A Deadline-Aware Multiple Workflows Scheduling Strategy In Multi-Processor Edge Servers Environment

Due to the rapid development of scientific computing and smart terminals,new mobile applications such as digital power grids,smart logistics,and drone inspections have attracted a lot of attention from users.These mobile applications usually consist of multiple tasks and there are sequential relationships and data dependencies among the tasks,so their execution process can be represented by workflows.Due to the low energy consumption and low latency of Mobile Edge Computing,more and more users are placing workflow applications on the Mobile Edge side,so the constrained performance of edge servers becomes a key issue in edge infrastructure.Rapid advances in hardware and software technologies have enabled modern edge servers to run compute-intensive applications such as multiplexed edge servers,which are typically equipped with multiple CPU sockets where additional parallel processing layers are implemented and multiple homogeneous CPUs collaborate with each other to increase the number of tasks that can be processed per unit of time.However,due to different user Quality of Service requirements and the NP-Hard nature of distributed resource mapping,it becomes a great challenge to execute multiple parallel workflows on a multi-processor edge server.Therefore,this thesis investigates multiple workflows scheduling in a multi-processor edge server environment with multiple homogeneous CPUs,and the main work and specific contributions are as follows:Therefore,this thesis investigates multiple workflows scheduling problem in the multi-processor edge server environment,with the goal of minimizing the average violation rate of Service Level Agreement of workflows and balancing the load among multiple CPUs.Specifically,the main work and specific contributions of this thesis are as follows:Firstly,this thesis models the multiple workflows scheduling problem in the multiprocessor edge server environment as a multi-objective optimization problem,which comprehensively considers the completion time of the workflow and the load balancing among multiple CPUs.By quantifying the Quality of Service of workflows and the change of CPUs' load,make an effective compromise between the execution efficiency of workflows and load balance of CPUs.Secondly,this thesis transforms the constructed optimization problem into a Markov Decision Process,and designs and implements a multiple workflows scheduling algorithm based on Deep Reinforcement Learning to solve the optimal strategy of the Markov Decision Process.It solves the inherent "over estimation" problem of traditional Deep Q Learning algorithm,and uses a new sample storage mechanism to improve the training speed of neural networks.Finally,this thesis conducts extensive simulation experiments based on five realworld scientific workflows,and evaluates the effectiveness of the proposed method by comparing it with a variety of existing scheduling algorithms.The experimental results show that,the proposed algorithm performs better in minimizing the average violation rate of Service Level Agreement of workflows and achieving CPU load balancing compared with the existing scheduling algorithms.