You can use TurboBench [1] to benchmark libdeflate against zlib, igzip, zlib-ng and others.<p>Download TurboBench from Releases [2]<p>Here Some Benchmarks:<p>- <a href="https://github.com/zlib-ng/zlib-ng/issues/1486">https://github.com/zlib-ng/zlib-ng/issues/1486</a><p>- <a href="https://github.com/powturbo/TurboBench/issues/43">https://github.com/powturbo/TurboBench/issues/43</a><p>[1] <a href="https://github.com/powturbo/TurboBench">https://github.com/powturbo/TurboBench</a><p>[2] <a href="https://github.com/powturbo/TurboBench/releases">https://github.com/powturbo/TurboBench/releases</a>
Based on my benchmarks, ISA-l/igzip is more than twice(!) as fast as libdeflate and zlib for decompression. I'm almost enamored with ISA-l because of its speed. And yes, it works on AMD and also has Assembler code for ARM, so probably also works on ARM.<p>For parallelized decompression of gzip, I recommended my own tool, rapidgzip. I have measured up to 10 GB/s decompression bandwidth with it (>20 GB/s if an index already exists). I'm currently working on integrating ISA-l for even more special cases into rapidgzip and hope to release version 0.9.0 in the next days. It will have another +30-100% performance boost for many cases, thanks to ISA-l.
One nice thing about this is that the compression factor seems quite a bit higher than standard gzip. This is very useful if you need compatibility with gzip/zlib formats (e.g. PNG files). For my use case, the compressed output was 8% smaller using libdeflate compression=12 than with gzip compression=9. The compression was also a lot faster with a single thread. It would be nice to see a comparison with a larger number of inputs, however.