Artifact Submission: NUMA-aware diffusion solver

This is a replication package containing code and experimental results related to the paper titled: HIGH-PERFORMANCE NUMA-AWARE IMPLEMENTATION FOR1 THE DIFFUSION EQUATION.

Overview

The artifact comprises the following directories:

• benchmark -- benchmarking scripts
• data -- the (bnd, cfg) model pairs on which the benchmarks were run
• plots -- scripts for generating results plots
• presented-results -- plots (including some that were not included in the paper), CSV data files with measurements and R script that generated the plots from the data
• diffusion-benchmarking -- all the implementations of diffusion solvers

Detailed artifact contents

diffusion-benchmarking/src directory contains the source files to the algorithm kernels. Notably, *_solver.h files each contain a single implementation of the diffusion solver. See the header comments for details on the implementation.

Requirements for running the experiments

Software requirements:

• cmake 3.22 or later
• R software for plotting the graphs (see details below)
• python3 for running the benchmarking scripts

Let us present a few commands for your convenience that will allow you to set up the environment quickly:

Installing all dependencies on Debian/Ubuntu:

sudo apt-get update && sudo apt-get install -y g++ cmake r-base python3

Installing all dependencies on RHEL-like distribution:

sudo dnf install -y cmake gcc-c++ R python3

R packages necessary for generating the plots:

R -e "install.packages(c('ggplot2', 'cowplot', 'sitools', 'viridis', 'dplyr', 'tidyr'), repos='https://cloud.r-project.org')"

Running the experiments

Our experiments are designed to provide a comprehensive analysis of the aforementioned algorithms running various combinations of parameters computing different sizes of input instances. Therefore, the overall duration of running the experiments is quite long (around 2 to 3 days on MareNostrum SC).

To provide a swift way to check the reproducibility of our experiments, we prepared a special script that runs only a subset of the benchmarks.

Kick the tires:

Just to see whether the code is working, run the following from the root directory:

./kick-the-tires.sh

The script should take just a few minutes to finish. The script runs a subset of the experiments.

After the script runs, it will generate results in a CSV format in the results directory. It should contain 2 CSV files for space-local and temporary-local algorithms respectively. Each CSV file contains self-documenting headers. Finally, the plotting script is executed generating a single plot in the plots directory. More details on how the CSV results rows are processed into plots can be found in the plots/plots-fast.R script.

The generated plot file will be named data-local-normalized.pdf and it shows the comparison of the aforementioned algorithms.

Complete set of measurements:

To run the complete benchmark, execute

./run-all.sh

Measured results

The measured data and plots were stored in the presented-results directory. The directory also contains plots.R script, which was used to plot the data. It can be executed by Rscript plots.R if you wish to re-generate the plots from the data yourself. Here we describe each figure in the presented-results/plots directory:

Plot	Description	Figure number in paper
all-best.pdf	Comparison of all algorithms	Figure 10
data-local-normalized.pdf	Comparison between temp- and space-local algorithms	Figure 6
full-blocking.pdf	Different configurations of blocked approach	Figure 9
temp-local-tile-transpose.pdf	Comparison of planar, tiled and tiled+transposed variants	Figure 8
temp-local-tile.pdf	Various tile sizes of tiled algorithm compared to the temp- and space- local variants	Figure 7