Home

Awesome

Bristol HPC Performance Portability Studies

This repository contains various data, scripts, and utilities for the study of performance portability (and productivity). It began with a collaboration between researchers at the University of Bristol and at Intel, and it is the original authors' hope that this material will help others in their exploration of this growing area of interest. We welcome contributions!

Organization

The repository is arranged as follows:

<root>
    AUTHORS                 # A list of contributors
    benchmarking/           # Performance portability data, organized by year and application, formatting in csv
    images/                 # A placeholder directory for output from scripts
    LICENSE                 # The software license governing the software in this repository
    metrics/
        data/               # Collated benchmark data in csv files, to be fed to scripts
        scripts/            # Processing scripts
            averages.py     # Compute different types of averages from datasets
            consistency.py  # Compute different types of variance/consistency-tracking scores from datasets
            heatmap.py      # Draw efficiency heatmaps
            pp_util.py      # Compute efficiency and PP, contains adaptive kernel estimation computations
            *.ipynb

Jupyter Notebooks

Jupyter is a web-based interactive framework for developing and presenting code, images, and equations in an integrated fashion. The .ipynb files in this repository are Jupyter notebooks used to produce the figures in the paper and presentations.

pp_vis.py

The pp_vis.py script contains implementations of the various visualization methods discussed in the companion paper. We hope it will prove useful to the community, both as a tool for evaluating datasets and as a springboard for new visualization and evaluation techniques.

The script can be used as a module for more sophisticated usage (as in the companion Jupyter notebooks), and can be used from the command line to generate plots.

This requires Python 3.7 or higher, Pandas 1.0 or higher, matplotlib, numpy, and scipy. These may be installed with pip or your local package manager.

Command-line usage is self-documenting with the $ ./pp_vis.py -h option, but to get started with the example data, try this:

$ ./pp_vis.py --raw-effs -F pdf ../data/synthetic.csv
Wrote ./synthetic_eff_cascade.pdf.
Wrote ./synthetic_estimated_density_chart.pdf.
Wrote ./synthetic_box_chart.pdf.
Wrote ./synthetic_binned_chart.pdf.

$ ./pp_vis.py --throughput -F pdf ../data/babelstream.csv
Wrote ./babelstream_eff_cascade.pdf.
Wrote ./babelstream_estimated_density_chart.pdf.
Wrote ./babelstream_box_chart.pdf.
Wrote ./babelstream_binned_chart.pdf.

$ ./pp_vis.py -F pdf ../data/neutral.csv ../data/minifmm.csv ../data/tealeaf.csv ../data/cloverleaf.csv
Wrote ./neutral_eff_cascade.pdf.
Wrote ./neutral_estimated_density_chart.pdf.
Wrote ./neutral_box_chart.pdf.
Wrote ./neutral_binned_chart.pdf.
Wrote ./minifmm_eff_cascade.pdf.
Wrote ./minifmm_estimated_density_chart.pdf.
Wrote ./minifmm_box_chart.pdf.
Wrote ./minifmm_binned_chart.pdf.
Wrote ./tealeaf_eff_cascade.pdf.
Wrote ./tealeaf_estimated_density_chart.pdf.
Wrote ./tealeaf_box_chart.pdf.
Wrote ./tealeaf_binned_chart.pdf.
Wrote ./cloverleaf_eff_cascade.pdf.
Wrote ./cloverleaf_estimated_density_chart.pdf.
Wrote ./cloverleaf_box_chart.pdf.
Wrote ./cloverleaf_binned_chart.pdf.

Data is expected to be in comma-separated csv format, with the first column being a list of platform names and each successive column the containing an application, with the results for each platform. An 'x' or 'X' may be used indicate that a platform did not run. By default, it is assumed that the input is in time-to-solution, but with the --throughput flag, this may be changed to be throughput. With the --raw-effs flag, the data is assumed to be in percentage efficiency already.

Multiple csv files can be passed in at once; all options are applied to each input..

Citing

The performance portability metric computed by these scripts was first proposed in the following two papers:

The datasets used here were originally collected for the following paper, which also served as the inspiration for the efficiency cascade plots: