expansion coefficient is as follows:

[0040] η=η max -(η max -η min )*(t / t max ) 2

[0041] η max and η min are the maximum and minimum scaling factors respectively. max is the number of current iterations. is the maximum number of iterations. By changing the previous formula, we can get the inversion point X * The calculation formula is:

[0042]

[0043] Extending the mathematical formula of inverse learning of convex lens imaging to the D-dimensional search space, we can obtain:

[0044]

[0045] X j and X j * X and X respectively * The j-th dimension vector of ub j and lb jare the jth vectors of decision variables. In the algorithm, the values of each dimension are mapped into space to obtain the inverse solution by performing convex lens imaging on the global optimal individual, which not only avoids interference between dimensions but also expands the search range of the algorithm.

[0046] Step 6.5. While the convex lens reverse learning strategy significantly improves the algorithm's solution accuracy and, to a certain extent, helps the population escape local optima, it cannot directly determine whether the generated new reverse individuals are superior to the original ones. Therefore, a greedy mechanism is used to compare the fitness values of new and old individuals and then select the optimal individual. This continuously yields better solutions and improves the algorithm's optimization capabilities. The mathematical model of the greedy selection mechanism is described as follows:

[0047]

[0048] Where t represents the number of current iterations. new (t) is the new position vector of the candidate gorilla. f(X) is the original position vector of the candidate gorilla. f(X * ) represents the position vector of the candidate gorilla learned through convex lens reverse learning.

[0049] Step 7. WHATO algorithm flow

[0050] Step 7.1 WHATO's optimization behavior simulates this social structure using five different behavioral strategies: two for the exploitation phase and three for the exploration phase. These behaviors include following the silverback gorilla, competing for mating rights, exploring unknown areas, migrating within familiar territory, and moving to other groups. In this model, each gorilla represents a potential solution, with the gorilla in the best position (the optimal solution) designated as the silverback gorilla.

[0051]

[0052] Step 7.2. In the above formula, t represents the number of iterations, GX(t+1) represents the candidate position of the gorilla in the next iteration. X(t) is the gorilla's current position vector, r1, r2, r3, and rand are random values from 0 to 1 updated in each iteration, and the parameter p is a number in the range of 0-1. b and L b Represent the upper and lower limits of the variable respectively. r is a member of a gorilla group randomly selected from the entire population, and is one of the candidate position vectors of the random gorilla. The following formulas can be used to calculate C, L, and H, respectively.

[0053]

[0054] F=cos(2×r4)+1

[0055] L=C×l

[0056] Step 7.3. In the above formula, It represents the current iteration number, MaxIt is the total number of iterations to perform the optimization operation, cos refers to the cosine function, r4 is a random value between 0 and 1, and varies with the number of iterations. l is a random value between -1 and 1. The following equation is used to model the leadership ability of the silverback gorilla:

[0057] H=Z×X(t)

[0058] Step 7.4. The Cauchy inverse cumulative distribution operator and tangent flight are both tangent functions. The former has a relatively uniform step size, with occasional large steps. The tangent flight has a smaller step size and a larger number of steps, sometimes with a large number of steps. This paper uses them as scaling factors to control the overall random walk in the GTO local search mode. The proposed WHATO method adds the Cauchy inverse cumulative distribution operator. With this addition, the distance between the gorilla and the silverback is reduced, allowing the final step size to quickly decrease to the optimal target value. Mathematically, the model can be expressed as follows:

[0059] GX(t+1)=GX(t)+L×M×(GX(t)-X silverback )

[0060]

[0061] L and M are calculated according to the original algorithm, and p is calculated using the formula p = randn(1, d).

[0062] Step 7.5. Meanwhile, GTO's second exploitative strategy, "Adult Female Battle," incorporates a tangent flight operator. The advantage of the tangent flight operator is that it balances exploration and exploitation search. It also controls the step size to avoid under-accuracy and improves the convergence of the GTO algorithm.

[0063]

[0064] Q=2×r5-1

[0065] Q simulates the impact force, v is a random number uniformly distributed in the range [0,1], and r5 is a random value in the range 0 to 1. The combination of the two operators improves the convergence speed and search ability of the WHATO method, making the search mechanism more stable, compensating for the shortcomings of local search, and quickly finding the optimal solution.

[0066] Step 7.6. Calculate the reverse elitist solution and recalculate the fitness value using the greedy selection algorithm. A new silverback gorilla is identified. Then, a convex lens reverse learning strategy is used to combine the newly generated reverse individual with the current best individual to generate a new candidate individual. Finally, the optimal silverback gorilla is identified and its fitness value is returned.

[0067]

[0068] Where ub and lb are the upper and lower bounds. Both search strategies can generate new individuals near the optimal individual, thereby increasing the search capability and finding more suitable individuals, thus escaping the local optimum.

