777.zstd_r
SPEC CPU®2026 Benchmark Description

Benchmark Name

777.zstd

Benchmark Program General Category

Data compression/decompression

Benchmark Authors

The open source project zstd has had many contributors; see the list at https://github.com/facebook/zstd/graphs/contributors.

Benchmark Description

Zstandard, or zstd as short version, is a fast lossless compression algorithm, targeting real-time compression scenarios at zlib-level and better compression ratios. It's backed by a very fast entropy stage, provided by Huff0 and FSE library. Zstandard's format is stable and documented in RFC8878. Multiple independent implementations are already available. This repository represents the reference implementation, provided as a library and a command line utility.

The benchmark runs zstd's internal benchmark mode for fixed number of iterations, measuring both compressing speed and decompressing speed at various compression levels.

code is derived from version 1.5.6 of zstd.

The code has been altered from original distribution in few aspects (all under SPEC pre-processor define):

• Removing platform specific code
• Alter internal timing routines to return a constant delta between calls
• If few iterations are needed, small block of random data is altered in the input stream to create a difference in the input stream between iterations

Input Description

The benchmark options used for each workload are in a "control" file in each workload input directory, which specifies the command-line options for the zstd executable, as well as the input file (if any).

Command Line Options

The following command line options are used by the benchmark. There are many others; for details run the benchmark executable with the --help option.

• --verbose: Selects verbose output
• -b<i>: selects benchmark mode using compression level <i> (default: 3)
• -e<j>: test all compression levels successively from <i> (specified with -b) to <j> (default: 1)
• -i<n>: sets number of iterations of the benchmark loop to run (default: 3). Note - this behavior is specific, due to the change in the timing routines

Workload Details

The refrate workloads are designed to provide a robust and balanced test of system performance. The specific compression levels and input data were chosen based on the following rationale:

• Algorithmic Diversity: The selected compression levels (-b3 through -b19) intentionally target a wide variety of Zstandard's internal compression strategies. This ensures the benchmark exercises a broad range of its codebase and exhibits a range of microarchitectural behaviors.
• Balanced Runtimes: The number of iterations (-i) for each workload was tuned on the SPEC Development Reference Machine. This balancing ensures that each compression strategy contributes a roughly equal weight to the final benchmark score, preventing any single workload from dominating the result.
• Representative Data: The input file cld.tar is a diverse archive containing a mix of files reflecting a variety of real-world data types. More details are below.

--verbose is used on every workload and will not be included in the table below.

The data file used for the refrate workload (cld.tar) is an archive consisting of commonly used files, such as documents (HTML, PDF, PostScript, and text), source code, binaries, images, and compressed files. Specifically:

• PDF documentation from SPEC
• HTML documentation and associated assets from SPEC
• A SPEC CPU®2006 result in PostScript format
• An x86 Linux copy of the run and reporting tools(Perl scripts and executables) from SPEC CPU 2006
• Image files contributed by SPEC members
• A short MPEG-4 video contributed by a SPEC member
• A zip file containing two x86 binaries of Perl
• A text copy of the English translation of the ancient Sanskrit text Mahabharata from India from sacred-texts.com

Output Description

The output file contains a line for each compression level tested, giving the compression level, number of compressed bytes, compression ratio, pseudo performance metrics, and the data source. The performance metrics claim to be measuring compression performance in MB/s, but as the benchmark does not do internal timing, and the number of benchmark iterations is constant, these numbers also do not change from system to system.

Programming Language

C

Threading Model

The SPECrate version is single-threaded. There is no SPECspeed version.

Known Portability Issues

None

Sources and Licensing

zstd is offered under either the BSD or GPLv2 licenses. SPEC is using it under the BSD license.

For the refrate workload:

• SPEC has been granted the permission to use the text file containing Mahabharata. The FAQ at sacred-texts.com indicates this English translation is available in the public domain. Additionally, of the 18 chapters in the work, only 17 are used in the benchmark text. This fits within usage guidelines under the commercial use section of the terms of service.
• As mentioned above, the refrate workload includes source code files that are compressed by the benchmark. They are not compiled; they are compressed. Among them are filepp, distributed by SPEC as 'specpp', and licensed under GPL v2 or later. You can find the sources on your SPEC ISO image at redistributable_sources/modified/tools.src/filepp-1.8.0.tar.gz
• The source code files that are compressed include the CPAN Module PDF:API2 version 2.15, licensed under LGPL v2, and available at https://metacpan.org/pod/PDF::API2
• All other contents of the refrate input file are covered by the SPEC CPU license.

References

• Project page at https://facebook.github.io/zstd/
• Wikipedia: https://en.wikipedia.org/wiki/Zstd
• GitHub: https://github.com/facebook/zstd

Copyright © 2026 Standard Performance Evaluation Corporation (SPEC®)