Managing performance and energy in large scale data centers .

摘要

Data centers are the computing facilities to process, store, and access information. Specifically, data centers serve as the key infrastructure for Cloud computing service providers. However, service providers have observed the trend of the under-utilization of production servers, which unnecessarily increases the total cost of ownership. The demand on managing the total cost of ownership is driving researchers to study both performance and energy issues in data centers, which are addressed in this thesis. For considering performance and energy, this thesis consists three contributions -- design of a comprehensive multi-tier data center simulation platform; energy management of multi-tier data centers; performance model for predicting running times of applications in data centers.;Design and analysis of large and complex distributed systems like data centers often suffer from the lack of an available physical infrastructure due to the cost constraints especially in the academic community. With this motivation, this thesis proposes the design of a comprehensive, flexible, and scalable simulation platform for in-depth analysis of multi-tier data centers. The potentials of such a simulation platform is demonstrated by its ability to simulate and measure the performance and power consumption of data centers with high accuracy. The second contribution is towards increasing the energy efficiency of multi-tier data centers using a multifacet approach, namely Hybrid, consisting of dynamic provisioning, frequency scaling and dynamic power management (DPM) schemes to reduce the energy consumption of multi-tier data centers, while meeting the Service Level Agreements (SLAs). The energy management scheme consists of two heuristics that utilize the Mean Value Analysis by modeling data centers as closed queueing networks. This scheme manages the energy consumption at the global and local levels. The global level management determines the sufficient number of servers for a service. At the local level, the proposed scheme dynamically exploits energy management techniques in individual servers. The third contribution is estimating the performance of data centers. This involves the development of performance models of shared service platforms with multiple resources contention given the fact that a typical data center is shared by multiple services that contend for multiple system resources. The proposed model can estimate the total completion time of jobs, which is the objective function of the scheduling problem. This work illustrates that an existing job scheduler can be enhanced by modifying its original objective function to the proposed model. To summarize, this thesis discusses the performance and energy implications in data centers, along with suggesting optimization techniques for improving performance and energy conservation in typical data centers.