[0069] Compared with the prior art, the present invention has obvious beneficial effects. It can be seen from the above technical solutions that:

[0070] The present invention adopts reverse elite learning to make the initialization population evenly distributed and increase the diversity of the population. Secondly, the degree of population aggregation in the local stage is analyzed, and a reverse learning strategy based on convex lens imaging is introduced to generate a dynamic reverse solution, which effectively improves the search accuracy and convergence speed of the algorithm. At the same time, a tangent flight strategy is added to guide the population to conduct an all-round search in the global optimal hunting search range, maintaining the diversity of the population while effectively avoiding convergence and reducing the probability of the algorithm jumping out of the local optimum. By comparing the 10 benchmark functions of CEC2019 and their Wilcoxon test and Friedman value as well as some CEC2017 functions, it is verified that the improved algorithm has higher optimization performance and stronger robustness. Finally, an optimization test analysis of the actual multi-objective optimization problem was carried out to verify the applicability of WHATO in actual engineering problems, providing a new way to solve complex engineering optimization problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 WHATO flow chart for resource scheduling in cloud computing;

[0072] Figure 2 Pseudocode diagram of the WHATO algorithm for resource scheduling in cloud computing;

[0073] Figure 3 This is the benchmark function diagram of the WHATO algorithm in CEC2017;

[0074] Figure 4 This is the benchmark function diagram of the WHATO algorithm in CEC2019;

[0075] Figure 5 The Wilcoxon test and Friedman value of the WHATO algorithm in the benchmark function;

[0076] Figure 6 This is the box plot of the WHATO algorithm;

[0077] Figure 7 A comparison chart of WHATO and the original algorithm on engineering problems;

[0078] Figure 8 This is the imaging principle diagram. DETAILED DESCRIPTION

[0079] The following is a detailed description of the specific implementation, structure, features and efficacy of the resource scheduling optimization method in cloud computing based on the intelligent algorithm according to the present invention, in conjunction with the accompanying drawings and preferred embodiments.

[0080] A method for optimizing resource scheduling in cloud computing based on an intelligent algorithm comprises the following steps:

[0081] Step 1. See Figure 1 , data preprocessing;

[0082] Step 1.1. Input the population size, the maximum number of iterations MaxT, the number of current iterations T, and other parameters such as β and p;

[0083] Step 2. Use the EOBL reverse elite algorithm to solve the initial solution of the population. Specifically, use the solution of the current individual and formula Calculate the reverse elite solution individual of the current iteration This dynamic boundary overcomes the disadvantage of fixed boundaries that it is difficult to save search experience, allowing the elite inverse solution to be located in a narrow search space, which is conducive to algorithm convergence. If the solution crosses the boundary and becomes infeasible, it can be reset using a random generation method. Correct the reverse elite solution;

[0084] Step 3. Calculate the fitness value X of the gorilla race;

[0085] Step 4. When the current number of iterations it is less than the maximum number of iterations MaxT, use Calculate C and use F = cos(2×r4)+1 and L = C×l to calculate L. Specifically, C, F, and L are all random parameters used to increase the range of population solutions and help escape from local optimal solutions;

[0086] Step 5. For each member of the population, calculate the current position (that is, the current solution) using the following formula:

[0087]

[0088] Specifically, the algorithm simulated this social structure using five different behavioral strategies: two for the exploitation phase and three for the exploration phase. These behaviors included following the silverback gorilla, competing for mating rights, exploring unknown areas, migrating within familiar territory, and moving to other groups. In this model, each gorilla represented a potential solution, with the gorilla in the best position (the optimal solution) designated as the silverback gorilla.

[0089] Step 5. Calculate the current gorilla's racial fitness value X again * :

[0090] Step 5.1. If the current fitness value X * If it is greater than the previous fitness value X, then replace X with X * , otherwise ignore this step;

[0091] Step 6. Update the population's location and search for Silverback Gorilla X silverback (optimal solution for the current number of iterations)

[0092]

[0093] Step 7. For each tribe member, if C is greater than or equal to 1, update the current position of each gorilla using the following formula, otherwise jump to step 8:

[0094]

[0095] Specifically, L, M, and p are all given or calculated parameters, and X silverback is the optimal solution for the current iteration round;

[0096] Step 8. If C is less than 1, update the current position of each gorilla using the following formula:

[0097]

[0098] Q=2×r5-1

[0099] Among them, r5 is a random parameter, The tangent flight operator

[0100] Step 9. Calculate the population fitness value again and update the new X silverback :

[0101] Specifically, it uses the nonlinear dynamic expansion coefficient to conduct a refined search near the optimal individual to improve the local optimization ability. The calculation formula is as follows:

