• Overview
  • Measurements
• Test Cases
  • Brainfuck
    • bench.b
    • mandel.b
  • Base64
  • Json
  • Matmul
  • Primes
• Tests Execution
  • Environment
  • Using Docker
  • Manual Execution
    • Prerequisites
• Contribution
  • Makefile guide
    • Binary executables
    • Compiled artifacts
    • Scripting language
  • README update
  • Docker image update

The benchmarks follow the criteria:

• They are written as the average software developer would write them, i.e.

  • The algorithms are implemented as cited in public sources;
  • The libraries are used as described in the tutorials, documentation and examples;
  • The used data structures are idiomatic.

• The used algorithms are similar between the languages (as the reference implementations), variants are acceptable if the reference implementation exists.

• All final binaries are releases (optimized for performance if possible) as debug performance may vary too much depending on the compiler.

My other benchmarks: jit-benchmarks, crystal-benchmarks-game

The measured values are:

• time spent for the benchmark execution (loading required data and code self-testing are not measured);
• memory consumption of the benchmark process, reported as base + increase, where base is the RSS before the benchmark and increase is the peak increase of the RSS during the benchmark;
• energy consumption of the CPU package during the benchmark: PP0 (cores) + PP1 (uncores like GPU) + DRAM. Currently, only Intel CPU are supported via the powercap interface.

All values are presented as: median_{±median absolute deviation}.

UPDATE: 2024-06-06

Testing brainfuck implementations using two code samples (bench.b and mandel.b). Supports two mode:

• Verbose (default). Prints the output immediately.
• Quiet (if QUIET environment variable is set). Accumulates the output using Fletcher-16 checksum, and prints it out after the benchmark.

Brainfuck

Mandel in Brainfuck

Testing base64 encoding/decoding of the large blob into the newly allocated buffers.

Base64

Testing parsing and simple calculating of values from a big JSON file.

Few notes:

• gason mutates input strings;
• simdjson requires input strings with batch of trailing zeros: a special zero padding for SIMD instructions;
• DAW JSON Link "NoCheck" skips some JSON structure correctness checks;
• DAW JSON Link, gason, default (not "Precise") RapidJSON, and D implementations except Mir-based have some inaccuracies in number parsing:
  • DAW JSON Link's number parsing issue
  • gason's number parsing issue
  • D stdlib number parsing issue

Json

Testing allocating and multiplying matrices.

Matmul

Testing:

• generating primes using the optimized sieve of Atkin;
• prefix search for their decimal numbers using Trie data structure.

Notes:

• All languages but V and Python use unordered hashmaps (V and Python don't provide those out of box, and their hashmaps use keys in the insertion order);
• The results are always sorted (could be unstable or stable though).

Primes

CPU: Intel(R) Xeon(R) E-2324G

Base Docker image: Debian GNU/Linux bookworm/sid

Build the image:

$ docker build docker/ -t benchmarks

Run the image:

$ docker run -it --rm -v $(pwd):/src benchmarks <cmd>

where <cmd> is:

• versions (print installed language versions);
• shell (start the shell);
• brainfuck bench (build and run Brainfuck bench.b benchmarks);
• brainfuck mandel (build and run Brainfuck mandel.b benchmarks);
• base64 (build and run Base64 benchmarks);
• json (build and run Json benchmarks);
• matmul (build and run Matmul benchmarks);
• primes (build and run Primes benchmarks);

Please note that the actual measurements provided in the project are taken semi-manually (via shell) as the full update takes days and could have occassional issues in Docker.

There is a ./run.sh that could be used to simplify Docker usage:

• ./run.sh build (build the image);
• ./run.sh make versions (run the image with the versions command);
• sudo ./run.sh shell (run the image with the `shell' command, sudo is required to read energy levels).

Makefiles contain recipes for building and executing tests with the proper dependencies. Please use make run (and make run2 where applicable). The measurements are taken using analyze.rb script:

$ cd <test suite>
$ ../analyze.rb make run
$ ../analyze.rb make run[<single test>]

Please note that the measurements could take hours. It uses 10 iterations by default, but it could be changed using ATTEMPTS environment variable:

$ ATTEMPTS=1 ../analyze.rb make run

Please use Dockerfile as a reference regarding which packages and tools are required.

For all (optional):

• Powercap for reading energy counters in Linux (Debian package powercap-utils).

For Python:

• NumPy for matmul tests (Debian package python3-numpy).
• UltraJSON for JSON tests (Debian package python3-ujson).

For C++:

• Boost for JSON tests (Debian package libboost-dev).
• JSON-C for JSON tests (Debian package libjson-c-dev).

For Rust:

• libjq for jq test (Debian packages libjq-dev, libonig-dev and environment variable JQ_LIB_DIR=/usr/lib/x86_64-linux-gnu/).

For Java, Scala:

• Coursier for downloading Maven artifacts.

For Haskell:

• network for TCP connectivity between the tests and the test runner.
• raw-strings-qq for raw string literals used in tests.

For Perl:

• cpanminus for installing modules from CPAN (Debian package cpanminus).

For Vala:

• JSON-GLib for JSON tests (Debian package libjson-glib-dev).

Please follow the criteria specified in the overview. Besides that please ensure that the communication protocol between a test and the test runner is satisfied:

• The test runner listens on localhost:9001;
• All messages are sent using TCP sockets closed immediately after the message has been sent;
• There are two messages sent from a test (it establishes the measurement boundary):
  1. The beginning message having the format name of the test/tprocess ID (the process ID is used to measure the memory consumption). Please note that the name of the test couldn't use Tab character as it's a delimiter;
  2. The end message with any content (mostly it's "stop" for consistency).
• The test runner could be unavailable (if the test is launched as is) and the test should gracefully handle it.

If the test is compiled into a single binary, then two sections of the Makefile require changes:

• append a new target (the final binary location) into executables variable;
• append the proper target rule.

If the test is compiled, but can't be executed directly as a binary, then three sections of the Makefile require changes:

• append a new target (the final artifact location) into artifacts variable;
• append the proper target rule to compile the test;
• append run[<target_artifact>] rule to run the test.

If the test doesn't require compilation, then two sections of the Makefile requires changes:

• append run[<script_file>] into all_runners variable;
• append run[<script_file>] rule to run the test.

TOC is regenerated using git-markdown-toc:

./run.sh toc