SGER: Metrics And Methodologies for High Performance System Energy Benchmarking

点击翻译按钮获取中文摘要

Failure to maintain exponential increases in supercomputer performance due to power limitations could endanger U.S. competitiveness. Fundamental improvements in parallel and distributed system design are needed, but the lack of standard metrics and methodologies to evaluate energy efficiency for supercomputers provides little incentive to design energy efficient systems. Developing new metrics and methodologies is wrought with challenges since previous performance-centric benchmarks do not accurately reflect differences in energy efficiency. Additionally, for new benchmarks and methodologies to gain acceptance, results must be easily verifiable and benchmark purveyors must have credibility and traction in the domain. In this project, the PIs are proposing to create novel metrics and methodologies to evaluate energy efficiency for supercomputers.They will designe a distributed software test-harness that enables automated, repeatable, and verifiable measurements of both performance and energy efficiency. They will designe new energy efficiency benchmarks for supercomputers using the software test harness. The PIs, who started Green500 List of energy efficient supercomputers, are going to work with vendors and end usersto encourage them to benchmark their systems using these new methodologies.