The International Journal of High Performance Computing Applications published an RCC-led paper online last month.
PhD student and RCC Systems Engineer Mark Endrei is the lead author of the paper titled, ‘Statistical and machine learning models for optimizing energy in parallel applications.’
Mark wrote the paper alongside RCC colleagues Dr Jin Chao, Dr Minh Dinh and Professor David Abramson, and with US-based Heidi Poxon and Dr Luiz DeRose of supercomputer manufacturer Cray, and Professor Bronis R. de Supinski of federal research facility Lawrence Livermore National Laboratory.
With rising power costs and constraints driving a growing focus on the energy efficiency of HPCs, their research involves statistical and machine learning models for tuning parallel applications, which usually are energy and time-intensive on HPCs.
“The International Journal of High Performance Computing Applications is a well-regarded journal in our field, so it is a great platform for publishing our work, made greater because it will be the CCDSC 2018 special edition,” said Mark.
The Clusters, Clouds and Data for Scientific Computing (CCDSC) 2018 workshop in France last September was an invitation-only event where HPC leaders evaluated current and future trends for parallel computing. RCC Director Professor Abramson presented the work, enabling the team to make the paper submission to the HPC journal’s special edition.
“Supercomputers use lots of electrical power, so techniques that improve energy efficiency are important for reducing operating costs and greenhouse gas emissions,” explained Mark.
“We developed models that allow end users to tune the performance and energy efficiency of their scientific applications running on HPC systems. We show that our approach can be used to make accurate trade-off predictions at low cost, identifying cases where energy efficiency improves by up to 45% with only a 10% drop in performance. Our work will lead to a practical tool that allows scientists to tune their energy use and performance.”
The paper was first published online on 25 April 2019. The CCDSC 2018 special edition is expected to be published either by the end of 2019 or early 2020.
