> In the past, atime updates used to ruin performance.  Thanks to work by Ted
> Ts'o and others, that penalty has been greatly reduced (but not eliminated)
> by two options:
> * relatime (on by default): atime is not updated unless atime<=mtime or
>    atime<NOW-24h
> * lazytime (off by default): atime and mtime updates are postponed by up to
>    30 minutes, unless there's memory pressure or the inode has to be written
>    for another reason

> But alas, there are more considerations than just performance:
> * atime is really nasty for any CoW filesystems (btrfs, zfs, lvm snapshots,
>    qcow2 snapshots, deduped thin storage, etc).  For a /usr-y mix of files,
>    it costs around 5% of disk space per snapshot.  relatime offers no help as
>    its threshold (24h) just happens to match the most popular snapshot and
>    cronjob frequency (once per day).  Even worse, it can make _reading_
>    things result in ENOSPC!
> * atime murders media with sharply limited write endurance, such as SD
>    cards.  For this reason, every SoC pre-made image I've seen ships with
>    noatime.

> And what do we get in return?
> * the "vote" field in popcon.  This somewhat benefits the distribution (lets
>    us spot low-use packages installed by default, such as xterm (53.60% inst,
>    6.67% vote) or tcpd (95.20% inst, 9.85% vote)) but gives no gain to
>    individual users.
> * some rare forensic uses
> * new mail notification on login