Research On DAG-based Task Scheduling In Mobile Edge Computing

With the rapid development of Internet of Things,5G and 6G,various new,computationintensive and latency-intensive applications emerge and supply us with great convenience,such as virtual reality and Vehicle to Everything,and so on.Simultaneously,these new applications bring great challenges to user equipments limited by computation power and battery lifetime.Mobile edge computing(MEC)provides rich computation resources and reliable computation services at the edge of the access network.Offloading tasks to nearby MEC servers is able to reduce execution latency of applications and energy consumption of user equipments,and improve execution reliability of applications.Practical applications are composed of multiple submodules with interdependency and can be modeled as directed acyclic graphs(DAG).In an MEC system,it is a crucial problem to coordinate computation and communication resources to decide all subtasks’ execution order and location to meet users’ requirements under the premise of ensuring dependencies among subtasks.In this thesis,in an MEC scenario with single user equipment and a set of heterogeneous MEC servers,oriented DAG applications and aiming at various task requirements,we carry out the following works:(1)To meet applications’ requirements for execution reliability and considering the applications’ sensitivity to execution latency and user equipments’ limited battery lifetime,a task scheduling problem is studied to maximize a DAG application’s execution reliability in this thesis.By designing a decomposition method,the constraints including the execution latency and energy consumption of the user equipment for the entire DAG application are broken onto each subtask.Finally,all subtasks are allocated to the processors that satisfies the given constraints and make the overall execution reliability the highest.Simulation results show that the proposed algorithm achieves higher execution reliability under the premise of satisfying the given constraints.(2)To satisfy users’ different requirements for execution latency and energy consumption of user equipments in different task scenarios,a multiobjective task scheduling problem is studied in this thesis.Objectives are to minimize an application’s execution latency and energy consumption of the user equipment.Considering the satisfying performance and disadvantages of multiobjective cuckoo search(MOCS)algorithm,we propose an improved multiobjective task scheduling algorithm by designing an encoding scheme for solutions,improving the update direction of Lévy flight,and designing an external archive and its update strategy.Simulation results demonstrate that the proposed algorithm obtains a series of task scheduling solutions and these Pareto optimal solutions perform excellently in the aspect of convergence,diversity,and uniformity.(3)To improve the generality of task scheduling policies and help users adjust task scheduling schemes adaptively,a task scheduling problem is studied for latency-sensitive applications in this thesis.The goal is to minimize a DAG application’s execution latency.By modeling the scheduling process of DAG applications as Markov decision process(MDP)and designing its key elements,we propose an adaptive task scheduling scheme and an algorithm based on deep Q-network(DQN).Simulation experiments show that the proposed algorithm is able to adaptively adjust scheduling strategies in different MEC environments,and obtain task scheduling schemes with lower latency.