Research On Scheduling Technology To Minimize MXNet Resource Lease Over Public Cloud

In recent years,deep learning technology has shown great advantages in image classification,speech recognition,and natural language processing.MXNet has outstanding performance in speed,memory saving,interface flexibility,portability,distributed efficiency and so on.In 2016,it was selected as the Amazon AWS official deep learning platform.However,the survey found that most enterprise and users currently use the MXNet for stand-alone version when processing deep learning tasks by using MXNet.This has great limitation on the current data volume,data throughput,and calculation iteration speed.As a new type of business model,cloud computing is widely welcomed by users because of its massive computing resources and pay-per-use service model.Therefore,users can rent cloud resources to handle local deep learning tasks on demand.It can avoid the limitations when dealing with deep learning tasks on a single machine.In real life,users often pay the most attention when renting cloud resources: handling local tasks successfully with the least amount of rental expenses.However,most existing research about task scheduling in the cloud computing environment is done in order to reduce the time span of execution.Only a few studies have taken into account both the time span and the cost of cloud resource leases,and the task scheduling algorithm used has a defect in performance.Therefore,when using MXNet to process deep learning tasks in the cloud,it is of great practical significance to implement a high-performance deep learning task scheduling algorithm that aims to shorten the time span for completing deep learning tasks and reduce the cost of cloud resource leases.In view of the deficiencies of the existing research,this dissertation mainly studies the deep learning task scheduling technology based on the MXNet platform in the cloud computing environment.It mainly solves the problem of minimizing the cost of cloud resource leases for users who use MXNet to handle deep learning tasks.The main tasks are as follows:1)The ant colony algorithm has good performance in solving combinatorial optimization problems,but the traditional ant colony algorithm also has some defects.For the shortcomings of the traditional ant colony algorithm,this dissertation proposes a corresponding optimization program,and the improved ant colony algorithm(GeneticAnd Ant Colony Optimization Algorithm,GAACOA)performs performance evaluation on traveling salesman problems.Experiments show that the results of GAACOA algorithm are superior to traditional genetic algorithms and traditional ant colony algorithms in the number of algorithm iterations and the optimal path.2)In order to implement the goal of minimizing the cost of cloud resource leases for users,when processing deep learning tasks in the cloud,this disertation establishes a deep learning task scheduling model based on GAACOA algorithm in the cloud computing environment,and GAACOA algorithm sloves the task scheduling problem for deep learning based on the MXNet platform in a cloud computing environment was verified by experiments.By comparing with the experment results of related algorithms,GAACOA algorithm can optimize the time span for the deep learning task to complete the implementation,and effectively reduce the cost of cloud resource leases.