Hybrid Resources Demanded Task Scheduling Method With Multi-Dimensional Configuration Requirements

Independent task scheduling with multi-dimensional configuration requirements in cloud environment is widely used in cluster management platforms such as Open Stack and Kubernetes.Multi-dimensional configuration requirements,execution environment requirements and hard deadline constraints of independent tasks from multi tenants are considered in this thesis.The problem of independent task scheduling in a single data center with the objectives of minimizing energy consumption and improving Qo S is studied.The main challenges of considered problem lie in:(i)Smaller granularity is set for containers than virtual machines while allocating CPU resources,which means that containers can use CPU time more flexibly.Delayed scheduling tasks running on containers would make workloads more balanced for servers while violation of the deadline constraints would be led.(ii)Since tasks are created with parallelism degree,allocating more CPU cores within the range of parallelism can speed up the overall execution time of the task with more energy consumption.According to the difference of the service capacity of data centers,the problem is divided into two subproblems:(1)Minimizing energy consumption with sufficient capacity for the data center,(2)Minimizing energy consumption while ensuring service quality under the condition of insufficient capacity.After analyzing the characteristics of the two sub-problems,the corresponding mathematical models are established respectively.For the first subproblem,an energy-aware scheduling algorithm for tasks with multidimensional configuration requirements(ESMCR in short)was presented in this thesis.The algorithm includes four components: tasks sequencing,host machine selecting,available time blocks searching and time block matching.Four strategies are proposed to generate the task sequence: maximum parallelism level first,minimum floating time first,maximum resource requirements first and random task sequencing based on deadline.In order to optimize the processing time of the algorithm,the host server is determined before resource allocation to avoid global search in the following steps.Three strategies are proposed to select host servers,i.e.minimum average utilization first,maximum computing capacity first and minimum average power first.After the host machine is selected,all available time blocks are searched by traversal.Finally,the appropriate time block is selected for a task.Four rules are presented for this component,mapping the task to the time block with the earliest start time,the latest start time,the lowest average utilization or the best fit memory.For the second subproblem,the performance index based on information entropy is modeled to measure the Qo S,and a decomposition-based multi-objective evolutionary algorithm combined with ESMCR(MOEA/D-ES in short)is proposed to for the problem.The algorithm mainly includes three core phases: initial population construction,sequence evaluation with optimization and task sequence adjustment.Firstly,the task scheduling sequence generation strategy proposed in the ESMCR algorithm is used to construct a diverse population,and the task sequence is evaluated according to the resource allocation strategy of the ESMCR algorithm.Meanwhile,the task sequence would be improved on the basis of the LDD(Later Due Date)rule.In the task sequence adjustment stage,differential evolution crossover and polynomial mutation operators are used to generate new scheduling sequences.To evaluate the performance of the proposed algorithm,the multi-factor analysis of variance(ANOVA)technique is used to calibrate the components or parameters of the two algorithms,and the optimal parameter combination of the algorithm was selected through a large number of experiments.By the component calibration of the ESMCR algorithm,it is found that in the task scheduling sequence generation phase,the proposed heuristic rules have no statically significant differences in the results,while different heuristic strategies have statically significant differences in the host server selection and the time block matching phase.As for the MOEA/D-ES algorithm,the relative percentage deviation of Inverted Generation Distance(IGD)was adopted to evaluate the Pareto solution sets,it is found that there is no significant difference in the influence of different parameter values on the evaluation indexes.At last,two algorithms for energy consumption on virtual machine placement(VMP)problems are modified according to the considered scenarios as the compared algorithms of the two presented algorithms.Through a large number of experimental verification,the two algorithms presented in this thesis outperforms the compared algorithms with different tenant numbers and different ratios tasks.