云服务系统中实时任务调度与资源动态调配方法研究

发布时间：2019-02-13 21:47

【摘要】：面对信息化战场中海量的战场信息和高动态并发的作战单元应用需求,云计算将为战场信息服务模式提供了一条新的途径。云计算作为分布式计算的最新发展趋势,它借助先进的虚拟化技术,将云计算数据中心大规模的计算、存储、网络等资源虚拟成巨大的资源池,为用户提供按需服务。在云计算模式下,用户只需将任务提交到云服务系统,云服务系统将自动分析任务特性、预测任务的资源需求,再根据云服务系统中底层资源的使用情况,将任务调度到相应的资源上执行,并在用户指定的时间内完成任务的执行。即,用户只需提交任务、服务质量要求和接收任务执行的结果,而中间的所有事情,云服务系统将自动完成。对于云服务系统而言,高效的任务调度和资源动态调配方法是提高其性能的关键技术之一。目前,已经存在大量关于云服务系统中任务调度和资源动态调配的研究成果。但是,已有的研究大部分集中于理想的调度环境:1)被调度的任务集合预先知道;2)任务的执行时间是确定值,并且在调度前可以获取;3)资源即时可用。然而,在实际的云服务系统中,存在大量动态、随机性因素。比如,任务到达率剧烈变化,任务执行时间具有随机性,资源可动态伸缩和启动资源需要时间开销等。云服务系统中这些动态和随机因素,往往使得预先生成的调度方案失去原有的优势或无法顺利实施,甚至使得初始调度方案不再可行。因此,云服务系统中实时任务调度和资源动态调配方法研究极具理论和现实价值,且富有挑战性。在实时任务调度和资源动态调配过程中,本文主要针对以下三种典型的情况:任务动态到达、任务执行时间是随机变量、主机和虚拟机启动时间不可忽略。本文的主要工作和创新点包括以下四点:(1)提出了一个可扩展的主机组织模式。针对云服务系统中大规模主机对传统主机组织模式,比如,集中式、分层式和分布式,提出的挑战,提出协同式组织模式,将大规模主机分为多个集群,每个集群都有一个独立的调度器,每个调度器负责本集群的任务调度和资源调配,同时调度器之间相互协调,共同调度任务和底层资源,从而提高云服务系统的可扩展性。(2)提出一个随机性感知的调度框架。针对云服务系统中实时任务的高动态、随机性和高时效性要求的特征,为每个集群提出一个随机性感知的调度框架,将大部分等待任务放置在全局等待队列中,并控制虚拟机上等待任务的个数,当虚拟机完成任务之后,等待任务就立即执行,然后优先调度全局队列中时效性要求较高的任务到虚拟机上等待,避免已经完成任务的随机性累加到当前调度的任务,从而提高调度的方案的稳定性和保障实时任务时效性的能力。(3)提出了一个随机性感知的调度算法PRS。针对云服务系统中实时任务动态到达、执行时间具有随机性的问题,在随机性感知调度框架的基础上,巧妙集成前摄性和反应式调度思想,提出一个在线调度算法PRS,该调度算法根据云服务系统的实际运行情况,不断为云服务系统生成新的任务和虚拟机调度方案,从而在保证实时任务时效性要求的条件下,提高云服务系统中主机资源的有效利用和降低能量消耗。(4)提出了一个机器启动时间感知的任务调度与资源动态调配算法STARS。在云服务系统中,实时任务的到达具有随机性和突发性,当云服务系统中的负载突增时,启动主机和创建虚拟机的过程会造成一定的时间开销,使得某些任务不能及时开始,从而延误了它们的截止期。针对以上问题,本文提出机器启动时间感知的任务调度与资源动态调配算法STARS,借助单个虚拟机CPU能力可以动态伸缩的能力,通过转移机器启动时间对截止期较短任务的影响,以减缓机器启动时间对突增任务时效性的影响,以提高云服务系统保障实时任务时效性的能力。
[Abstract]:The cloud computing will provide a new way for battlefield information service mode in the face of the massive battlefield information in the information field and the application demand of the high-dynamic and concurrent operation unit. As the latest development trend of distributed computing, cloud computing, with the help of advanced virtualization technology, virtual computing, storage, network and other resources of the cloud computing data center into a huge resource pool, providing the user with the on-demand service. In the cloud computing mode, the user only needs to submit the task to the cloud service system, and the execution of the task is completed within the time specified by the user. That is, the user only needs to submit the task, the quality of service requirements, and the result of the execution of the receiving task, and all the things in the middle, the cloud service system will be automatically completed. For the cloud service system, the efficient task scheduling and resource dynamic allocation method is one of the key technologies to improve its performance. At present, there are a lot of research results on the task scheduling and the dynamic allocation of resources in the cloud service system. most of the existing studies, however, focus on the ideal scheduling environment: 1) the scheduled task set is known in advance; 2) the execution time of the task is a determination value and can be acquired prior to scheduling; and 3) the resources are available immediately. However, in the actual cloud service system, there are a lot of dynamic and random factors. For example, the task arrival rate changes dramatically, the task execution time is random, the resources can be dynamically expanded and the resources need time overhead, and the like. These dynamic and random factors in the cloud service system often make the pre-generated scheduling scheme lose the original advantage or can not be implemented smoothly, and even the initial scheduling scheme is no longer feasible. Therefore, the research of real-time task scheduling and resource dynamic allocation in the cloud service system is of great theoretical and practical value and is challenging. In the process of real-time