| Abstract: |
|
Recent exascale studies point to power consumption becoming the major design constraint in high-performance computing (HPC). Current scientific benchmarks, such as LINPACK, only evaluate HPC systems when running at full throttle, i.e., 100% workload, resulting in more of a focus on performance than on power and energy consumption. In contrast, efforts like SPECpower evaluate the energy efficiency of a server at varying workloads. This is analogous to evaluating the fuel efficiency of an automobile at varying speeds. However, the applicability of SPECpower to HPC is limited.
Given the absence of a scientific benchmark to evaluate the energy efficiency of HPC at different workloads, we propose GBench (Green Benchmark), a methodology to evaluate the energy efficiency of supercomputers enabling a more rigorous study of energy efficiency in HPC. We use LINPACK as a case study and demonstrate the efficacy of our methodology by identifying application parameters impacting performance and providing a systematic methodology to vary the workload of LINPACK. |
|
