Re: A 2025 NewYear present: make dpkg --force-unsafe-io the default?

[Bernhard Schmidt <berni@debian.org>]

Am 03.01.25 um 00:36 schrieb Ángel:
> On 2025-01-02 at 17:11 +0300, Michael Tokarev wrote:
> > 02.01.2025 03:00, Aurélien COUDERC wrote:
> >
> > > Sure but I wouldn’t know how to do that since I’m calling apt and
> > > force-unsafe-io seems to be a dpkg option ?
> >
> > echo force-unsafe-io > /etc/dpkg/dpkg.conf.d/unsafeio
> >
> > before upgrade.
> >
> > /mjt
>
> Beware: this should actually be
>
>   echo force-unsafe-io > /etc/dpkg/dpkg.cfg.d/unsafeio
>
> :)

Out of curiosity I've made a benchmark of my own. In my experience you 
hardly notice slowdown on desktop machines with modern NVMe drives, but 
you notice it on bad storage. I.e. I just happen to have a small (3 old 
nodes, 9 osd with cheap SSD) old CEPH cluster at hand and a bookworm VM 
with 42 pending updates on it. Small 2 vCPU/2GB RAM VM. Snapshotted 
after downloading all packages locally.

Shared infrastructure of course. Note that this includes an update of 
the initramfs, which is CPU bound and takes a bit on this system. You 
can take around 45s off the clock for the initramfs regeneration in each 
run. I did a couple of runs and the results were pretty consistent.

/dev/mapper/proxy1-root on / type ext4 (rw,relatime,errors=remount-ro)

time apt -y dist-upgrade; time sync

The following packages will be upgraded:
  base-files bsdextrautils bsdutils distro-info-data initramfs-tools 
initramfs-tools-core iputils-ping libblkid1 libc-bin libc-l10n libc6 
libcurl3-gnutls libfdisk1 libglib2.0-0 libglib2.0-data
  libmount1 libnghttp2-14 libnss-systemd libpam-systemd libpython3.11 
libpython3.11-minimal libpython3.11-stdlib libsmartcols1 libsqlite3-0 
libssl3 libsystemd-shared libsystemd0 libudev1
  libuuid1 locales locales-all mount nscd openssl python3.11 
python3.11-minimal systemd systemd-sysv udev util-linux util-linux-extra 
util-linux-locales

Default settings:

real    1m50.451s
user    0m52.333s
sys     0m33.574s

real    0m0.236s
user    0m0.007s
sys     0m0.001s


force-unsafe-io

real    1m47.309s
user    0m52.072s
sys     0m32.334s

real    0m0.324s
user    0m0.005s
sys     0m0.001s


eatmydata

real    1m36.407s
user    0m51.810s
sys     0m34.764s

real    0m0.334s
user    0m0.001s
sys     0m0.010s

So removing the time needed for the initramfs we have around 65s 
unchanged vs. 62s with force-unsafe-io vs. 51s with eatmydata. Or, < 5% 
speedup with force-unsafe-io on this system.

I have the snapshot still at hand, if there are other options you'd like 
to test drop me a note.

Bernhard