Graduation Year

2020

Document Type

Dissertation

Degree

Ph.D.

Degree Name

Doctor of Philosophy (Ph.D.)

Degree Granting Department

Computer Science and Engineering

Major Professor

Ken Christensen, Ph.D.

Committee Member

Miguel Labrador, Ph.D.

Committee Member

Yicheng Tu, Ph.D.

Committee Member

Nasir Ghani, Ph.D.

Committee Member

David Rabson, Ph.D.

Keywords

Metrics, Multi-Tenant Data Centers, Performance Evaluation, Power Overloading, Power Provisioning

Abstract

Data centers contribute to approximately 1% of the global electricity consumption, and billions of dollars are spent annually worldwide in construction of new data centers to meet the rising demand for cloud-based services. Given the high cost of construction, the power infrastructure in a data center is typically oversubscribed. Power oversubscription leads to efficient use of data center power hierarchy while simultaneously reducing the power provisioning cost. Power overload situations can occur in oversubscribed data centers. Power overload can lead to power capping of servers or even power outages – both of which degrade the performance of the services offered by the data center. In this dissertation, we address key open problems in the area of safe power oversubscription of data centers.

First, we quantify the level of safe power oversubscription possible for servers characterized by energy proportionality metric and workload distribution. By developing a theoretical framework to model the relationship between server energy proportionality and possible power oversubscription, we show how increasing server energy proportionality opens up the opportunity for more power oversubscription. Second, we develop a real-time dynamic power pricing mechanism to enable safe power oversubscription of multi-tenant data centers, an often neglected but important type of data center. Simulation results show that our new mechanism benefits both the tenant, by decreasing leasing costs, as well as the operator, by decreasing capital expense, and achieves the goal of keeping total power consumption under the data center power limit while reducing overall energy use. Third, we propose a coordinated priority-aware battery charging algorithm to tackle the problem of distributed battery charging in oversubscribed data centers. By coordinating the charging process, we charge the batteries according to the priorities of applications running on the servers and we are able to meet reliability service level agreements, while satisfying given power constraints.

Finally, we highlight the importance of taking workload characteristics into account when seeking to identify the most energy efficient server and develop a new server energy efficiency metric that is both linear as well as reliable in ranking of servers. The findings presented in this dissertation, if widely used, can result in energy savings, as well as capital cost savings, of millions of dollars per year.

Download

Share

COinS