Dynamic Task Scheduling Algorithm And Platform Based On Reinforcement Learning

As big data receives more and more attention,the rich value contained in big data is increasingly favored by many industries.By analyzing big data,researchers can extract information that is difficult to obtain under the traditional data scale.And due to the application of cloud computing,more and more enterprises and scientific research institutions tend to use big data analysis platforms to complete data analysis work.The big data analysis platform is for storage,computing and presentation purposes,allowing users to run pre-written analysis tasks in the cloud,or use the services provided by the cloud.The execution of computing tasks often requires the output of other tasks as input,which can be abstracted as a Directed Acyclic Graph(DAG)representation.The big data analysis platform needs to schedule DAG tasks submitted by users based on its computing resources and the resources required by the tasks,so that we can take advantage of the computing resources of the platform and provide users with high-speed and reliable computing services.Based on the above background,this article has launched a research on the task scheduling mechanism under the background of big data cloud computing.In this paper,starting from the scheduling basis of task scheduling,combined with incremental learning,the task execution time prediction is explored,and then based on the reinforcement learning algorithm,the task scheduling under static and dynamic scheduling scenarios is modeled,and the advantages of asynchronous are proposed.The task scheduling model of the action evaluation algorithm,and finally the algorithm model is applied to the actual project to build a big data task scheduling platform.The main work of this paper is as follows:(1)Task time prediction model based on ensemble incremental learning: Task scheduling requires the execution time of the task on each node as the scheduling basis,and the time required for task execution is unknown before the actual execution.Therefore,this paper proposes a task time prediction model based on ensemble incremental learning.The model can model the characteristics of the executed tasks and the task execution environment,and optimize the effect of the model through incremental training.The model uses an ensemble learning algorithm,merges the gradient boosting decision tree model and the neural network model to form the final task time prediction model.(2)Task scheduling algorithm based on asynchronous advantage actor-critic evaluation:Task scheduling only needs to consider the situation of the task itself and the load situation of each computing node when tasks being scheduled.The scheduling process can be abstracted as a Markov decision process.Therefore,this paper applies the reinforcement learning algorithm to DAG task scheduling,models the scheduling process,analyzes the advantages and disadvantages of the DQN algorithm and the policy gradient algorithm,and then combines the advantages of the two to establish a static algorithm based on the asynchronous advantage actor-critic algorithm.Task scheduling model and extend the algorithm to dynamic scheduling scenarios.(3)Big data task scheduling platform:In order to provide task management and scheduling services for the big data analysis platform,this paper builds a big data task scheduling platform based on cloud computing and combines task time prediction model and task scheduling algorithm,and applies the theoretical model studied in this paper to actual projects.As a subsystem of the big data analysis platform,the big data task scheduling platform includes a task management module,a task execution time prediction module,and a task scheduling module.In addition to task scheduling,the platform also provides functions for tasks such as saving,loading,modifying,and executing.Based on a heterogeneous cloud computing environment,this article has carried out experiments for a variety of task type cluster configurations.Experiments show that the task time prediction model can predict the task execution time more accurately,and the incremental learning method can ensure the accuracy of the model while reducing the model training time.Compared with a single model,the ensemble model has improved performance.At the same time,the task scheduling algorithm based on asynchronous dominant action evaluation has better convergence,and can get better scheduling results than traditional heuristic algorithms and ordinary DQN and policy gradient algorithms.At the same time,the development of the big data task scheduling platform enables the full use of system resources during the execution of big data tasks,which is of great value and significance for the research of algorithms and the solution of engineering problems.