CAREER: Commonality Exploiting Architectures for Energy Efficiency

Princeton University, Princeton NJ

Investigators

Abstract

An increasing amount of computation is moving into the Cloud and large-scale data centers. Hosted web-applications, such as social networks, hosted email, and photo sharing, are all examples of applications, which have moved into the Cloud. In order to control costs and save energy, maximizing the computational efficiency of datacenter computers is of paramount importance. This research project investigates computational efficiency in datacenter processors by looking for commonality between all of the applications that are executing in the Cloud. This project explores how to change a microprocessor core to save energy by exploiting the commonality available between applications in the Cloud. In addition, this project utilizes an integrated education plan that will disseminate information about computer architecture inside of the University, across the New York region, and across the world through the use of a Massively Open Online Course (MOOC). Accelerating data center and big data applications can have large societal impact by helping humans understand data that has been collected, thus enabling the understanding of social trends and public health challenges. The success of the research project will create more efficient data centers that will enable humanity to answer societal big data questions and have richer Internet experiences. This project investigates how to create computer architectures that can exploit similarity between programs executing on different cores across the data center. A software system identifies likely candidates that exhibit commonality. Once likely commonality has been identified, micro-architectural information is sent between different cores or within a single core to enable the execution of one program to reduce the energy needed to execute a second similar program. This project investigates what is the best micro-architectural information to share between cores executing common programs and how it is best to send that information between cores. In particular, this project investigates customizing on-chip networks specialized for sending commonality information and how to modify a manycore processor to harvest common micro-architectural information.