저전력 프로세서 기반 슈퍼컴퓨터 활용 기술 타당성 연구

Abstract

In the operation of leadership-class supercomputing facilities, efficient management of power consumption is becoming a challenging issue that can affect not only the total cost but also the amount of space for cooling systems. Moreover, with the upcoming era of Exascale computers hat are expected delivered in 2019-2020, we need to devise effective solutions that can fulfill 20MW of power consumption. However, traditional multi-core processors will not be sufficient to satisfy the limited energy consumption requirement so that many-core coprocessors such as GPU or Xeon Phi should be appropriately utilized as we can see from the Top 500 list of supercomputers. An alternative approach can be leveraging low-powered and lightweight mobile processors which have been mainly used in portable devices such as smartphones, tablet PCs. Although mobile processors are not originally designed to support High-Performance Computing, they are optimized to maintain sustainable performance under a certain degree of power consumption (i.e. batteries). Therefore, the European project called Mont-Blanc has been conducted to design a new type of system built from energy efficient solutions used in embedded and mobile devices. Additionally, relatively low energy consuming ARM server processors are also becoming prevalent. Despite the potential computing powers of emerging mobile processors for HPC should be more investigated, effectively leveraging low-powered and lightweight processor technologies to build a supercomputing cluster can substantially minimize the overall system size along with the cooling device. In this study, we present our empirical study of building a small scale cluster called SLAP (Scalable, Low-powered, Autonomous & Robust, Pluggable) based on mobile processor technologies under different use cases and scenarios. Although mobile processor based cluster can have advantages in terms of physical space and energy efficiency, they have not mainly used for HPC servers so that we especially focus on the reliability and usability of our system. Through our empirical study, we could find potential benefits, possibility along with limits of applying embedded technology for supercomputing area. However, as the mobile rocessor and ARM server technologies evolve, energy efficient processors can become a competitive alternative to the road of Exascale computing.