CS Events

Qualifying Exam

Efficient Datacenter Management Using Deep Reinforcement Learning

 

Download as iCal file

Monday, March 06, 2023, 02:00pm - 03:30pm

 

Abstract:

Datacenters are expensive to build and operate, with the power infrastructure and electricity beingtwo significant construction and operational expenses. Further, carbon emission is a significant concernbecause of the large amount of electricity consumption. Thus, it is highly desirable to build managementsystems that maximize the efficiency and sustainability of datacenters.

Our research group is exploring the use of deep reinforcement learning (DRL) as a robust approach fordesigning and building datacenter management systems. As I will explain in my talk, accurately modelinga datacenter to build a simulator for training of a DRL agent is challenging and time consuming. I will givea brief overview of our current DRL DC management system, GreenDRL, and explain why simulationis necessary. I will then describe my work to model a microdatacenter’s (Parasol) cooling system forGreenDRL.

I will then present a high-level overview of an DRL approach that either does not require a detailedsimulator or can tolerate a high-level simulator that is much less effort intensive to implement andmaintain. I will explain the most important challenges and ideas for addressing them. For example,without simulation, it is hard to evaluate actions without data supported. Finally, I will conclude the talk with the current status on this research direction.

 

Speaker: Ning Gu

Location : CoRE 301

Committee

Professor Thu Nguyen (Advisor)

Professor  He Zhu

Professor Ulrich Kremer

Professor David Pennock

 

Event Type: Qualifying Exam

Abstract: See above

Organization

Rutgers University

School of Arts and Sciences

Department of Computer Science

Contact  Professor Thu Nguyen