Research On Container-Based Task Scheduling And Optimization In Edge Computing

In recent years,with the rapid development of the Internet of Things in many fields,the number of various terminal devices and the demand for data processing tasks have increased sharply.Under the traditional cloud computing paradigm,the transmission of massive data generated by these devices in the core network will aggravate the network load on the one hand,and on the other hand,delay sensitive tasks(such as automatic driving,health monitoring,etc.)will not receive timely response.Therefore,as a supplement and extension of cloud,computing edge computing emerged as the times require.By deploying small servers on the edge of the network close to devices or data sources,edge computing can provide high bandwidth,low latency,and high efficiency computing services to nearby users.But on the other hand,compared with cloud computing data centers,edge servers have relatively limited computing and storage resources,and each edge server can only configure a small number of services to handle corresponding tasks at one time.At the same time,the tasks requested by terminal devices are increasingly complex and varied,which can be divided into independent tasks and complex tasks(also known as DAG tasks,or jobs)composed of multiple interdependent tasks.In order to be processed,each task must be configured with the corresponding container service on its assigned server.Therefore,when a task or job request reaches the edge computing system,it is a challenge to optimally schedule each task according to the optimization objective while considering the container services required by the task and the dependent constraint relations among the tasks.(1)To solve the scherduling problem of online multiple independent task requests and offline single DAG task requests in edge computing,this dissertation proposes online task scheduling algorithm and offline DAG task scheduling algorithm respectively based on Proximal Policy Optimization(PPO),and realizes the joint optimization of task or job completion time and energy consumption.Specifically,the scheduling process of online task request and the scheduling process of DAG task according to priority order are abstractly modeled as Markov decision model.The PPO-based task scheduling mechanism and the corresponding task information input method are designed.The effectiveness and reliability of the proposed algorithm are verified by simulation experiments under different scenarios and different optimization purposes.(2)To solve the scheduling problem of online multiple DAG task requests,we propose a simple and effective online DAG scheduling algorithm CBASGA based on the genetic algorithm architecture.In the initialization population,for each task,the algorithm uses the idea of the earliest completion time to select the server which can finish it the earliest,thus generating the initial optimal scheduling scheme for the whole job and ensuring the lowest performance of CBASGA.Then,at the same time of the crossover and variation of the species,CBAS is used to find a good scheduling scheme with a purpose,which is added to the species for coevolution,so as to improve the probability of generating a good scheduling scheme.Based on the real data set of Alibaba in 2018,we conducted a comprehensive performance simulation experiment on it.The experimental results show that the average job completion time of CBASGA is at least 27.6%shorter than that of the contrast algorithm.At the same time,we also study the effects of different parameter settings on the performance of CBASGA,and prove that the performance of CBASGA is stably better than that of the comparison algorithm.