TurboBench

Why it's ranked here

Few open tools put this many codecs behind one command line. The build pulls each codec in as a submodule, enables it only when that source is present, and CI compiles the result with GCC and Clang on x86, ARM, RISC-V, macOS and Windows before benchmarking a standard corpus. That makes it a practical way to settle a codec choice on your own data rather than trusting a vendor chart. The case is weakened by the marketing: the README calls itself the only benchmark of its kind, and the comparison table against lzbench is written by the author.

What's good

Every decoded buffer can be compared byte for byte against the input, so a codec that silently corrupts data shows up as an error rather than a fast number. The decode check can also abort the process, which the source notes is for fuzzing. Codec parameters pass through on the command line, so you can test a tuned LZMA dictionary or match finder, not only preset levels. Preset groups are plain lines in a config file you can edit. Each nightly result table marks the speed and ratio frontier in bold, which is the useful reading of a benchmark.

Tradeoffs

Heap peak tracking works by replacing the system allocator on Unix builds. On Windows and static builds the heap peak functions are compiled to return zero, which undercuts the README promise of peak memory reporting there. The author's own codec, LzTurbo, sits at the top of the codec list and in the default groups, but its plugin header lives outside the repository, so public builds cannot reproduce those rows. Building everything needs CMake, and some codecs need Cargo or Zig; CI itself turns several off. All compiler warnings are suppressed. The codec table flags Iguana as not reliable and mzip as too slow.

How to use it well

Use it when you are choosing a codec for a specific data shape, such as logs, columnar numbers or web assets, and want numbers from your own hardware. Clone with submodules, build, list compiled codecs first, then run a preset group on a representative sample and rerun with tuned parameters for the two or three finalists. Treat memory figures as Unix only. Keep a separate result file per experiment set so merges stay comparable. It does not cover streaming, multithreaded throughput tuning in production, or compatibility between versions of a format.

Observed

License

GPL v2 (header in turbobench.c)

Language

C and C++ (turbobench.c, plugin.cc)

Interface

Command-line executable

Install

Recursive git clone then make, or a prebuilt release binary

Build dependencies

make and CMake; Cargo and Zig for some optional codecs (makefile)

CI platforms

Linux x86_64 and ARM (GCC, Clang), RISC-V, macOS, Windows MSYS2 (.github/workflows/build.yml)

Codec inclusion

Each codec compiled only when its submodule directory is present (makefile)

Heap peak memory

Compiled to zero on Windows and static builds (turbobench.c)

Preset groups

FAST, MAX, OPTIMAL, BWT, WEB and others defined in turbobench.ini

Output formats

Text, HTML, Markdown, CSV, SVG charts