Research topic and goals
The SSP (Sustained System Performance) metric is used to measure the performance of existing and future supercomputer systems at NERSC, NCSa, the Australian Bureau of Meterology and other sites. The SSP metric takes into account the performance of various scientific applications and input data sets (aka a “benchmark”), which represent some part of the sites’ workload.
In this collaboration, we propose the SSSP (Simplified Sustained System Performance) metric that makes performance projection using a set of simple existing benchmarks to the SPP metric for real applications. The benchmarks used as the “simple” may be existing simple benchmarks such as HPCC benchmark, HPL, parts of the SPECFP benchmark, and other simplified pseudo benchmarks which data already exist or easy to be measured.
The first objective of this collaboration is to study candidates and methodologies of SSSP benchmark. The important claim of the SSP benchmark is to measure the sustained performance by using real applications that are proportional, reliable, consistent, easy to use and repeatable, which will be useful for users. Although it is important to meet the requirements of various real-applications, not benchmarks, simple benchmarks are easy to port, optimize, execute and estimate their performance on various kinds of systems.
The second objective is to construct a set of benchmarks including traditional benchmarks for the SSSP and define the SSSP metric. We can investigate the appropriateness of the existing benchmarks by comparing the SSSP and SSP metrics on several systems.
The third objective is to investigate the appropriateness of the existing benchmarks that has been used for many years. Based on this investigation, we will update the SSSP by adding relatively new benchmarks and/or adjusting the weight of each benchmark.
We execute mini-applications and benchmarks on various systems including K-Computer at RIKEN and Blue Waters at NCSA and investigate the relationship between the SSP and SSSP metrics.
We consider a machine learning methodology to give appropriate weighting factors for the SSSP metris.
Results for 2016/2017
We have proposed the idea of Simpelified SSP and evaluated its consistency with the original SSP over several systems.
Visits and meetings
Miwako Tsuji visited NCSA in January 2017.
Impact and publications