服务质量感知的云计算任务调度方法研究

发布时间：2018-05-31 14:49

本文选题：云计算 + 服务质量　；参考：《北京工业大学》2014年硕士论文

【摘要】：云计算是从分布式计算、并行计算和网格计算进一步发展而来的基于互联网的新的服务模式。许多知名的IT公司如Google，Microsoft， IBM， Amazon等都在进行云计算方面的研究，开发出的应用平台有的也已经投入商业运营。任务调度是云计算领域所要研究的核心问题，主要研究最优的任务分配策略，即如何将任务最优地分配到云系统管理的资源上执行，以使得任务得到均衡的分配或使得每个任务的执行代价降到最低或系统的总体性能达到最优。当前，任务调度的方法主要针对的目标包括处理时间、系统负载、经济原则等，在以上研究工作的基础上，针对云计算提供强大并行处理能力的本质和特征，本文分别从服务质量的可靠性、响应时间出发研究出新的任务调度方法，主要研究工作包括以下几个方面： 1、从目前具有一般特征的云计算系统框架出发，建立云计算任务调度的抽象系统模型，特别是抽象系统模型中各组件的数学模型。该系统模型将云计算系统抽象成用户、调度器、计算节点三个层面：各个用户彼此独立地产生任务，服从泊松分布；调度器从用户接受任务，把任务分解成任务分片后发送到相应计算节点上执行；计算节点执行任务分片，并将每个计算节点看作一个M/G/1排队系统。 2、在抽象系统模型的基础上，以博弈论为工具，建立以可靠性为目标的任务调度方法。可靠性，即计算节点的计算能力提供的稳定性。该方法以计算节点上稳定状态的提供能力为目标，以调度器上的任务分片策略为博弈策略，确定系统中各个调度器的任务分片方案。通过与均衡任务调度方法比较可以看出，在计算节点计算能力均衡或不均衡、不同系统规模、不同系统负载的情况下，该任务调度算法都能使系统具有较高的可靠性，同时该算法具有计算速度快的优点且能使各个调度器都能获得公平的机会。 3、在抽象系统模型的基础上，以调节熵函数法为工具，建立以响应时间为目标的任务调度方法。调度器将任务分解成任务分片以后发送到计算节点上并行执行，该方法将任务的各个任务分片执行最长的时间作为该任务的响应时间，每个调度器对任务进行分解调度时，，总是以任务的响应时间最小为目标建立数学模型。通过调节熵函数法，我们给出该模型的解并以此为基础设计出任务调度方法。实验表明，该算法优于均衡算法且该算法下的响应时间明显低于博弈算法；同时发现，比均衡算法目标函数值较优的博弈算法在并行时并不一定具备优势。
[Abstract]:Cloud computing is a new service model based on Internet which is developed from distributed computing, parallel computing and grid computing. Many well-known IT companies such as Google Microsoft, IBM, Amazon and so on are conducting cloud computing research, and some of the developed application platforms have been put into commercial operation. Task scheduling is the core problem in cloud computing field. It mainly studies the optimal task allocation strategy, that is, how to allocate tasks to the resources of cloud system management. In order to achieve a balanced assignment of tasks or minimize the execution cost of each task or optimize the overall performance of the system. At present, task scheduling methods mainly aim at processing time, system load, economic principles, etc. On the basis of the above research work, it aims at the essence and characteristics of cloud computing to provide powerful parallel processing capability. In this paper, a new task scheduling method is developed from the reliability and response time of QoS. The main research work includes the following aspects: 1. Based on the general cloud computing system framework, the abstract system model of cloud computing task scheduling is established, especially the mathematical model of each component in the abstract system model. The system model abstracts the cloud computing system into three layers: user, scheduler and compute node: each user generates tasks independently from each other, and the scheduler accepts tasks from users. The task is partitioned and sent to the corresponding computing node, which executes the task partition and treats each computing node as a M/G/1 queueing system. 2. Based on the abstract system model and game theory, a task scheduling method with reliability as its goal is established. Reliability is the stability provided by the computational power of the computing node. The aim of the method is to calculate the ability to provide stable state on the node, and the task partitioning strategy on the scheduler is used as the game strategy to determine the task partitioning scheme of each scheduler in the system. By comparing with the balanced task scheduling method, it can be seen that the task scheduling algorithm can make the system more reliable under the conditions of balanced or unbalanced computing power, different system size and different system load. At the same time, the algorithm has the advantages of fast computation and fair opportunity for all schedulers. 3. On the basis of abstract system model, a task scheduling method with response time as its target is established with the method of adjusting entropy function as a tool. The scheduler decomposes the task into pieces and sends them to the computing node for parallel execution. The method takes the longest time to execute each task as the response time of the task. When each scheduler decomposes the task, A mathematical model is always built with the aim of minimizing the response time of a task. By adjusting the entropy function method, we give the solution of the model and design a task scheduling method based on it. The experimental results show that the algorithm is superior to the equalization algorithm and the response time of the algorithm is obviously lower than that of the game algorithm. At the same time, it is found that the game algorithm which is better than the objective function of the equalization algorithm is not necessarily superior to the game algorithm in parallel.
【学位授予单位】：北京工业大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP393.09;TP301.6

【相似文献】