We (NetBSD) still have a historical construct in config(5) that a bunch of system-wide limits like MAXFILES are calculated from a presumed average or median amount of those resources per user, expressed as multiples of “maxusers [n]” in config(5).
We may wish to survey typical applications now in use and see what the distribution of {file, socket} descriptors “per user” looks like (standard distribution? bi-modal? tri-modal?) and see if our current multipliers still reflect actual use. Or perhaps use a different scheme/model for sizing kernel & system resource limits than “n per user." I’ve had to contend with our assumed limits by significantly raising the per-process soft limits for programs like pkgsrc/www/privoxy wherein descriptor use is proportional to traffic/clients rather than to number of processes. something to consider, Erik Fair > On Apr 18, 2018, at 02:27, Thomas Klausner <t...@giga.or.at> wrote: > > On Wed, Apr 18, 2018 at 11:10:37AM +0200, Martin Husemann wrote: >> On Wed, Apr 18, 2018 at 11:08:49AM +0200, Thomas Klausner wrote: >>> Did anyone else notice something similar? >> >> Check with fstat(1) ? > > Good idea. Right now, the top ones seem nearly ok: > > # fstat | sed "s/ [0-9].*$//" | sort | uniq -c | sort -n > ... > 24 root pbulk-build > 26 wiz at-spi2-registry > 30 wiz at-spi-bus-launc > 34 wiz zsh > 40 wiz dbus-daemon > 42 root sh > 45 root X > 49 bulk sh > 56 bulk cc1plus > 62 root sshd > 73 wiz transmission-gtk > 92 root master > 115 wiz syncthing > 142 wiz firefox > > I wonder about some of the numbers (master 92, sshd 62) but I don't > see anything eating thousands. Perhaps it's one particular package. > Thomas