Milen Nikolov

In 2013, following up on an inquiry and a problem posed by my adviser Prof. Zygmunt Haas, I drafted a research proposal and a strategy for dynamically right-sizing servers in data centers. Right-sizing referes to predictive policies for turning servers on/off so that cloud applications demands and reliability are met, utlizing the least number of powered servers possible at a given time. Essentially, it is a combinatorial/stochastic optimization problem. A proper right-sizing technique would decrease data centers energy consumption and also make their operation "greener". The research proposal was awarded a NSF EAGER Grant, and the PI is my thesis adviser Prof. Zygmunt Haas, starting January 2014.

Right-sizing techniques are currently under research; below is an excerpt from the proposal abstract:
"Energy consumption is one of the most important practical and timely problem associated with data centers for cloud computing. The urgency of this problem has been exposed by both, the governmental agencies and the industry. While policies that consider the physical design of the data centers have been studied in the technical literature rather thoroughly, the operational characteristics of the data centers (e.g., the required performance of the executed applications, such as the type of services and the applications being supported, their QoS requirements, associated background server maintenance processes, etc.) have rarely been accounted for. In particular, there is a need to investigate the implications of exploiting the operational characteristics of the data centers in the context of power management policies, as those could lead to potentially significant energy savings and operational cost reduction of a computer cloud.

A particular interest of our study will be the design of such policies in the context of distributed data centers, as provided by the cloud-computing paradigm. In particular, we will investigate the problem of Power Management Policies in heterogeneous data centers for the case of independent/individual data centers from the perspective of job profiling, such as failure and latency tolerance. We will study the possibility of decomposition of the peak loads to a data center into high latency tolerant requests that need to be migrated to different, less busy data centers, or high failure tolerant requests that require less server resources being activated at peak times. Similarly, we will study the approach of grouping jobs, as to reduce the demand variability on servers. Such decomposition and grouping will be dynamically performed."