Cloud task scheduling method based on competitive particle swarm optimization

Abstract

The invention provides an improved competitive particle swarm optimization (CSO) algorithm in order to solve the problems of load unevenness and task completion time optimization in cloud computing task scheduling. Aiming at the problems of slow convergence, easy falling into local optimum, neglect of load balance and the like when the traditional CSO is used for processing the problems, three innovations are carried out: 1, the global exploration capability of the algorithm is effectively enhanced by introducing Gaussian mutation operation to implement position disturbance on winner particles; 2, designing a dual-mode updating strategy, and guiding the population to escape from a local optimal region in an evolution stagnation stage through a differentiated particle updating mechanism; and 3, constructing a multi-objective fitness function, and integrating task completion time (Makespan) and a load balancing index (LB) to realize collaborative optimization of scheduling efficiency and resource utilization. By reconstructing an updating mechanism of the competitive particle swarm, the comprehensive performance of the algorithm in the aspects of scheduling performance and resource utilization rate is remarkably improved, and a more competitive solution thought is provided for the task scheduling problem in the cloud computing environment.

Description

Technical Field

[0001] This invention belongs to the field of cloud computing resource scheduling technology, specifically relating to a task scheduling scheme suitable for cloud computing environments. This scheme is based on an improved competitive particle swarm optimization algorithm, aiming to improve the efficiency of task allocation and resource management. This method can be widely applied to various scenarios such as data centers and edge computing, thereby achieving efficient utilization and optimized configuration of computing resources. Background Technology

[0002] Cloud computing technology provides computing resources over a network, enabling users to conveniently access the services they need on demand without managing the underlying infrastructure. This model, by reducing maintenance costs and improving resource accessibility, is the core advantage of cloud computing. In a cloud computing environment, when users initiate service requests, how to achieve reasonable task scheduling becomes a key issue in ensuring efficient resource utilization. The core objective of task scheduling is to reduce wasted time, improve performance, ensure that the response time of received tasks is within an acceptable range, and maximize resource utilization efficiency. When the tasks and resource configurations involved are relatively simple, the scheduling process is also relatively easy; however, as user demand increases and service quality requirements rise, the complexity and difficulty of task scheduling also increase significantly. To meet these requirements, most task schedulers strive to adopt efficient scheduling techniques.

[0003] Currently, scheduling methods are mainly classified into three categories: heuristic scheduling, metaheuristic scheduling, and hybrid scheduling. Among them, metaheuristic algorithms are particularly suitable for handling large-scale, dynamically changing scheduling problems due to their powerful global search capabilities, good adaptability, and multi-objective optimization potential. However, when faced with high-dimensional and complex problems, the search efficiency of these algorithms often decreases, making it difficult to efficiently coordinate the allocation relationships between a large number of tasks and resources in a cloud computing environment. This, to some extent, limits their application effectiveness in complex cloud environments. Summary of the Invention

[0004] Task scheduling is inherently NP-hard, with its computational complexity increasing exponentially with problem size. Traditional metaheuristic algorithms (such as Particle Swarm Optimization (PSO) and Genetic Algorithm (GA)) often struggle to achieve an effective balance between optimizing both task completion time (Makespan) and resource utilization, leading to a significant contradiction between computational efficiency and goal balance. Against this backdrop, current Competitive Particle Swarm Optimization (CSO) algorithms still suffer from the following key shortcomings when dealing with such complex scheduling problems:

[0005] Easily trapped in local optima: Although CSO improves population diversity through competition, it may still get trapped in local optima in high-dimensional, multimodal problems.

[0006] Low update efficiency: It is difficult to guarantee the quality of winners and losers, which may lead to the risk of suboptimal solutions.

[0007] Single optimization objective: Existing methods mainly focus on minimizing task completion time (Makespan), but do not explicitly model and optimize load balancing, making it difficult to further improve resource utilization efficiency.

[0008] With the continuous development of cloud computing technology, the number of users and their demand for computing resources are constantly growing, leading to a concentration of large-scale data processing and computing tasks in cloud data centers. In response to this trend, this invention integrates recent research findings on intelligent algorithms and applies them to the optimization of cloud computing scheduling strategies. While ensuring overall system coordination, a competitive particle swarm optimization algorithm is introduced to improve the cloud scheduling strategy, aiming to shorten task execution cycles while maintaining a balanced system load. Subsequently, experiments verify the effectiveness of this method.

