This repository contains the source code of the Rio (Run-in-Order) runtime and benchmarks for both Rio and StarPU measuring their ability to execute fine-grained Sequential Task Flow (STF) programs.

The repository is organized as follow:

• rio: A Rust project containing two crates:
  • rio: The source code of the Rio runtime.
  • experiments: The source code of experiments used for benchmarking Rio.
• starpu: The source of experiments usef for benchmarking StarPU.
• bare_metal: A single threaded C implementation of the Rio and StarPU experiments, used for reference.
• scripts: Python scripts to automate the benchmarks.
• bench: A tool for collecting statistics about programs using performance counters, used by the python scripts.

Compiling and running the benchmarks require the following dependencies (in parenthesis are the versions we used):

• A C compiler (gcc 10.2.0)
• A Rust compiler (rustc 1.51)
• StarPU (1.3.8)
• hwloc (2.4.1)
• python (3.9.6)

The experiments are run using the benchmark scrips scripts/benchmark_efficiency.py and scripts/benchmark_workers.py. Both script produce a json file that can be used to plot different figures using the scripts/plot_efficiency.py and scripts/plot_workers.py scripts. All scripts accept a -h or --help argument that can be used to learn about configuration options.

To plot the efficiency decomposition, the following commands can be used:

python scrips/benchmark_efficiency.py -f efficiency.json --timeout 100 --average-on 3 --nb-threads 24 --experiment counter_deps --verbose
python scripts/plot_efficiency.py -f efficiency.json

And to plot the execution time for different number of workers with a miximum of 64 (2⁶) workers:

python scripts/benchmark_workers.py -f workers.json --timeout 100 --average-on 3 --nb-threads 6 --size 22 --task-size 20
python scripts/plot_workers.py -f workers.json