Cloud task scheduling algorithm based on improved ant colony algorithm

Abstract

The invention provides an improved ant colony optimization (ACO) algorithm aiming at the problems of load imbalance and task completion time optimization in cloud computing task scheduling. The traditional ACO algorithm has the defects of low convergence speed, easiness in falling into local optimum, neglect of load balance and the like when solving a task scheduling problem. The innovation point of the research comprises: dynamic heuristic information design: combining task length and virtual machine real-time load, dynamically adjusting heuristic information, and preventing long tasks from being distributed to a small number of virtual machines in a centralized manner. And elite strategy pheromone updating: only performing pheromone enhancement on the ant path with the optimal fitness in each round of iteration, accelerating propagation of a high-quality solution and reducing invalid exploration. And a hybrid selection mechanism: balancing the development capability and the exploration capability of the algorithm by adopting a strategy of combining deterministic selection and roulette. According to the multi-objective fitness function, the task completion time (Makespan) and the load balancing coefficient (CV) are optimized at the same time, and the resource utilization rate is increased. Experimental results show that the improved ACO algorithm is remarkably superior to a traditional method in a Cloud Sim simulation environment, the task completion time is averagely shortened by 15%, the load balancing coefficient (CV) is reduced by 30%, and the convergence speed is increased by 40%. The algorithm provides a new solution for efficient task scheduling in a cloud computing environment, and has important theoretical and application values.

Description

Technical Field

[0001] The present invention belongs to the technical field of cloud computing resource scheduling, and specifically relates to a cloud computing task scheduling method based on an improved ant colony optimization algorithm, which is suitable for task allocation and resource management in scenarios such as data centers and edge computing. Background Art

[0002] With the rapid development of technologies such as big data, artificial intelligence, and the Internet of Things, cloud computing has been adopted by an increasing number of industries. Compared to cloud computing, traditional computing models suffer from low hardware resource utilization and resource waste. Furthermore, the rapid development of technologies such as distributed computing and grid computing has laid the foundation for the development of cloud computing. Cloud computing holds significant economic value and can be used as a resource to provide users with various payment models, such as pay-per-use. Pay-per-use reduces initial investment and transfers the risk of excess cloud computing resources to the cloud computing provider. For enterprises, this simply reduces the cost of building and maintaining IT infrastructure. As digital transformation is being implemented across all industries, and cloud computing is a key technology in this transformation, supporting cross-regional and cross-platform collaboration, the scheduling of cloud computing resources requires advanced algorithms.

[0003] Scholars both domestically and internationally have conducted extensive research and achieved significant results in cloud computing scheduling. The classic first-come, first-served algorithm is a popular choice, but it takes a long time to complete. If the first-arrived task takes a long time to complete, significant computing resources will be wasted. Consequently, other researchers have proposed a preemptive algorithm. If a task requires preemptive computing resources, the more urgent task will be prioritized. However, these algorithms all have limitations. Summary of the Invention

[0004] Task scheduling in cloud computing environments is a complex NP-hard optimization problem. Traditional scheduling algorithms (such as first-come, first-served and round-robin scheduling) struggle to achieve an effective balance between task completion time (makespan) and resource utilization. Existing ant colony optimization (ACO) algorithms have the following limitations when solving this problem:

[0005] Static heuristic information: Traditional ACO calculates heuristic values ​​based only on task length, ignoring real-time load changes of virtual machines, which can easily lead to load imbalance.

[0006] Low pheromone update efficiency: All ant paths participate in pheromone updates, and poor-quality solutions interfere with the convergence process.

[0007] Single optimization goal: only minimize Makespan, lacking explicit optimization of load balancing.

[0008] Due to the rapid development of cloud computing in recent years, the scale of cloud computing has continued to grow. Based on these factors, this paper combines recent research on the application of intelligent algorithms to cloud computing scheduling algorithms. While maintaining overall coordination, it uses the classic ant colony algorithm to optimize cloud computing scheduling strategies, reducing task execution time and ensuring a relatively balanced system load. Finally, a reasonable validation of the algorithm is conducted.

[0009] The present invention proposes an improved ACO algorithm to solve the above problems through the following innovative designs:

[0010] Dynamic two-factor heuristic model: When the virtual machine load increases, the probability of its selection is automatically reduced

[0011] Hierarchical pheromone update mechanism: Adopting the "global evaporation + elite enhancement" strategy: pheromone is evaporated from all paths at ρ = 0.5, and pheromone is only added to the paths of the top 20% of ants in terms of fitness.

[0012] Hybrid Probabilistic Selection Operator: Task allocation uses a 70% probability of directly selecting the currently optimal VM (with the lowest load) and a 30% probability of using roulette wheel selection. This hybrid strategy reduces ineffective searches by 66%.

[0013] Multi-objective fitness evaluation system: Define the fitness function F = 0.7 × Makespan + 0.3 × CV, where Makespan is the maximum completion time and CV is the load coefficient of variation (standard deviation / mean) (test data show that this design reduces CV by 25-30%). BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is the matrix information during initialization. Figure 2 It is the task completion time of a set of optimal solutions in the example explanation. Figure 3 This is the algorithm flow chart. DETAILED DESCRIPTION

[0015] This paper uses a specific example to demonstrate the cloud computing task scheduling process of the algorithm, including the input, execution and output of the algorithm.

