Research Of Adaptive And Green Control Thechnologies In Data Center And Its Applications

Data center, as the core of big data and cloud computing, plays an important role in nearly every sector of the society. They provide a wide variety of services from Internet services to high-performance computing and huge storage. As an increasing explosion of development in Internet and information society, the scale of modern data centers are increasing significantly. There are multi-faceted challenges to the effectively resource management of a modern data center, such as the high dynamical of Internet workload,the continuous expansion of complexity of Web applications, the resource contention caused by virtualization technology. Furthermore, it poses an important issue of building green data center, since the high cost of energy and environment problems. According to the above reasons, this dissertation firstly analyzes and summarizes the related work in the resource management and power saving of data center. Then does depth researches in adaptive and green control approaches for data center in terms of resource management and power saving. Finally, depend on the cooperation project with the institute of optics and electronics, the Chinese academy of sciences, and the simulation on the adaptive optics(AO), we apply the research results of adaptive and green control technologies to the research of delay error management and parallel computing of adaptive optics simulation system(AOSS). The main contributions and works of this dissertation include:1. For the data center, existing works consider the performance and power saving in two separate ways, and the traditional power saving approach cannot be directly applied to virtualized server environment. We propose a novel two-layer control architecture solution that guarantees the performance and power saving in a coordinate way, this solution adopts the virtualization technology and dynamical voltage and frequency scaling(DVFS) technology. In the first layer, it adopts a multi-input-multi-output(MIMO)control approach to maintain the workload balance among different VMs. Thus, all of the VMs have approximately the same response time. In the second layer, for meeting the power saving target, it dynamically manipulates the physical frequency that keeps the average response time of all VMs to a desired level. Experiment results show the effectiveness of our solution.2. Due to the resource contention between co-located VMs, and different with existing modeling methods that do not consider uncertainties of real-world web servers,in here the web server system is considered as the stochastic time-varying system and formulate as a coupled MIMO control problem. Meanwhile, a time-varying Autoregressive moving-average model with exogenous inputs(ARMAX) model is used for system modeling. The experiment results show that the effectiveness of AMRAX model and the performance of AMRAX model is better than the popular used AMRX model in terms of accuracy and stability.3. For the requirement of online control for the performance and power consumption of the virtualized server in the faces of dynamical workload case, we propose a autonomic power-aware resource control approach for virtualized Web server via robust control, it is used to deal with the problems of previous studies seriously relied on models that were trained offline for specific workload cases, and those approaches unsuitable for time-varying workloads. This approach adopts the adaptive Linear Quadratic Gaussian algorithm with stochastic method(ALQGw S) to calculate the optimal resource allocation solution. The obvious advantage of this approach is that the control decision is made based on minimizing an average cost function among a set of models, which are generated according to a Gaussian distribution. Experiments on our Xen-based testbed server demonstrate that our robust control approach guarantee the performance of each Web server, while minimizing the overall power consumption of physical serve, its performance outperforms existing solutions under dynamical workloads in terms of control accuracy and system stability.4. For the management requirement of virtualized servers cluster in terms of resource utilization, performance and power consumption, we propose a power-aware performance stochastic control of virtualized servers cluster that copes with the existing control solutions show the instability and inefficiency when facing the high dynamical and burst Internet workload. This approach adopts a constrained stochastic linear quadratic control method(c SLQC) that simultaneously guarantees power and meets performance specifications with flexible tradeoffs for virtualized servers cluster. Firstly, it considers the resource allocation between different VMs as a dynamic optimization problem and set the probabilistic constraints on response time according to the features of Web server, then it solves this problem by Semidefine Programming. Secondly, for coping with inaccuracy of estimated model due to the unexpected burst workload, it integrates a proportional controller into the control architecture. Finally, experiments on Xen-based testbed server with a variety of real-world Internet workload traces show that compared with existing solution, the proposed control solution has strong robustness in faces of the burst workload, and it has elasticity.5. For solving existing control solutions cannot be applied in the AOSS for real-time computing problem, based on proposed online control and load balancing control methods in the servers resources management, we propose an adaptive predictive controller to reduce the time delay error of AOSS, the simulation results show the effectiveness of proposed controller. Furthermore, in order to realize the parallel computing of different AOSS, by applying the power-aware resource control approach for virtualized Web server, we design a novel AOSS parallel computing platform based on the virtualized server. Experiments on the Xen-based server show that high efficiency of proposed control algorithm, it realizes the parallel computing of different AOSS while improves the server resource utilization significantly and achieves the power saving target.