[0009] To address the above challenges, this invention proposes an improved CSO algorithm, the specific innovative measures of which are as follows:

[0010] • Gaussian Mutation Particle: A Gaussian mutation operation is introduced to perturb the position of the winning particle, enhancing the algorithm's versatility and global search capability. The update formula for the winning particle is as follows:

[0011] x W (t+1)=x W (t)(1+C(0,1))

[0012] In the formula: x W (t+1) is the new particle after Gaussian mutation, and C(0,1) is a random variable that follows a Gaussian distribution.

[0013] • Dual-mode update strategy: When the algorithm evolution process enters a stagnant state, the failed particles update their position and velocity based on the information of the top P% of individuals, as shown in the following formula:

[0014] v L (t+1)=r1v L (t)+r2(x P (t)-x L (t))

[0015] x L (t+1)=x L (t)+v L (t+1)+r1(v L (t+1)-v L (t))

[0016] In the formula: v L(t) represents the velocity of the failed particle in the t-th iteration; x L (t) represents the position of the failed particle in the t-th iteration; x P (t) represents the average position of the top P% of individuals; r1 and r2 are random numbers in the range [0,1].

[0017] When the algorithm evolution process is in a normal state, the failed particles update their position and velocity based on the information of the winning individuals, as shown in the following formula:

[0018] v L (t+1)=r4v L (t)+r5(x W (t)-x L (t))

[0019] x L (t+1)=x L (t)+v L (t+1)+r4(v L (t+1)-v L (t))

[0020] In the formula: x P (t) represents the position of the victorious particle in the t-th iteration; r4 and r5 are random numbers in the range [0,1].

[0021] • Multi-objective fitness evaluation framework: Construct a comprehensive evaluation function F = 0.5 × Makespan + 0.5 × LB, where Makespan represents the latest time of completion of all tasks and is used to measure the total execution time of task scheduling; LB is a load balancing index, which reflects the balance of resource load distribution in the system by calculating the variance of the expected processing time of each virtual machine. Attached Figure Description

[0022] Figure 1 This is the algorithm flowchart. Detailed Implementation

[0023] This article illustrates the execution process of the algorithm in cloud computing task scheduling through a specific example. The system demonstrates the configuration of input parameters, processing flow, and final output results. This example is not derived from simulation experiments, but aims to help understand the core ideas and operating mechanisms of the algorithm in a more intuitive way.

[0024] The initial task set contains 5 tasks to be processed, labeled T0 to T4. The system is configured with 3 virtual computing nodes, numbered VMO, VM1, and VM2, with corresponding processing capacities (MIPS) of 1000, 1500, and 500, respectively. Task lengths are set to 2000, 3000, 1500, 2500, and 1800, respectively. For the employed competitive particle swarm optimization algorithm, the total number of particles is set to 4, and the maximum number of iterations is 2, to more clearly demonstrate the algorithm's execution process.

[0025] The algorithm is then executed. First, the particle information is initialized; the initial scheduling scheme for the four particles is as follows:

[0026] The initial information for x1 is [VM1, VMO, VM2, VM1, VM0].

[0027] The initial information for x2 is [VMO, VM1, VMO, VM1, VMO].

[0028] The initial information for x3 is [VMO, VM2, VM1, VM2, VM1].

[0029] The initial information for x4 is [VM1, VMO, VM2, VM1, VM1].

[0030] Next, the particles are randomly divided into two groups to compete: x1 and x2 in one group, and x3 and x4 in another. The two particles in each group compete against each other. The fitness value of each particle is calculated. Taking x1 as an example:

[0031] The VMO executes tasks T1 and T4, with a completion time of (3000 / 1000 + 1800 / 1000) = 4.8 seconds.

[0032] VM1 executes tasks TO and T3, with a completion time of (2000 / 1500 + 2500 / 1500) = 3.0 seconds.

[0033] VM2 executes task T2, and the completion time is: (1500 / 800) = 1.875s.

[0034] Therefore, the maximum completion time (Makespan) of particle x1 is 4.8s, and the load balancing metric is: The value of 3.225 represents the average of the expected processing times of the three virtual machines. Based on the calculation rule of the multi-objective fitness function F = 0.5 × Makespan + 0.5 × LB, the final fitness value of this particle is 3.4965.

[0035] Similarly, we can deduce that the fitness value of particle x2 is 4.5695, the fitness value of particle x3 is 5.3428, and the fitness value of particle x4 is 2.9221. Based on the principle that lower fitness values ​​are better, the winning particles are x1 and x4, and the losing particles are x2 and x3.