[0102] η=η max -(η max -ηmin )*(t / t max ) 2

[0103] By the maximum scaling factor and the minimum scaling factor η max and η min and the current number of iterations t max Changing the previous formula, we can get the inversion point X * The calculation formula is:

[0104]

[0105] The mathematical formula of convex lens imaging reverse learning is extended to the D-dimensional search space. By performing convex lens imaging reverse learning on the global optimal individual, the values of each dimension are mapped into the space to obtain the inverse solution. This avoids interference between dimensions and expands the search range algorithm. We can obtain:

[0106]

[0107] After recalculating the fitness value using the greedy selection algorithm, a new silverback gorilla is identified. Then, a convex lens reverse learning strategy is used to combine the newly generated reverse individual with the current best individual to generate a new candidate individual. Finally, the silverback gorilla in the best position is identified and its fitness value is returned:

[0108]

[0109] Step 10. Find the silverback gorilla (optimal solution) for the current number of iterations and calculate its usage value f(X * ), and returns the fitness value and the optimal solution;

[0110] The optimal result found each time is saved and the reconstructed CEC test function image is output.

[0111] Specifically, in the process of training the CEC standard test function to find the optimization target, the relatively good results and parameters of each indicator are saved, and the data used for real verification are input into the saved population model, so as to obtain the generated image data using the better population parameters, such as Figure 3 、 4 , 5, 6, and 7, where Figure 3 、 4 , 5, 6 are function images on the standard test set, Figure 7 This is the real data image reconstructed in this embodiment.

[0112] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A method for optimizing resource scheduling in cloud computing based on an intelligent algorithm, comprising the following steps: Step 1. See Figure 1, data preprocessing; Step 1.

1. Input the population size, the maximum number of iterations MaxT, the number of current iterations T, and other parameters such as β and p; Step 2. Use the EOBL reverse elite algorithm to solve the initial solution of the population. Specifically, use the solution of the current individual and formula Calculate the reverse elite solution individual of the current iteration This dynamic boundary overcomes the disadvantage of fixed boundaries that it is difficult to save search experience, allowing the elite inverse solution to be located in a narrow search space, which is conducive to algorithm convergence. If the solution crosses the boundary and becomes infeasible, it can be reset using a random generation method. Correct the reverse elite solution; Step 3. Calculate the fitness value X of the gorilla race; Step 4. When the current number of iterations it is less than the maximum number of iterations MaxT, use Calculate C and use F = cos(2×r4)+1 and L = C×l to calculate L. Specifically, C, F, and L are all random parameters used to increase the range of population solutions and help escape from local optimal solutions; Step 5. For each member of the population, calculate the current position (that is, the current solution) using the following formula: Specifically, the algorithm simulated this social structure using five different behavioral strategies: two for the exploitation phase and three for the exploration phase. These behaviors included following the silverback gorilla, competing for mating rights, exploring unknown areas, migrating within familiar territory, and moving to other groups. In this model, each gorilla represented a potential solution, with the gorilla in the best position (the optimal solution) designated as the silverback gorilla. Step 5. Calculate the current gorilla's racial fitness value X again * : Step 5.

1. If the current fitness value X * If it is greater than the previous fitness value X, then replace X with X * , otherwise ignore this step; Step 6. Update the population's location and search for Silverback Gorilla X silverback (optimal solution for the current number of iterations) Step 7. For each tribe member, if C is greater than or equal to 1, update the current position of each gorilla using the following formula, otherwise jump to step 8: Specifically, L, M, and p are all given or calculated parameters, and X silverback is the optimal solution for the current iteration round; Step 8. If C is less than 1, update the current position of each gorilla using the following formula: Q＝2×r5-1 in, r5 is a random parameter, The tangent flight operator Step 9. Calculate the population fitness value again and update the new X silverback : Specifically, it uses the nonlinear dynamic expansion coefficient to conduct a refined search near the optimal individual to improve the local optimization ability. The calculation formula is as follows: the=the max -(or max -or min )*(t / t max ) 2 By the maximum scaling factor and the minimum scaling factor η max and η min and the current number of iterations t max Changing the previous formula, we can get the inversion point X * The calculation formula is: The mathematical formula of convex lens imaging reverse learning is extended to the D-dimensional search space. By performing convex lens imaging reverse learning on the global optimal individual, the values of each dimension are mapped into the space to obtain the inverse solution. This avoids interference between dimensions and expands the search range algorithm. We can obtain: After recalculating the fitness value using the greedy selection algorithm, a new silverback gorilla is identified. Then, a convex lens reverse learning strategy is used to combine the newly generated reverse individual with the current best individual to generate a new candidate individual. Finally, the silverback gorilla in the best position is identified and its fitness value is returned: Step 10. Find the silverback gorilla (optimal solution) for the current number of iterations and calculate its usage value f(X * ), and returns the fitness value and the optimal solution.

Images