Fedora 34 Cleared For Btrfs Zstd Compression By Default, DNF/RPM Copy-On-Write

[Hi-Angel]

darkbasic I just upgraded to 5.10.9 kernel, and I can tell indeed apparently something was fixed, because results are changed. Now I see no difference in time or CPU usage when I build libinput with or without reflinks being used.

[S.Pam]

Oddly enough, CoW not necessarily is faster than creating a new file. At least when dealing with small-sized ones, which would be the case for package installations.

Incidentally, yesterday I benchmarked ccache with and without CoW (they added an opt-in feature to use reflinks instead of copying files). What I found is that building a libinput with reflinks is consistently 30% slower than building without.

I should also mention that I asked question about that on #btrfs IRC channel, but nobody replied.

Yes from 4.1 to 4.3 seconds.. maybe also some other bottleneck involved .there was a regression fixed in 5.10.7 and you benchmarked on 5.10.4. Try 5.9 or 5.10.7.

[Hi-Angel]

Yes from 4.1 to 4.3 seconds..

From 2.2 to 3.2. I don't know what those 4.05s-4.30s numbers refer to, but it's not the duration. You can check that by running e.g. `time sleep 1`.

[S.Pam]

From 2.2 to 3.2. I don't know what those 4.05s-4.30s numbers refer to, but it's not the duration. You can check that by running e.g. `time sleep 1`.

Perhaps redo those benchmarks with 'time' ?

[Hi-Angel]

Perhaps redo those benchmarks with 'time' ?

I think there's some miscommunication, because `time` is what I used for benchmarks. Btw, in case you missed my comment: I re-measured with Linux 5.10.9, apparently some regression did cause the difference. Now with and without reflinks there's no difference, which is definitely better.