面向能耗优化的云平台调度策略

发布时间：2018-11-17 08:01

【摘要】：云计算由于其先进的理念、方便的使用方式，越来越受到各大厂商和用户的青睐。随着云计算的广泛使用，数据中心和集群的能耗问题越来越受到关注。服务器目前消耗的能源已经达到惊人的程度。大部分服务器在大部分时间内都处于空闲状态，这是对能源极大的浪费。但是如果降低能耗，服务器性能也会降低，这是用户所不愿意看到的。通常来说，云服务提供商和客户之间会签订服务等级协议定义SLA（Service Level Agreement），规定了服务提供商提供服务的最低条件。为了降低数据中心的能源消耗和满足服务等级协定（SLA）的要求，通常会在空闲时间关闭无用服务器。但是，服务器从关闭状态回到工作状态需要一定的时间。而在这一段时间内，服务器的响应时间可能会违反SLA的要求。在这篇文章中，我们首先总结前人的能耗模型并进行相应的实验，提出了单机上的与CPU频率和CPU利用率都相关的能耗模型。随后我们利用排队模型研究了集群上的服务器频率、负载到达率和能耗之间的关系。发现服务器集群上也存在某一个负载下的最优频率，使得整个集群的能耗最少，同时SLA得到满足。随着负载的变化，动态的开关机或调整服务器的频率，都是降低整个集群能耗的手段。如何在认识到存在最优频率的基础上有效结合这两种手段，达到能耗最优和满足SLA的要求，是前人没有的工作，本文继续围绕这一问题展开研究。在文献[1]的基础上，我们提出了Enhanced AutoScale （EAS）技术，该技术将最优频率作为一个重要参数加入到策略AutoScale中，也就是将开关机和CPU调频调压（DVFS）结合在一起。在EAS中，，调节频率被用来弥补AutoScale中关机造成的性能损失。该论文提出了两种EAS的实现方式，集中化的EAS（CEAS）和分布式的EAS （DEAS）。并给出了相应的算法和具体实现。实验结果表明这种方法能够有效的降低响应时间，同时只增加很少的能源消耗。我们的方法每瓦特性能（P P W）值在某些负载下甚至比AutoScale高了50%。
[Abstract]:Cloud computing is more and more popular among manufacturers and users because of its advanced concept and convenient way of use. With the wide use of cloud computing, the energy consumption of data centers and clusters has been paid more and more attention. Servers are currently consuming a staggering amount of energy. Most servers are idle most of the time, which is a great waste of energy. But if you reduce energy consumption, server performance will also be reduced, which users do not want to see. Typically, a service level agreement between a cloud service provider and a customer defines the SLA (Service Level Agreement), as a minimum condition for service delivery by a service provider. In order to reduce the energy consumption of the data center and meet the requirements of the Service level Agreement (SLA), useless servers are usually shut down during idle time. However, it takes some time for the server to return to work from shutdown. During this time, the server's response time may violate SLA's requirements. In this paper, we first summarize the previous energy consumption models and carry out corresponding experiments, and put forward the energy consumption model which is related to both CPU frequency and CPU utilization on a single machine. Then we use queuing model to study the relationship among server frequency, load arrival rate and energy consumption. It is found that there is an optimal frequency under a certain load on the server cluster, so that the energy consumption of the whole cluster is the least, and the SLA is satisfied. With the change of load, dynamic switching machine or adjusting the frequency of the server are the means to reduce the energy consumption of the whole cluster. How to effectively combine these two methods to achieve optimal energy consumption and meet the requirements of SLA on the basis of recognizing the existence of optimal frequency is a work that has not been done before. On the basis of reference [1], we propose the Enhanced AutoScale (EAS) technique, which adds the optimal frequency as an important parameter to the strategic AutoScale, that is, combining the switch machine with the CPU frequency modulation and voltage modulation (DVFS). In EAS, frequency regulation is used to compensate for performance losses caused by shutdown in AutoScale. This paper proposes two ways to implement EAS, centralized EAS (CEAS) and distributed EAS (DEAS). The corresponding algorithm and implementation are also given. Experimental results show that this method can effectively reduce the response time and increase only a small amount of energy consumption. Our method has a (P P W) value per watt that is even 50 per watt higher than AutoScale under some loads.
【学位授予单位】：上海交通大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP368.5

【参考文献】