Research On Divisible-load Scheduling Algorithms In Dispersed Computing

In recent years,with the rapid development of manufacturing process and mobile communication technology,the performance of Internet of Things(Io T)devices has been gradually improved and their applications have become increasingly rich.However,the Io T devices are geographically dispersed and collaborate only within independent systems.Such system-to-system isolation results in a large number of devices with underutilized computing resources.The Dispersed Computing(DCOMP)architecture proposes that a wide range of devices with computing,communication,and storage modules can be used as general-purpose programmable computing nodes by using DCOMP middleware technology.A well-defined task scheduling strategy among geographically dispersed computing resources can significantly enable the system to provide high throughput,low latency,and low energy consumption computing services to the end users.Therefore,scheduling strategies are the key problem of current research in the field of the Dispersed Computing.This paper is based on the theory of divisible-load,with the optimization objectives of minimizing the computational latency and minimizing the average energy consumption of the system.Mainly for the following two aspects.For the optimization objective of minimizing task computation latency,we construct a dynamic scheduling model of divisible tasks subject to transmission reliability constraints in a dispersed network environment.This model comprehensively considers a range of factors that affect task scheduling decisions,including differences in node computing capabilities,heterogeneous communication modes,and random generation and the arrival of computing tasks.Further,it breaks the limitation of single-hop offloading and can utilize the task recursive partitioning property to support multi-hop offloading.We then propose and design a distributed decision algorithm for this model based on the heuristic thought after analyzing the time complexity of the model solution.The algorithm optimizes the population initialization and crossover operator on the basis of the traditional genetic algorithm.It can reduce the solution space dimension and accelerate the convergence speed of this algorithm,and also solve the problem of cyclic scheduling of tasks among computational nodes.Finally,by analyzing the simulation experimental results and comparing with the current classical scheduling algorithms,it is proved that our designed algorithm is efficient in multiple evaluation dimensions,such as global resource utilization,system throughput,and average computation delay of tasks.For the optimization objective of minimizing the average energy consumption of the system,we construct a scenario model of dispersed computing in an energy-constrained disaster area environment as an emergency computing support application on the basis of the previous chapter.And furthermore,this model takes into account random variables such as computational node mobility and time-varying wireless channel.In this case,the unmanned aerial vehicles and the mobile smart devices within the disaster area form a dispersed network to collaborate on the computational tasks issued by the drone control vehicle.Under the constraint that the queue backlog is stable for a long term,we present two scheduling phase optimization objective functions of maximizing the task distribution and minimizing the average energy consumption of all computing nodes in the system.Subsequently,the two stochastic optimization problems were transformed into general optimization problems under the support of Lyapunov's optimization theory.In the two scheduling phases,we propose scheduling algorithms based on deep reinforcement learning and convex optimization techniques,respectively,and analyze the performance loss of the algorithms.Finally,by designing simulation experiments,it is verified that this application scenario can minimize the average system energy consumption with stable and controllable computing service quality by collaborative decision making of two-phase scheduling algorithms.