Applications that need to be processed on resources of shared server recently show various resource request forms (CPU speed oriented, memory band oriented, storage size oriented), request processing time is several seconds several seconds, several days several days, multiple times Interest in building and operating a large-scale computing infrastructure is exploding increasing as it is displayed on an inter-scale. As a result, management of the latest data centers, which possess hundreds to tens of thousands of high-performance computing servers, is one of the core technologies of the modern IT industry. Data Center has a market size of more than US$ 15 billion on 2016 basis. However, an explosive increase in the energy consumption cost due to the improvement of the computing capacity of the data center has emerged as a serious problem. The modern data center is the same as 25,000 households energy consumption of the level required. Based on the general o詮긟e environment, data centers are reported using the same area of up to 100-200 times more power. The energy consumption cost required for the data center is doubled every 5 years. In addition, the enormous energy consumption cost accompanying the increase in the amount of power usage is a major obstacle to the expansion of the scale of the data center and the improvement of the applied service quality.
In this dissertation, we classify the method for reducing the energy consumption cost of the data center into four major categories. Initially it is a method to adaptively control the amount of electric power to be supplied to the data center by forcibly setting a logical budget (power budget). In this method, Allocate center computing resources logically adjust the size of the computing server by mon-itoring and analyzing the usage patterns and power consumption of resources of the applications being received and processed. This method guarantees ﬁne granularity, but for large data centers It is diﬃ-cult to exhibit a fast response speed. Secondly, the way to utilize feedback-based control techniques to achieve real-time power usage adjustment has received much attention. However, this method is based on multi-tenant based writing on virtualized resources. For business management, it is diﬃcult to apply it all at once. The third method is to actively utilize renewable energy, generated from nature such as the sun, wind power, which is inexpensive to produce, to supply electricity from the data center. Renewable energy. . . etc, however, Supply quantity changes based on the geographical position of the center (on-site basis), and the supply pattern has a form of a graph oscillating along the ﬂow of time. These supply instability and intermittency not only make it diﬃcult to achieve reliable application service quality as-surance. The lifetime of the data center hardware infrastructure can also be adversely aﬀected. The ﬁnal approach is to use a polymorphic approach to diﬀerentiate the quality of service by providing diﬀerent modes of operation based on the priority of the application assigned to the data center. This method has the advantage of eﬃciently reducing the energy consumption of the data center and being able to respond to the variable power budget and the application workload in real time. In this dissertation, we will focus on the last method and ﬁnally restrictive power budget provided a management framework that can achieve eﬃcient energy consumption reduction of data centers owned.
In this dissertation, we propose power consumption control technology to ensure sophisticated real-time performance of large-scale cloud data center and explain the modeling method and algorithm accompanying this. The purpose of the method proposed in this dissertation is to give to the cloud data center the maximize utilization of the computing server while observing the constraints of power budget. In order to secure the service level agreement (SLA) required by the user of the cloud service, the proposed method uses a power budget, it can be eﬃciently distributed and extended to each server. In this dissertation, the main items that contributed technically are as follows.
• Hybrid cloud data center power management system
– Cloud data center power Balancing Manager, Power Distributor Manager (PDM): PDM mon-itors the size of the power budget, the performance of applications that are operated in the computing server that can be used in the dynamically changing data center Most It aims to derive the variance of enemy’s ﬁghting resources. For this reason, the PDM periodically sees resource allocation of virtual machine instances (virtual machine instances) placed on the cloud server’s gen hypervisor (Xen-hypervisor)
– Cloud application performance - Cloud server power consumption converter, Resource To Power (RTP) converter: RTP converter converts the performance level of cloud service ap-plication of diﬀerent metric and the power consumption of allocated server It plays a role of mapping. Through the mapping function deﬁned in the RTP converter, the cloud data cen-ter administrator can elaborate estimate the power supply amount compared to the required application service quality.
• In case Dynamic Allocation of Power Transmission Path in Integrated Data Center Based on Adaptive Uninterruptable Power Supply (UPS) Switching.
– Switching Techniques UPS providing recognizes the current state of idle servers (idle server) and the activation server for each server racks allocated (server rack) (active server), dynam-ically distributed UPS connection Reconstruct. For this reason, the technique derives the optimal connection or release policy of the circuit of the DC / AC converter located inside each UPS distributed via the central controller Automatic Transfer Switch (ATS).
• Cloud server power management
– Provide a server power consumption model based on the performance of heterogeneous com-puting servers in the cloud data center and the size of the computing resources allocated to each virtual machine instance. For this reason, we deﬁne the CPU as an important component of the power consumption of the server Based on the high CPU frequency level (frequency level), elaborate power consumption.
– Provide a local power controller, Local Power Controller (LPC) in conjunction with PDM and UPS which is the central controller. This controller proposes Gain scheduling based on the technology of the high level feedback control theory and assigns to each server Thereby making it possible to set an appropriate physical resource corresponding to the speciﬁed logically usable power resource.
– Provide vCPU Tuner which can adjust the performance component of virtual CPU. For each instance of each virtual machine, determine the virtual CPU core frequency level and the CPU credit level at the same time to obtain the target power consumption To be derived. For this reason, we provide Multi Input Single Output (MISO) based power modeling with frequency and credit variables as input values and output size of power consumption.
The scope of this dissertation is to improve the power consumption eﬃciency of the data center under the constraint of the power budget. Aim for energy center eﬃcient data center in cloud environment. The publication position is based on the power management level of the data center, that is, the size of the data center and the individual server I put emphasis. For example, we used a hybrid data center power management system with two levels of control system to handle data center level power eﬃciency. Here, the PDM operates at the microscopic level of operation of the data center, and UPS integration contact Gun Bop manages the data center active and shutdown UPS module via the dynamic power allocation path of the integrated data center based on adaptive UPS switching. On the other hand, we use two main components, power and control at the server level Model possibilities. There are MISOSPM and LPCFreqSchd for each power modeling and power control. We use Resource To Power (RTP) converters, which are used to convert the virtual machine’s resource requirements to the typical amount of power shown in the power demand Is deﬁned. The biggest advantage of this RTP converter is that the computing requirements of the virtual machine are recognized as the amount of power used to distribute power between servers in the data center.