RE: arm64 pkg building is taking longer

Diego Linke <diego_at_bsd.com.br> wrote on
Date: Mon, 10 Feb 2025 12:07:19 UTC :

> I noticed that ARM64 packages take a few days longer to build and become
> available on the official FreeBSD servers compared to AMD64 and i386.

The amd64 systems are generally faster and there are
a lot more of them used as package builders.

There are only 3 aarch64 build systems: ampere[1-3].

By contrast, there are 10 amd64 build systems, 3 being
fairly modern and vastly faster than the ampere*'s
(far more hardware threads, for example).

ampere1 cycles through building and distributing:
141arm64-quarterly
141releng-armv7-quarterly
1341arm64-quarterly
134releng-armv7-quarterly

amd64 systems do not build that many variations on the
same machine: in fact each normally only builds one
variation, no waiting for other cycle members to
finish.

As each takes longer, the next time it gets back to the
same type of build, it is likely that far more packages
that need to be built (compared to the analogous amd64
context). It is not as extreme for quarterly as it is
for default (a.k.a. latest).

ampere3 is similar (default here is a.k.a. latest):
141arm64-default
141releng-armv7-default
1341arm64-default
134releng-armv7-default

ampere3 likely builds the most per poudriere bulk run
compared to ampere1 above, taking the largest to get
back to the next build of the same member of the cycle.

ampere2 has only 2 cycle members (as stands, main is 15.0):
main-arm64-default
main-armv7-default

So that makes 10 separate variations overall, spread
over the 3 ampere*'s.

amd64/i386 has 10 separate variations as well, but
1 per amd64 system that does the type build.
141amd64-quarterly , 141amd64-default , and
141i386-default are what get the fastest 3 builder
machines.

> For example, last Monday, Mat committed the Bind 9.2 (dns/bind920) update
> to 9.20.5 in the quarterly branch, which has two security fixes. This
> update was available on amd64 and i386 2 days later. It's still not
> available (after 7 days) for ARM64.

Expected sort of result, given the resources available to put
to use.

> Could we prioritize building packages with security updates, especially for
> the quarterly branch?

Already done: The more and bigger "default" builds do not
complete for the machine time on ampere1. Mixed on the
same machine the "default" builds would further delay the
quarterly builds.

> Is anyone aware of any initiatives to improve this
> process?

Unless the aarch64/armv7 system resources are considered
as part of the "process": it is not basically a process
problem. (I'm not intending to imply that "no optimization
is possible", just that such is not likely to lead to
a large change of scale for how long things take in
order to reach similar times to amd64 now takes.)

> PS: I’m aware that I can set up my own package build infrastructure using
> Poudriere and am considering it as a fallback option. However, I’d like to
> explore whether we can improve the process for everyone.

That last note repeats here: it is not basically a process
problem for what is the major constraint.



===
Mark Millard
marklmi at yahoo.com