Ubuntu 21.10 Compressing Debian Packages With Zstd

[discordian]

Are there any benchmarks for this? Intuitively I would have thought that decompression times would be insignificant compared to download times (and the actual disk writes) and a slower but higher ratio compressor would thus lead to better performance overall?

If you would download/decompress in parallel (potential on more cores) then yes. They way dpkg/apt work is however completely serial so decompression comes afterwards, and no parallelism in sight. Further it calls sync quite often in an attempt to improve reliability (stalls any pending writes).

So no, decompression takes up alot of time. At the expense of using more memory you could mitigate that.

[uid313]

The drawback of this is that while Zstd decompresses faster, it also compresses less so the files do get bigger.

I wonder what compression algorithm Apple and Microsoft use for their OS updates.

[iruoy]

Are there any benchmarks for this? Intuitively I would have thought that decompression times would be insignificant compared to download times (and the actual disk writes) and a slower but higher ratio compressor would thus lead to better performance overall?

https://lists.archlinux.org/pipermail/arch-announce/2020-January/000400.html

Recompressing all packages to zstd with our options yields a total ~0.8% increase in package size on all of our packages combined, but the decompression time for all packages saw a ~1300% speedup.

They were using XZ compression before. I'm not sure what Debian/Ubuntu are using at the moment.

[david-nk]

The drawback of this is that while Zstd decompresses faster, it also compresses less so the files do get bigger.

Yes, but the increase is so marginal it doesn't matter. If we assume the 1% average increase in package size is accurate and we take a Firefox-sized package, we save 0.05s of download time on a 100mbps connection when using XZ instead of zstd, but we pay more than 3 seconds of additional decompression time. On a slow CPU or some shitty VPS, it will be well above 10 seconds of additional decompression time. Not to mention that the vast majority of package installations happens inside data centers with 1/10/100 Gbps lines and Debian packages are served from the local network instead of the official Debian/Ubuntu servers.

I wonder what compression algorithm Apple and Microsoft use for their OS updates.

Deflate/zlib and Apple theoretically also supports bzip2. Wouldn't take lessons from either of them if I were you.

[pranav]

there are quite a few people who really hate zstd because company behind its implementation. same as people who shun selinux for the very same reason.

ELI5 on this comment would be nice. Selinux is not good because...

[oleid]

haha, I decompress the zstd archive just to compress it again with btrfs. If only I could skip decompression