Centmin Mod compression benchmarks

[SolidComp]

Hi all – The Centmin Mod project ran a very interesting set of benchmarks that includes memory use during compression and decompression. It includes brotli, zstd, pigz, gzip, and more.

He also includes CPU use, but it looks like it's 99-100% for all codecs so that stat must be erroneous.

I wish every compression benchmark reported memory and CPU...

(Centmin Mod is a project that bundles an optimized nginx build with PHP, MariaDB, and a few other pieces.)

[Cyan]

Updated version, with graph :
https://community.centminmod.com/thr...-xz-etc.18669/

[SolidComp]

Oh that looks nice. Zstd is pulling away from brotli for some reason.

[Jarek]

Let's put the plot here:

[schnaader]

Some nitpicking from my side:

- In the tables, the column "compress size (MB)" is actually in MiB, e.g. the xz level 9 result is around 46.47 MiB => ~46.47 * 1024 * 1024 => ~48.7xx.xxx bytes.
- Automatically determined block sizes destroy multi-threading on higher levels for some algorithms (which can be seen in the table column "compress cpu %"). This can be reproduced/improved when explicitly setting block sizes (see xz results below). So some comparing apples to oranges here (e.g. plzip vs. xz on highest level is similar ratio, but 200% vs. 100% cpu), but at least it's clear what happens in the tables - and it's actually a good test on how well compressors perform when used without any extra command line parameters.

Results from my quick tests (to verify results and compare with Precomp):

Command line				compressed size		compression time	decompression time	notes
precomp					46.783.211 (44.62 MiB)	1 min 36 s		6 s			1
precomp -t+				48.920.911 (46.65 MiB)	1 min 31 s		5 s			1, 2
precomp -lm4000				46.630.387 (44.47 MiB)  2 min 49 s		6 s			3
xz -9					48.765.608 (46.51 MiB)	2 min 35 s		4 s
xz -9 -T 0				48.812.012 (46.55 MiB)	2 min 34 s		4 s
xz -9 -T 0 --block-size=100663296	48.911.256 (46.65 MiB)	1 min 24 s		4 s			4

notes:
test machine: cloud server, 2 vCPU, 2 threads, Skylake, ~2.1 GHz, 4 GB RAM
1) automatically determined block size: 96 MiB on this machine, can be different on others
2) this is pure LZMA2 without recompression, should be very similar to multithreaded xz -9
3) allowing more memory usage improves compression ratio, but leads to 192 MiB block size, so no multithreading boost anymore
4) 100.663.296 bytes = 96 MiB to make it comparable to the first Precomp result

[SolidComp]

@schnaader That's a good cloud instance. Where do you get it and how much does it cost?

[Bulat Ziganshin]

@schnaader That's a good cloud instance. Where do you get it and how much does it cost?

something like that is 5 euro/month at https://www.hetzner.com/cloud

[schnaader]

something like that is 5 euro/month at https://www.hetzner.com/cloud

And that's exactly the one I use

[SolidComp]

something like that is 5 euro/month at https://www.hetzner.com/cloud

Interesting. That would cost $20 - $40 per month at Vultr or Digital Ocean in the US. And it might not be Skylake, but Haswell or something instead.