Formal Metrics for Large-Scale Parallel Performance

"Formal Metrics for Large-Scale Parallel Performance." Kenneth Moreland and Ron Oldfield. In High Performance Computing, July 2015. DOI 10.1007/978-3-319-20119-1_34.


Performance measurement of parallel algorithms is well studied and well understood. However, a flaw in traditional performance metrics is that they rely on comparisons to serial performance with the same input. This comparison is convenient for theoretical complexity analysis but impossible to perform in large-scale empirical studies with data sizes far too large to run on a single serial computer. Consequently, scaling studies currently rely on ad hoc methods that, although effective, have no grounded mathematical models. In this position paper we advocate using a rate-based model that has a concrete meaning relative to speedup and efficiency and that can be used to unify strong and weak scaling studies.

Full Paper

Formal Metrics for Large-Scale Parallel Performance

Supplemental Material

You can easily use any spreadsheet program (such as Microsoft Excel) or any other plotting program to generate plots based on the metrics in this paper. The plots in this paper were generated with a Python module called toyplot. I built the scripts as self-documenting IPython notebooks and provide them here as supplemental material for examples on how to compute and use these metrics. Even if you do not plan to use the same tools I am using, you might find detail useful when replicating the detail yourself. You can download the archive of scripts, data, and results or you can browse the material in the following web pages.

You can also download the presentation slides I used at ISC 2015. These slides do not give enough explanation to really understand the concepts (that is what the speaker and paper are for), but it might be useful if you want to present this information to others.

SAND 2015-4836 W