On Mon, Dec 03, 2018 at 10:33:44AM +0100, Alessandro Selli wrote:
> On 03/12/18 at 10:05, Adam Borowski wrote:
> > realtime greatly reduces atime writes, but it's still too much.
>
>   I wouldn't say so.  Since relatime updates atime only relative to the
> present ctime and mtime, it's only changed when one of those two is
> changed.  That is, updating atime does not require a separate write
> operation.

That was the original design -- but alas, it was later changed so the atime
is updated at least once a day.

> > Case in
> > point: it's the likely culprit for wasting the SD card that started this
> > thread (on a mostly-read load).
>
>
> if that filesystem was mounted with the relatime option (or with no
> option at all, since relatime is the default), then it's very unlikely
> it caused any more writes that if it was fully disabled.

You still rewrite every inode once per day.

> > And, update frequency of 1/day happens to
> > match the typical backup schedule, making it ruin snapshots just the same
> > as strictatime would.
>
>   Uh?  How can atime "ruin snapshots"?

Every inode has to be updated. That causes a lot of metadata churn, and
even takes significant space. Changing this single number tends to cost
far more than a page worth of space -- usual snapshot tools (btrfs, lvm,
etc) CoW more than just the inode. Thus, for some common loads we're
looking at 5% wasted disk space just for atimes.

> > So it's time to kill the nasty thing.
>
>   If only it was any nasty.

<troll>
Some folks say systemd isn't nasty.
</troll>


Meow!
--
⢀⣴⠾⠻⢶⣦⠀
⣾⠁⢠⠒⠀⣿⡁ Ivan was a worldly man: born in St. Petersburg, raised in
⢿⡄⠘⠷⠚⠋⠀ Petrograd, lived most of his life in Leningrad, then returned
⠈⠳⣄⠀⠀⠀⠀ to the city of his birth to die.