Research On Task Scheduling For Serverless Edge Computing

With the proliferation of Internet of Things(IoT),edge computing has emerged as a key driving force to cope with the explosive growth of data by delivering computing and storage resources in close proximity to the network edge.Meanwhile,benefit from the development of serverless computing,an edge server can be configured as a carrier of limited serverless functions,in the way of deploying Docker runtime and Kubernetes engine.As a new paradigm,serverless edge computing allows users to execute their differentiated applications,especially workflow applications with complex inter-task dependency,without managing the underlying servers.In serverless edge computing scenarios,resources are managed and orchestrated by the third-party provider.Due to the resource-finite nature of edge infrastructure,it has limitations for hosting diverse services,which are decomposed into stateless functions.Each edge computing node can only deploy a portion of services.Consequently,different from traditional edge computing scenarios,computation tasks can no longer be executed on any accessible edge computing resources but can only be executed on edge computing nodes where corresponding services are deployed.In view of this,this paper focuses on how to effectively schedule computation tasks by considering the distribution of edge resources and the deployment of serverless functions.The main contributions are summarized as follows.(1)Static multi-objective workflow scheduling is investigated.A scheduling model and computation model is proposed and the scheduling can be divided into three phases:the offloading,the execution and data transmission.Workflow applications are structed as a Directed Acyclic Graph(DAG)composed of dependent tasks.Considering serverless computing charges for per-subsecond use of computational resources,the solution is obtained by minimizing the makespan of the workflow,the energy consumed by user device and the overhead for users,turning the initial scheduling problem into a multi-objective optimization problem.Finally,a new hybrid PSO-GA algorithm based on the Particle Swarm Optimization(PSO)and Genetic Algorithm(GA)using critical path is introduced,which can reduce the execution time and improve the efficiency.(2)Workflow scheduling in dynamic and time-varying environment is investigated.Facing the dynamic change of communication resources,computation resources and quantized workloads of servers,the workflow scheduling problem is modeled as a Markov Decision Process(MDP).A heterogeneous scheduling priority calculation method is designed,which makes the scheduling order more reasonable.Considering the number of system states is enormous and the choice of allocated position is a discrete action,inspired by Deep Reinforcement Learning(DRL)method,a Dueling Double Deep Q-network(D3QN)-based scheduling scheme is described.Furthermore,a global load balancing factor is introduced to make the load on each computing node relatively balanced.This paper compares the proposed algorithms with various algorithms through simulation experiments.Experimental results show that the two proposed algorithms can make rational scheduling decisions in their corresponding scenarios.