ReBench: Execute and Document Benchmarks Reproducibly

ReBench is a tool to run and document benchmark experiments. Currently, it is mostly used for benchmarking language implementations, but it can be used to monitor the performance of all kinds of other applications and programs, too.

The ReBench configuration format is a text format based on YAML. A configuration file defines how to build and execute a set of experiments, i.e. benchmarks. It describes which executable was used, which parameters were given to the benchmarks, and the number of iterations to be used to obtain statistically reliable results.

With this approach, the configuration contains all benchmark-specific information to reproduce a benchmark run. However, it does not capture the whole system.

The data of all benchmark runs is recorded in a data file for later analysis. Important for long-running experiments, benchmarks can be aborted and continued at a later time.

ReBench focuses on the execution aspect and does not provide advanced analysis facilities itself. Instead, the recorded results should be processed by dedicated tools such as scripts for statistical analysis in R, Python, etc, or Codespeed, for continuous performance tracking.

The documentation for ReBench is hosted at https://rebench.readthedocs.io/.

Goals and Features

ReBench is designed to

• enable reproduction of experiments;
• document all benchmark parameters;
• provide a flexible execution model, with support for interrupting and continuing benchmarking;
• enable the definition of complex sets of comparisons and their flexible execution;
• report results to continuous performance monitoring systems, e.g., Codespeed;
• provide basic support for building/compiling benchmarks/experiments on demand;
• be extensible to parse output of custom benchmark harnesses.

ReBench Denoise

Denoise configures a Linux system for benchmarking. It adapts parameters of the CPU frequency management and task scheduling to reduce some of the variability that can cause widely different benchmark results for the same experiment.

Denoise is inspired by Krun, which has many more features to carefully minimize possible interference. Krun is the tool of choice if the most reliable results are required. ReBench only adapts a subset of the parameters, while staying self-contained and minimizing external dependencies.

Non-Goals

ReBench isn't

• a framework for (micro)benchmarks. Instead, it relies on existing harnesses and can be extended to parse their output.
• a performance analysis tool. It is meant to execute experiments and record the corresponding measurements.
• a data analysis tool. It provides only a bare minimum of statistics, but has an easily parseable data format that can be processed, e.g., with R.

Installation

ReBench is implemented in Python and can be installed via pip:

pip install rebench

To reduce noise generated by the system, rebench-denoise depends on:

• sudo rights. rebench will attempt to determine suitable configuration parameters and suggest them. This includes allowing the execution of rebench-denoise via sudo without password and with the permission to set environment variables (SETENV).
• cpuset to reserve cores for benchmarking. On Ubuntu: apt install cpuset

Please note that rebench-denoise is only tested on Ubuntu. It is designed to degrade gracefully and report the expected implications when it cannot adapt system settings. See the docs for details.

Usage

A minimal configuration file looks like this:

# this run definition will be chosen if no parameters are given to rebench
default_experiment: all
default_data_file: 'example.data'

# a set of suites with different benchmarks and possibly different settings
benchmark_suites:
    ExampleSuite:
        gauge_adapter: RebenchLog
        command: Harness %(benchmark)s %(input)s %(variable)s
        input_sizes: [2, 10]
        variable_values:
            - val1
        benchmarks:
            - Bench1
            - Bench2

# a set of executables for the benchmark execution
executors:
    MyBin1:
        path: bin
        executable: test-vm1.py %(cores)s
        cores: [1]
    MyBin2:
        path: bin
        executable: test-vm2.py

# combining benchmark suites and executions
experiments:
    Example:
        suites:
          - ExampleSuite
        executions:
            - MyBin1
            - MyBin2

Saved as test.conf, this configuration could be executed with ReBench as follows:

rebench test.conf

See the documentation for details: https://rebench.readthedocs.io/.

Support and Contributions

In case you encounter issues, please feel free to open an issue so that we can help.

For contributions, we use the normal GitHub flow of pull requests, discussion, and revisions. For larger contributions, it is likely useful to discuss them upfront in an issue first.

Development Setup

For the development setup, the currently recommended way is to use pip install --editable . in the root directory of the repository. You may also want to use a virtual environment to avoid conflicts with other Python packages.

For instance:

git clone https://github.com/smarr/rebench.git
cd rebench
pip install --editable .

Unit tests and linting can be run with:

python -m pytest
python -m pylint rebench

Use in Academia

If you use ReBench for research and in academic publications, please consider citing it.

The preferred citation is:

@misc{ReBench:2018,
  author = {Marr, Stefan},
  doi = {10.5281/zenodo.1311762},
  month = {August},
  note = {Version 1.0},
  publisher = {GitHub},
  title = {ReBench: Execute and Document Benchmarks Reproducibly},
  year = 2018
}

Some publications that have been using ReBench include:

• Transient Typechecks are (Almost) Free, Roberts et al. 2019.
• Efficient and Deterministic Record & Replay for Actor Languages, D. Aumayr et al. 2018.
• Fully Reflective Execution Environments: Virtual Machines for More Flexible Software, G. Chari et al. 2018.
• Building efficient and highly run-time adaptable virtual machines, G. Chari et al. 2017.
• Improving live debugging of concurrent threads through thread histories, M. Leske et al. 2017.
• Adaptive Just-in-time Value Class Optimization for Lowering Memory Consumption and Improving Execution Time Performance T. Pape et al. 2016
• Cross-Language Compiler Benchmarking---Are We Fast Yet?, S. Marr et al. 2016.
• Tracing vs. Partial Evaluation: Comparing Meta-Compilation Approaches for Self-Optimizing Interpreters, S. Marr, S. Ducasse. 2015.
• Zero-Overhead Metaprogramming: Reflection and Metaobject Protocols Fast and without Compromises, S. Marr et al. 2015.
• Pycket: a tracing JIT for a functional language, S. Bauman et al. 2015.
• Adaptive just-in-time value class optimization: transparent data structure inlining for fast execution, T. Pape et al. 2015.
• Meta-tracing makes a fast Racket, C. F. Bolz et al. 2014.
• Cloud PARTE: Elastic Complex Event Processing based on Mobile Actors, J. Swalens et al. 2013.
• Identifying A Unifying Mechanism for the Implementation of Concurrency Abstractions on Multi-Language Virtual Machines, S. Marr, T. D'Hondt. 2012.
• Insertion Tree Phasers: Efficient and Scalable Barrier Synchronization for Fine-grained Parallelism, S. Marr et al. 2011.