【作者】 刘阳；

【导师】 林伟伟； 邓健爽；

【作者基本信息】 华南理工大学 ， 工程硕士（计算机技术）（专业学位）， 2020， 硕士

【摘要】 近年来云计算规模不断扩大,伴随着容器技术的迅速成熟,采用容器即服务(Container as a Service,CaaS)的云数据中心市场占用率愈来愈高,由此带来的能耗问题日趋明显。在云数据中心内部,容器负载和主机负载是影响其能耗变化的关键因素,精准的把握负载变化趋势,对于资源的合理调度有促进作用,使得有效降低云数据中心能耗成为可能。但是云数据中心的负载种类繁多,需要探索更为通用和广泛的模型,来对多种负载进行预测。而且在基于CaaS的云数据中心中,调度过程需要同时考虑容器和虚拟机两个因素,才能使得云数据中心的调度策略,同时考虑到节能和服务质量。在此背景下,本文所做的主要工作有:(1)本文在调研了时间序列常用的分析方法和预测模型后,结合云环境下负载数据的特性来考虑,提出了一种基于CNN-LSTM模型和EEMD分解的预测方法(简称EEMD-CNN-LSTM)。该方法首先将云环境下的负载时序数据通过集成经验模态分解为若干个本征模态函数(IMF),选取前三个IMF合并为新的分量,利用带注意力机制的CNN-STM模型进行预测,对于其他分量采用LSTM算法进行预测。CNN-LSTM在加入了注意力机制之后,较好的改善了对序列较长和波动较大的时序数据预测效果。通过实验结果表明,EEMD-CNN-LSTM方法能够预测多种类型的云环境下负载数据,并且预测精度良好。(2)为了有效降低基于CaaS架构的云数据中心能耗,本文提出了基于负载相似性和改进遗传算法的容器整合策略(简称CSLIGA)。该策略首先根据EEMD-CNN-LSTM方法的预测结果,判断物理主机是否过载;然后在过载物理主机上,根据容器与主机的负载相似性大小,选择容器进行迁移,在此过程中,对于容器副本,在进行负载预测后判断是否扩缩容,有针对性的增加或者减少。接着对于迁移出的容器,通过改进遗传算法的容器放置策略,将容器放置到合适的目标主机上;最后策略会通过欠载主机处理算法,将判定为欠载的主机进行休眠,实验在Cloud Sim平台上进行,结果表明,CSLIGA策略能够有效降低云数据中心的能耗(少消耗4%的能量),而且保证服务质量(SLA违反率小于5%)。更多还原

【Abstract】 The scale of cloud computing has been continuously expanding in recent years,with the rapid development of container technology,the market share of cloud data centers that use Container as a Service(CaaS)has been expanding rapidly as well,which leads to a prominent problem of high energy consumption.In the inner part of the cloud data center,container loads and host loads are the key factors that affect its change of energy consumption.Having an accurate control of the load change’s trend contributes to the reasonable scheduling of resources,making it possible to reduce energy consumption of the cloud data center efficiently.However,there are many types of loads in cloud data centers,resulting in a need for exploring more general and extensive models to predict multiple loads.Moreover,in the cloud data center based on CaaS,both the container and the virtual machine need to be taken into consideration in the scheduling process so that scheduling strategy of the cloud data center can take energy saving and service quality into account.Based on this background,the main contribution of this article are made as follows:(1)Having investigated the commonly used analysis methods and prediction models of time series,and taken the characteristics of load data in the cloud environment into consideration,this paper proposes a prediction method based on CNN-LSTM model and EEMD decomposition(EEMD-CNN-LSTM).The method first decomposes the load data in the cloud environment into several Intrinsic Modal Functions(IMFs)through Ensemble Empirical Mode Decomposition,selects the first three IMFs and merges them into new components and then uses CNN-LSTM with attention mechanism to predict.For the other low-frequency components,LSTM is used to predict.With the attention mechanism added,CNN-LSTM has better improved the prediction effects of time series data with long sequences and large fluctuations.The experimental results show that the EEMD-CNN-LSTM method can predict various types of load data in the cloud environment with high prediction accuracy.(2)In order to effectively reduce the energy consumption of cloud data centers based on CaaS architecture,this paper proposes a Container consolidation Strategy based on Load similarity and Improved Genetic Algorithm(CSLIGA).The strategy first determines whether the host is overloaded according to the prediction result of the EEMD-CNN-LSTM method;then on the overloaded host,the container is selected for migration according to the load similarity between the container and the host.Meanwhile,for the container duplication,it judges whether to expand or shrink the capacity after load forecast so that the container duplication can increase or decrease to the point.Next,for the migrated containers,they are placed on the appropriate destination host by Container Placement of Improved Genetic Algorithm;finally,the strategy will make the host sleep which is determined to be underloaded by Low-Loading Host Processing Algorithm.The experiment was performed on the Cloud Sim platform with the experimental results showing that the CSLIGA strategy can effectively decrease the energy consumption of cloud data centers(consume 4% less energy than other),and ensure the quality of service(SLA violation rate is less than 5%).更多还原