On 25/07/2026 13:39, Bruno Haible wrote:
Pádraig Brady wrote:
That probably makes sense. I'm quite surprised with the systemd runtime
overhead:
$ strace -c /bin/who 2>&1 | grep total
100.00 0.003110 6 459 15 total
$ strace -c git/coreutils/src/who 2>&1 | grep total
100.00 0.000000 0 140 8 total
My main point here is that the systemd variant had over 3x the number of
syscalls.
Indeed, libsystemd appears to read the per-session files multiple times,
and there are also many redundant fstat() calls:
$ strace 9.11-systemd/bin/who
For general info,
I also notice these extra systemd calls from ls,
even though it doesn't link directly with systemd!
It's coming from user db lookup which connects to systemd
(over polled sockets) though nss config.
You can see the difference between the default:
$ sudo strace -c /bin/ls -ld / 2>&1 | grep total
100.00 0.000488 2 179 17 total
and with systemd disabled in nss:
$ sed 's/systemd//' /etc/nsswitch.conf > /tmp/nsswitch-nosystemd.conf
$ sudo unshare -m -- bash -c '
mount --bind /tmp/nsswitch-nosystemd.conf /etc/nsswitch.conf
strace -c /bin/ls -ld /' 2>&1 | grep total
100.00 0.000614 4 146 17 total
Now systemd is providing support for DynamicUser=, homed, or nspawn containers
but it's unfortunate that it's quite heavy weight for all users
even with basic command like ls.
cheers,
Padraig
