Research On Key Technologies For Energy Efficient Virtual Machines Load Balancing In Cloud Data Center

With the emergence of virtualization technology of cloud computing, mature commercial cloud data center thoroughly changed the whole IT by providing on-demand configuration of elastic computing resources. However, HPC is often one of the most important evaluation metrics in the data center, lack of awareness of the energy consumption and carbon emissions. In this context, the scale of the cloud data center is getting bigger than before, the consumption of electricity is very breathtaking, not only bring the high operation cost, but also produce the high carbon dioxide emissions.The main reason of the huge power consumption in data center is the low-efficient utilization of computing resources. Large amounts of data show that most of the data center server equipment is in under-load condition, i.e.10%-50%, while the equipment consumes 70% of the full load energy consumption. Not only that, but also brings additional energy consumption of cooling system. In the decade of advocating green, saving energy, reducing carbon emissions, so as to realize the sustainable development, the cloud data center should not only focus on QoS, but also pay more attention to the energy consumption impacted by imporoving the resources utilization, realize the sustainable development of green energy-saving efficient target. However, unless the development and application of advanced energy conservation and resource management solutions have been taken, otherwise these data will grow fast.This dissertation focus on the resources management problem with the green energy efficient technologies of cloud data center. By establishing the energy consumption model to analyze new features of the host load data of cloud environment, we proposed a series of host workload prediction and virtual machine placement algorithms. Using host workload data of a cloud environment from real world in the simulation experiment, we find that the proposed multiple combination algorithms have improved obvious performance compared with the benchmark algorithms.Specifically, the main work and innovation points are as follows:1. In view of the high energy consumption, we proposed a taxonomy of the energy efficient technology of cloud computing system. We launched in-depth analysis of the multiple levels include the static and dynamic power management, hardware and software levels, operating system level, virtual machine and data center levels. In the process of research, we focus on the system resource, the optimization goal, energy-saving technology, load characteristic, and many other aspects. This paper analyses three aspects of key technology which are the operating system level, the virtual machine and energy-efficient cloud data center level. The taxonomy has a strong guiding significance further to green energy conservation as the goal of dynamic virtual machine equalization algorithm research.2. We proposed a set of distributed dynamic virtual machine balancing heuristic algorithms in cloud computing environment. Specifically, we leverage the median absolute deviation and Interquartile Range to improve the traditional static threshold policy, and use local regression and robust local regression algorithms to host overload detection. We use the minimum migration time, random and maximum correlation policies for virtual machine migration. We define the virtual machine placement problem in cloud computing environment as a bin-packing problem, design Best Fit Decreasing Power Aware algorithm(BFD-PA) for VM placement. Using Planetlab cloud data center workload data simulation experiments, the results show that the proposed combination of LR algorithm and the MMT policy significantly outperforms the other dynamic VM consolidation algorithms, and can greatly reduce the SLA violation cases and reduce the number of virtual machine migrations.3. We proposed a combination of phase space reconstruction method (PSR) and group method data handling based on evolution algorithm(EA-GMDH). As we know, it is the first time to use it in the host load prediction field of the cloud computing environment. Using the PSR method, one dimensional time series of workload was reconstructed into the time sequence of multidimensional space. Then the final prediction result is obtained by training and learning in EA-GMDH neural network. Using the real Google cloud data center workload data simulation experiments, the results show that the EA-GMDH algorithm proposed in this paper, outperforms the benchmark algorithms:Bayes, EMA, LWMA, LMA, AR, ANN and PP, etc.4. We proposed a novel algorithm to predict host load in cloud environment. The method uses the autoencoder coding network and softmax classifier, converting the regression problem into a classification problem, once again raises the accuracy of load prediction. Through unsupervised learning method, we successfully extract the features of historical load data in the window. Based on the extracted features, we leverage the classifier, successfully obtained the prediction results of load data. Also, we use the Google cloud data center the workload of data as a data source. Compared with some benchmark algorithms such as ANN, Bayes and EA-GMDH algorithm, the new algorithm we proposed can achieve higher performance with more accuracy.