Power-Efficient Autonomic Management and Control for Virtualized Data Centers
Data centers, the backbone of cloud computing and big data processing provide a wide variety of services, from Internet to high performance computing and storage. With the explosion of Internet applications and information processing, the scale of modern data centers is increasing significantly. Consequently, there is a growing interest in reducing cost by decreasing human involvement in the management of data centers and diminishing energy usage through autonomic computing and control.
Virtual machine technology, as an important enabler of cloud computing, has been widely employed in modern data centers. The last decade has seen an explosion of the development in the area of virtualization (e.g., Amazon EC2, Google AppEngine), as it offers improved servers utilization and reduced power consumption via consolidating multiple applications to several powerful servers. However, in recent years the server farm has been in a state of continuous expansion, both in scale and complexity. As a result it is increasingly difficult for system administrators to maintain servers manually. Hence there is a growing interest in decreasing human involvement in management of modern data centers through automatic computing. The benefits of this proposal accrue to the effective management of modern data centers by optimizing resource allocations and minimizing overall power consumption while guaranteeing the service level agreements of users.
This project aims to develop an autonomic computing and control framework with a self-managed power-efficient structure for virtualized environments. We will focus on improving the performance of resources and power control for virtualized web servers at different levels. A comprehensive and hierarchical management solution that exploits unique features of control systems will be established to improve the energy efficiency of servers with performance assurance. In particular, we propose to characterize web server complexity using nonlinear modeling. The main idea is to identify the key factors that affect the service quality of web applications. For co-located multitiered web applications, we propose to design robust control laws to address instability issues with existing resource and power control approaches. Resource bottleneck switch and trade-off between optimal control and computational complexity will be sought. To guarantee end-user response time in the face of dynamical and bursty workloads, a stochastic control framework to coordinate resource and power management of web servers is proposed. For the server cluster level, we propose to design coordinated distributed controllers for resource and power consumption under dynamical workloads. The proposed algorithms will be tested experimentally on in-house developed test-beds. Preliminary results obtained demonstrate the method’s viability.
The research plan is complemented by a strong educational plan, with particular emphasis on the cultivation of a diverse undergraduate and graduate population and inclusion of under-represented minorities. The teaching plan will include improvements in essential courses and creation of undergraduate capstone design projects. Proven effective teaching techniques will focus on both project-oriented integrated systems design and widespread dissemination of results through presentations and publications.