| With the explosion of Internet application and big data technology,the cloud computing industry is booming,and users can obtain the resources they want from the cloud computing environment according to their own needs.Cloud computing platform needs to schedule and execute a large number of tasks submitted by users.How to design an effective task scheduling scheme to improve the efficiency of task execution and resource utilization,so as to meet the goals of both users and cloud service providers,is a key issue to be solved in the field of cloud computing for further development.The existing research methods can be roughly divided into three categories: traditional scheduling,heuristic scheduling and hybrid scheduling algorithms.Most of them only consider a single cloud computing task scheduling scenario and an optimization objective,which have certain drawbacks.Aiming at the task scheduling problem in the above cloud computing environment,this paper considers the scene characteristics of the task arriving at the center of cloud computing cluster,and proposes two algorithms for different task scheduling scenarios.For the scenario of static task scheduling in cloud computing environment,several optimization goals are set up,and a particle swarm algorithm based on dynamic adjustment parameters is proposed.For the scenario where tasks dynamically and continuously reach the cluster center in cloud computing environment,trust region policy optimization based on task dynamic visit is presented using a real-time scheduling reinforcement learning method.The main contents of this paper include the following:(1)Aiming at the static task scheduling scenario in cloud computing environment,Particle Swarm Optimization Algorithm based on Dynamic Parameter Adjustment is proposed.The optimization objectives of the system are described mathematically in three aspects: the total execution time of the tasks,the total cost of the system and the total execution cost of the tasks.The original particle swarm algorithm is improved,and a dynamic correction learning factor and an adaptive inertia weight optimization strategy are proposed,which enable the algorithm to modify the relevant parameters by itself in the training so as to make the original particle swarm algorithm better.The algorithm can jump out of the local optimal solution and search for the global optimal solution in a larger space.(2)Trust Region Policy Optimization based on Dynamic Task Scheduling is proposed for cloud computing environments where tasks visit cluster centers dynamically and continuously.Different from the static task scheduling scenario,in this scenario,tasks continuously arrive at the cloud computing center dynamically,and an algorithm with faster real-time response speed and closer interaction with the environment is required.Therefore,Trust Region Policy Optimization in reinforcement learning is adopted.Firstly,Markov modeling is carried out based on the task scheduling problem model,and mathematical definitions are made in three factors,namely state space,action space and reward function.For the state space,the idle resources of the cluster and the resources required by the task are represented discretely by a matrix.For the action space,in order to reduce the dimension of the action space,the strategy of running multiple actions in the same time step for many times and taking action for only one task at a time step is adopted.The training of the current time step is finished until no task can be scheduled.In the training process,Taylor expansion is used to replace the objective function and constraint optimization approximatively to accelerate the convergence of the algorithm.In the experiment of this paper,the existing Cloud Sim cloud environment simulation system and the self-simulated cloud computing task scheduling simulation environment are respectively adopted.Under the strict control constraints,the two algorithms proposed in this paper are compared with the mainstream task scheduling algorithms and the experimental results are evaluated.Compared with baseline algorithms,Particle Swarm Optimization Algorithm based on Dynamic Parameter Adjustment has an average optimization effect of 16.93% in the makespan,and an increase of 24.60% in the task average resource cost.Compared with baseline algorithms,Trust Region Policy Optimization based on Dynamic Task Scheduling has an average of 80.11% improvement in reward.The experimental results show that the DPA-PSO algorithm and DTS-TRPO algorithm presented in this paper can perform better under different problem scales and constraints. |
