Research On Key Technologies Of Task And Service Scheduling In Edge Computing Systems

Edge Computing(EC)extends the powerful cloud computing ability to the network edge in the proximity of performing application,which can compensate for the deficiencies of current cloud computing in terms of ultra-low latency and ultra-high reliability,thus satisfying the requirements of emerging Io T applications which are sensitive to performances and resources simultaneously.Therefore,EC has been widely recognized as one of the critical technologies to realize various emerging Io T scenarios.However,schedule of computing tasks,services and related resource managements are vital to guarantee and improve the quality of service(Qo S)and system service capacity due to the heterogeneity of resource-constrained edge systems and performance-sensitive applications.Therefore,this dissertation investigates the approaches to improve the capability of EC systems and the Qo S of application from the perspective of scheduling the transmission of computation offloading,task allocation,edge service placement,and related resource allocation.The main work and contributions of this dissertation are summarized as follows.This dissertation investigates centralized single-server EC systems that employ channel-competed wireless communication protocols.Then,to cope with the transmission collision issue of short-packet computing results,this dissertation proposes a channel-reserved medium access control(Ch RMAC)protocol,which can significantly increase the transmission efficiency during edge computation offloading within the radio access network.The Ch RMAC protocol enables the access point to reserve idle transmission slots for the following computing result by broadcasting the data reception acknowledgment frame,reducing the conflict probability and backoff delay of computing results.Besides,a latency-constraint-aware reservation admission mechanism and a virtual backoff recovery mechanism are designed to improve channel utilization during work procedures and avoid second-time collision issue.Finally,the Ch RMAC is implemented in the ns-3 network simulator based on a designed cross-layer implementation architecture.Simulation results reveal that the proposed Ch RMAC protocol can effectively improve the ratio of obtaining results that satisfy the latency constraint and reduce the service latency.This dissertation investigates edge-cloud systems and proposes a dynamic task allocation and service migration(DTASM)approach based on deep reinforcement learning(DRL),which can realize the DTASM to satisfying vast users’ dynamic computing requirement while meeting the service downtime constraint during service migration.The dissertation comprehensively considers the constraints of Qo S,service downtime,and resource capacities to formulate the DTASM problem in edge-cloud systems.Then,the DTASM problem is decomposed into the sub-problem of task allocation and the sub-problem of user selection according to their resource requirements on each edge server(ES).After that,the user selection problem is reformulated as a knapsack problem and addressed by dynamic programming.Finally,the deep deterministic policy gradient(DDPG)method is exploited to cope with the task allocation problem of massive Io T users with a vast discrete action space.Simulation results show that the proposed approach can realize service-seamless DTSAM while ensuring both Qo S and resource constraints and performs significantly superior to other methods.This dissertation considers an edge computing system where multiple-facilities work corporately with the support of multiple edge servers,and proposes DRL-based joint caching and computing edge service placement(JCCESP)approaches,which can optimize the JCCESP of multiple sensing-data-driven applications in a heterogeneous EC system without the prior knowledge of underlay communication networks.The JCCESP requirement of each sensing-data-driven application is described as a variable-length sequence-to-sequence mapping process from a sequence of requested edge services to a JCCESP action.After that,the dissertation models the JCCESP implementation process of multiple applications as a Markov Decision Process with finite steps and then explores the DRL to tackle the complexity caused by system heterogeneity and limited prior knowledge.A JCCESP policy model based on the Encoder-Decoder model is constructed to realize the functionality of mapping variable-length edge service sequences to corresponding JCCESP actions.Then,the JCCESP policy is separately trained by the REINFORCE-based and DDPG-based methods,in which a weight-averaged-twin-Q-delayed(WATQD)method is proposed to alleviate the training instability caused by the sudden Q estimation bias in the later training stages of DDPG method.Simulation results reveal that the proposed approaches can satisfy the JCCESP requirement of multiple sensing-data-driven applications in a heterogeneous EC system with limited prior knowledge and performs superior to traditional methods.In addition,the WATQD method can significantly improve the training stability.This dissertation investigates a distributed EC system and proposes joint task and computing resource allocation(DJTCRA)approaches based on multi-agent-DRL(MA-DRL)method,which can realize DJTCRA among collaborative ESs with limited prior system knowledge and partially observed system state to optimize the system Quality of Experience(Qo E)of users.This dissertation considers the nonlinear correlation between Qo E and Qo S to formulate the Qo E-maximized DJTCRA problem in the considered distributed EC system and decomposes the problem into a distributed system task allocation problem and a computing resource allocation problem for maximizing system Qo E.Then,the Sigmoidal Programming optimizer is employed to obtain the Qo E-maximized computing resource allocation result after executing a task allocation policy.Meanwhile,this dissertation designs an off-policy centralized training and distributed execution framework based on MA-DRL,in which the DJTCRA policies are trained on remote cloud servers and distributed executed on ESs.Then,the MADDPG approach and an MA Max Entropy approach are separately employed to train the DJTCRA policies.Simulation results reveal that the proposed method can effectively optimize the DJTCRA in distributed EC systems,and the MA Max Entropy performs significantly superior to the MADDPG in terms of training efficiency and converged performance.There are totally 39 figures,6 tables,and 196 references in this dissertation.