[0036] The particles are updated. Winning particles (x1, x4) undergo Gaussian mutation, while losing particles update their positions and velocities based on the information from their winning counterparts. After the update, the particle scheduling scheme is as follows:

[0037] x1 is updated to [VM1, VMO, VM2, VMO, VMO]

[0038] x2 is updated to [VM1, VMO, VM1, VM1, VMO]

[0039] x3 has been updated to [VMO, VMO, VM1, VMO, VM1]

[0040] x4 updated to [VM1, VMO, VM2, VM1, VM1]

[0041] This completes the first iteration of the update process.

[0042] In the second iteration, the particles are randomly divided into two groups to compete: x1 and x4 in one group, and x2 and x3 in another group. The two particles in each group compete against each other.

[0043] Calculations show that the fitness values ​​for particle x1 are 4.44, x2 is 3.918, x3 is 4.826, and x4 is 2.9221. Based on the principle that lower fitness values ​​are better, x2 and x4 are the winning particles in this round, while x1 and x3 are the losing particles.

[0044] Suppose that the evolution process of the algorithm has stalled at this point, and the parameter p = 1, meaning that the failed particle will update its position and velocity based solely on the information of the globally optimal individual x4.

[0045] The updated particle scheduling scheme is as follows:

[0046] x1 is updated to [VM1, VMO, VM2, VMO, VM1]

[0047] x2 is updated to [VM1, VMO, VMO, VM1, VM2]

[0048] x3 has been updated to [VMO, VMO, VM2, VM1, VMO]

[0049] x4 is updated to [VM2, VMO, VM1, VM1, VM1].

[0050] The fitness values ​​of each particle are recalculated: the fitness value of particle x1 is 4.116, the fitness value of particle x2 is 3.0601, the fitness value of particle x3 is 4.1537, and the fitness value of particle x4 is 2.7658. The optimal solution is x4, and its scheduling scheme is: [VM2, VMO, VM1, VM1, VM1].

[0051] Analysis of the execution process and results of this example reveals that the Gaussian mutation operation effectively breaks the limitations of local optima by perturbing the position information of particles, enhancing the algorithm's population diversity and global search capability. The dual-mode update strategy, during periods of stagnant algorithm evolution (such as slowed convergence or getting stuck in local optima), accelerates the convergence process through a differentiated strategy (e.g., failed particles preferentially refer to the global optimum). These improvements work synergistically, not only avoiding insufficient local search capability due to premature convergence but also effectively overcoming the efficiency degradation caused by overexploration.

Claims

1. A cloud task scheduling method based on an improved competitive particle swarm optimization algorithm, the specific process of which is as follows: (1) Environment modeling: Initialize and configure the virtual node resources in the cloud service, and build a resource list, including basic parameters such as virtual machine computing power (e.g., MIPS) and task set workload. (2) Algorithm parameter initialization: Set task set parameters, virtual machine node related parameters, population size and preset iteration limit, and initialize particle position and velocity information. (3) Fitness value calculation, which calculates two core indicators: a. Maximum Task Completion Time (Makespan): The latest time for a task to complete across all virtual machines, used to measure the overall efficiency of the scheduling scheme. b. Load Balancing Metric (LB): The standard deviation of the expected processing time of a node, reflecting the degree of balanced distribution of tasks across different virtual machines. Through proper scheduling, the resource utilization of each node can be made more balanced. The multi-objective optimization problem described above is transformed into a single comprehensive fitness value by using a weighted summation method, which serves as a unified standard for evaluating the quality of scheduling schemes. (4) Competition between particles: The population is divided into m / 2 pairs (assuming the population size m is an even number). The two particles in each pair are judged by evaluating the fitness index: the particle with the better fitness value wins, and the other particle loses. (5) Winning Particle Update: This mechanism aims to improve the algorithm's global exploration capability while avoiding the risk of getting trapped in local optima. By performing a Gaussian mutation operation on the winning particle, its position information is perturbed, thereby effectively enhancing the diversity of the population. (6) Failure Particle Update: a. Normalization phase: Failed particles update their position and velocity based on information from the corresponding winner; b. Evolutionary Stagnation Phase: When the algorithm enters a stagnant state, failed particles update their own state based on the information of the top P% of high-quality individuals in order to accelerate convergence. (7) Update the global optimal solution: After completing the particle state update, recalculate the fitness value of each particle and update the global optimal solution and its corresponding scheduling strategy accordingly. (8) Termination Condition and Result Output: When the iteration process reaches the preset maximum number of iterations, the algorithm stops running. At this time, the position information of the current best particle is output, and the final task scheduling strategy is generated. Key output metrics include: maximum task completion time (Makespan), load details for each virtual machine, and load balancing metrics (LB).

Images