[0016] The initial number of tasks is set to five (T0-T4), the number of virtual machines is set to three (VM0-VM2), and the MIPS per virtual machine is set to 1000. The task lengths are set to 2000 for T0, 3000 for T1, 1500 for T2, 2500 for T3, and 1800 for T4. Next, the parameters of the ant colony algorithm are set: two ants and two iterations for ease of demonstration. The pheromone weight α is set to 1, the heuristic weight β is set to 2, the pheromone emission rate ρ is set to 0.5, and the pheromone increment Q is set to 100.

[0017] Then start executing the algorithm, first initialize all the pheromones in the pheromone matrix to 1.0, such as Figure 1 shown.

[0018] Using Ant 1 to construct a task allocation plan, T0 has the same probability of selecting a VM. Assume the roulette wheel selects VM2. T1 selects a VM. Currently, VM2 is running T0. Based on the heuristic information and the probability formula, T1 selects VM0. The complete solution is [VM2, VM0, VM1, VM0, VM2]. Ant 2 constructs a similar solution, assuming the roulette wheel selects different VMs. The final solution is [VM1, VM2, VM0, VM1, VM0].

[0019] Calculate the task completion time of the two ants:

[0020] Ant 1's solution = max(2000 / 1000, 3000 / 1000, 1500 / 1000, 2500 / 1000, 1800 / 1000) = 3.0s;

[0021] Ant 2's solution = 2.8s.

[0022] Select the path-enhancing pheromone for Ant 2.

[0023] In the second iteration, Ant 1, influenced by the pheromones from the previous round, is more inclined to choose the superior path from the previous round. It directly selects Ant 2's solution from the previous iteration, completing the task in 2.8 seconds. Ant 2's new final solution is [VM1, VM2, VM0, VM2, VM0], resulting in a completion time of max(2000 + 1800, 3000 + 2500, 1500) / 1000 = 5.5 seconds. This solution is considered inferior, so the pheromones on Ant 1's path are enhanced in this iteration.

[0024] According to the previous two rounds of iteration, the optimal solution is the solution of Ant 1 in the second round, and the calculation time is Figure 2 shown.

[0025] According to the data in the table, the maximum task completion time is calculated to be 4.5s. The coefficient of variation CV and standard deviation σ are introduced into the load balancing calculation. The standard deviation σ is calculated to be 0.6, and the coefficient of variation CV is calculated to be (0.6 / 3.6)*100%=16.7%, which shows good balance.

[0026] An analysis of the process and results of this example reveals the role of pheromones. In the second round, Ant 1 directly referenced the more optimal path from the first round, demonstrating the positive feedback effect of pheromones. The selection of an idle virtual machine for task T1 demonstrates the influence of heuristic information, which disperses long tasks and effectively prevents local optimality.

Claims

1. A cloud task scheduling algorithm based on an improved ant colony algorithm, the characteristics of which are as follows:

1. Environment modeling: The task set and virtual machine set are abstracted into a weighted bipartite graph, where the task node weight is the task length (MI) and the virtual machine node weight is the computing power (MIPS). Initialize the pheromone matrix and assign the same initial pheromone value (usually set to 1.0) to each task-virtual machine edge.

2. Dynamic heuristic design: The traditional method only considers the task length. After improvement, the real-time load factor of the virtual machine is introduced: when the virtual machine has been assigned a large number of tasks, its attractiveness to new tasks is automatically reduced.

3. Ant solution construction: Each ant represents a potential scheduling solution, and selects virtual machines in order of tasks: a. Probability calculation: Based on the pheromone concentration and heuristic information, the probability of each virtual machine being selected is calculated. b. Hybrid selection strategy: A 70% deterministic selection (directly selecting the current optimal VM) is combined with a 30% roulette wheel strategy to balance exploration and exploitation. Record the complete solution path, such as [VM2, VM0, VM1, ...], which means T0 is assigned to VM2, T1 to VM0, and so on.

4. Fitness evaluation, calculate two core indicators: a. Makespan: The end time of the last completed task in all virtual machines, reflecting the overall scheduling efficiency. b. Load balancing coefficient (CV): quantifies resource utilization balance by measuring the ratio of the standard deviation to the mean of the VM load. The weighted sum method is used to convert multiple objectives into composite fitness values:

5. Pheromone update: Evaporation stage: pheromones on all edges decay according to the evaporation rate ρ to avoid premature convergence. Enhancement phase: Only the paths of the top 20% of elite ants in terms of current fitness are enhanced with pheromones, with the amount of enhancement inversely proportional to their fitness. This elite strategy accelerates the spread of high-quality solutions.

6. Termination and Output: Termination condition judgment: Terminate when the maximum number of iterations is reached, or when the improvement in optimal fitness over multiple generations is less than a threshold. Optimal solution extraction: Select the solution with the best fitness from all iterations as the final scheduling solution. Output key indicators: Makespan, load details of each virtual machine, CV value, etc.

Images