Research On Multi-objective Evolutionary Algorithm Based On RBF Surrogate Model And Adaptive Local Search For Cloud Workflow Scheduling

Cloud computing is a model that uses the Internet and the central remote server to provide users with dynamic and scalable services,just with little underlying software or hardware management and interactions between service providers and end users.It can quickly configure and release the configurable computing resources(CPU,storage devices,applications,networks,etc)in the shared resource pool.Users can access cloud computing services in real time and conveniently at any location and any terminal without knowing the details of key technologies such as internal cloud storage and virtualization.When cloud service providers provide users massive virtual resources,the most critical problem is how to schedule resources economically and efficiently,and how to complete the execution of workflow with the minimum cost(execution time,budget,transmission cost,resource utilization,etc.).In the traditional cloud workflow scheduling algorithms,it is often optimized for a single objective of time or cost.However,according to the characteristics of pay-on-demand in cloud computing environment,the single objective optimization of cloud workflow scheduling algorithms can't meet the growing demand of users,and users prefer to get the best service at the lowest cost,that is,to improve the service quality at the same time to minimize operating costs.In order to implement the synchronous optimization of execution time and multiple execution costs of scientific workflow scheduling in cloud environment,aiming at the problem of multi-objective workflow scheduling in IaaS environment,the Surrogate-assisted Adaptive Local Search Multi-objective Optimization Algorithm(S-ALSMO)is proposed.The specific work is as follows:(1)The implementation and performance analysis of cloud workflow scheduling algorithm based on evolutionary multi-objective optimization are made.In order to achieve the multi-objective simultaneous optimization of the completion time,execution cost and transmission cost of scientific workflow scheduling in cloud environment,the process and characteristics of cloud workflow scheduling are analyzed,the real coding mechanism of virtual machine type are designed.Besides,the fitness functions corresponding to different scheduling models are defined,and two different cloud workflow scheduling models are constructed,respectively based on the execution time and transmission cost,the execution time and execution cost.In addition,based on the evolutionary algorithms(MOEA/D and NSGA-?),the population diversity is implemented by introducing a variety of different crossover and mutation operators in the evolutionary optimization process to explore the influence of different combinations of genetic operators on optimization performance.Four types and three scales of real scientific workflows are used as test data to complete multiple groups of comparative experiments and performance analysis of the algorithms.(2)S-ALSMO,a cloud workflow scheduling algorithm based on RBF surrogate model and adaptive local search,is proposed.The algorithm is applied to two cloud workflow scheduling models based on execution time and transmission cost,execution time and execution cost.While solving the expensive multi-objective optimization problem with high-dimensional decision variables,MOEA/D algorithm can't get the distributed uniform solution.So,based on the solution idea of MOEA/D algorithm,the RBF surrogate model is utilized,and the unknown solution is predicted by establishing the approximate model of the real complex objective function,so as to get a more uniform solution,and greatly reduce the cost of function evaluations.In addition,to accelerate the convergence speed of the algorithm while getting the uniformly distributed solution,and make the Pareto optimal solution set more stable,the adaptive local search strategy is introduced in the surrogate-assisted evolutionary algorithm.(3)Through the simulation experiments of four kinds of real scientific workflows(Cybershake,Epigenomics,Inspiral and Montage)on the simulation platform of workflowsim,the performance of the proposed S-ALSMO algorithm is compared with that of the traditional MOEA/D algorithm and NSGA-? algorithm.The experimental results show that the S-ALSMO algorithm based on surrogate model and adaptive local search achieves faster convergence speed and more uniform Pareto solutions in the combination optimization of data transmission cost and task execution time,task execution cost and task execution time.That is to say,some improvements have been made in reducing task execution time,data transmission cost and other costs,which can better satisfy the needs of users.