基于Kriging的云任务调度及工程优化云平台建设

发布时间：2018-03-02 09:04

本文关键词： 工程优化 Kriging 期望提高信息熵云计算任务调度　出处：《大连理工大学》2015年博士论文　论文类型：学位论文

【摘要】：利用数据中心的高性能软硬件资源,云计算能够以“所付即所用”的方式对用户提供高质量和高可靠的服务,这种“以服务的形式提供计算资源”的新型模式已经在很多领域得到了广泛运用。工程优化问题具有复杂的应用背景,建设针对工程优化应用的云计算平台缺乏现有的成熟方法。首先,云平台需要使用高性能的工程优化方法完成云应用的开发和部署以节约用户成本；其次,常规的云任务调度方法通常具有普适性,并不适应工程优化云任务的调度需求；另外,工程优化云应用响应时间和计算成本的密集特征也对用户使用体验和平台资源分配提出了新要求。针对上述问题,本文展开了如下工作：1.提出了基于信息熵的期望提高(EEI)加点准则及其Kriging并行优化方法。使用信息熵原理与加权形式的期望提高准则结合,可在每次优化迭代时求得最优加权系数,使用该系数计算得到的样本点可同时满足优化点的最大期望特征和加权系数的最优特征；同时使用并行计算技术,按照投入运算的并行进程数分割样本组和多点加点的加权系数,可在较高的粒度上拆分整个优化过程。使用EEI准则的Kriging并行优化方法可以在保证优化计算精度提高的同时,获得理想的并行加速比,具有较高的计算性能。2.提出了基于Kriging代理模型的动态云任务调度方法。从工程优化云应用的微观计算特征出发,针对其具有的计算平稳期和过渡期特征,提出了使用Kriging代理模型对平稳期进行计算资源优化的策略。以每个计算平稳期的资源分配组合为设计变量值,以任务响应时间和计算成本的多目标最小化为目标函数值,优化计算后得到平稳期的最优资源分配方案。云任务的动态调度不仅有利于工程优化云任务的快速计算和成本降低,也同时提高了云平台计算资源的利用效率。3.提出了基于Kriging的云任务预测和分配方法并建设了工程优化云平台。针对工程优化云应用的宏观计算特征,构建了以应用部分和计算资源部分组成设计变量,以云任务的响应时间和计算成本为目标函数的优化模型。利用Kriging建立的设计变量和响应时间函数关系,给出用户的新任务响应时间预测值,任务成本给定条件下的计算资源分配为带约束的响应时间最小化问题。响应时间的预测可以避免用户不必要的等待过程,成本给定前提下计算资源的分配可以避免用户使用平台计算资源基于经验值的盲目性,也有助于用户依据其计算成本总量合理安排各个计算任务。使用虚拟化技术创建了平台组件,提出了核心功能的实现方式,最终建设了针对工程优化应用的云计算平台。
[Abstract]:Using the high performance hardware and software resources of the data center, cloud computing is able to deliver high-quality and reliable services to users in a "pay-as-you-go" manner, This new model of "providing computing resources in the form of services" has been widely used in many fields. Building cloud computing platform for engineering optimization application lacks existing mature methods. Firstly, cloud platform needs to use high-performance engineering optimization method to complete cloud application development and deployment to save user cost. Conventional cloud task scheduling methods are generally universal and do not meet the requirements of engineering optimization for cloud task scheduling; in addition, The dense features of response time and computing cost for engineering optimization cloud applications also put forward new requirements for user experience and platform resource allocation. In this paper, the following work is carried out: 1. The addition point criterion based on information entropy and its Kriging parallel optimization method are proposed. The information entropy principle is combined with the expectation enhancement criterion in weighted form. The optimal weighting coefficient can be obtained at each optimization iteration. The sample points obtained by using this coefficient can satisfy both the maximum expected characteristic of the optimization point and the optimal characteristic of the weighting coefficient, and the parallel computing technique is used. According to the number of parallel processes in the input operation, the whole optimization process can be split at a higher granularity by dividing the sample group and the weighting coefficients of the multi-point addition points. The Kriging parallel optimization method using the EEI criterion can ensure the improvement of the accuracy of the optimization calculation at the same time. A dynamic cloud task scheduling method based on Kriging agent model is proposed. Based on the microscopic computing characteristics of engineering optimization cloud application, the dynamic cloud task scheduling method is proposed. According to the characteristics of computing stationary period and transition period, a strategy of using Kriging agent model to optimize computing resources in stationary period is proposed. The resource allocation combination of each computing stationary period is taken as the design variable value. Taking the multi-objective minimization of task response time and computing cost as objective function value, the optimal resource allocation scheme for stationary period is obtained after optimization. The dynamic scheduling of cloud tasks is not only conducive to the rapid calculation and cost reduction of engineering optimization cloud tasks. At the same time, the utilization efficiency of cloud platform computing resources is improved. 3. A cloud task prediction and allocation method based on Kriging is proposed and an engineering optimization cloud platform is constructed. In this paper, an optimization model is constructed, which consists of application part and computing resource part, and takes the response time and computing cost of cloud task as objective function. The relationship between design variable and response time function is established by using Kriging. The prediction value of the user's new task response time is given. The computing resource allocation under the given task cost is a constrained response time minimization problem. The prediction of the response time can avoid the unnecessary waiting process of the user. The allocation of computing resources under a given cost can avoid the blindness of using platform computing resources based on empirical values. It also helps users reasonably arrange each computing task according to their total computing cost. Using virtualization technology, the platform components are created, and the core functions are implemented. Finally, the cloud computing platform for engineering optimization applications is built.
【学位授予单位】：大连理工大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TP393.09

【参考文献】