task scheduling and resource dynamic allocation, this paper mainly focuses on three typical situations: the dynamic arrival of the task, the execution time of the task is a random variable, and the starting time of the host and the virtual machine is not negligible. The main work and innovation points of this paper include the following four points: (1) An extensible host organization model is proposed. in that light of the challenge of the large-scale host in the cloud service system to the traditional host organization mode, such as centralized, layered and distributed, propose cooperative organization mode, the large-scale host machine is divided into a plurality of clusters, each cluster has an independent scheduler, each scheduler is responsible for the task scheduling and resource allocation of the cluster, and meanwhile, the schedulers are in coordination with each other, and the tasks and the bottom layer resources are jointly dispatched, so that the expandability of the cloud service system is improved. (2) a random-sensing scheduling framework is proposed. aiming at the characteristics of high dynamic, random and high timeliness requirements of real-time tasks in a cloud service system, a random-sensing scheduling framework is proposed for each cluster, a large part of the waiting tasks are placed in a global waiting queue, and the number of waiting tasks on the virtual machine is controlled, after the virtual machine completes the task, the waiting task is executed immediately, and then the task of higher timeliness requirement in the global queue is preferentially dispatched to the virtual machine to wait, so that the randomness of the completed task is prevented from being accumulated to the currently scheduled task, so as to improve the stability of the scheduling scheme and the capability of ensuring the timeliness of the real-time task. (3) A stochastic perceptive scheduling algorithm (PRS) is proposed. Aiming at the problem of the dynamic arrival of real-time tasks in the cloud service system and the random problem of the execution time, on the basis of the random-aware scheduling framework, the proactive and reactive scheduling ideas are skillfully integrated, an on-line scheduling algorithm PRS is proposed, According to the actual operation condition of the cloud service system, the scheduling algorithm continuously generates a new task and a virtual machine scheduling scheme for the cloud service system so as to improve the effective utilization and the energy consumption of the host resources in the cloud service system under the condition of ensuring the timeliness requirement of the real-time task. (4) The task scheduling and resource dynamic allocation algorithm STARS is proposed. In the cloud service system, the arrival of real-time tasks is random and bursty, and when the load in the cloud service system suddenly increases, the process of starting the host and creating the virtual machine can lead to a certain amount of time overhead, so that some tasks can not start in time, thus delaying their cut-off period. In view of the above problems, this paper puts forward the task scheduling and resource dynamic allocation algorithm STARS of the machine start-up time perception, and the ability of dynamic expansion can be realized by the ability of the single virtual machine CPU, and the influence of the start time of the transfer machine on the short task of the cut-off period is achieved. so as to reduce the effect of the starting time of the machine on the timeliness of the sudden increase task so as to improve the capability of the cloud service system to guarantee the timeliness of the real-time task.
【学位授予单位】：国防科学技术大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP393.09

【参考文献】