Re: ampere2's main-armv7 crashed after only 3 min 51 sec, nothing started to build; libdm.so.{6->7} issues

On Oct 22, 2024, at 14:02, Philip Paeps <philip@freebsd.org> wrote:

> On 2024-10-22 16:52:26 (-0400), Ronald Klop wrote:
> 
>> Ampere2 is building armv7 again
>> https://pkg-status.freebsd.org/ampere2/build.html?mastername=main-armv7-default&build=p94c4ac6b071b_sc87b3f0006
>> BTW: I made a graph that shows how old a pkg repo is per architecture.It shows that arm/aarch64 is significantly older than i386/amd64.  But also that armv7 is way behind the rest now.
> 
> It takes aarch64 considerably longer to build a set than i386/amd64.

That is per "poudriere bulk -a" run. But there is also:

ampere1 builds a cycle of 4 distinct quarterly's
ampere3 builds a cycle of 4 distinct latest's

Both cover the 13.3/14.1 and arm64/armv7 combinations currently

ampere2 builds a cycle of 2 distinct latest's
(main for arm64 and for armv7)

> The builds are scheduled at the same frequency, but we don't start a new build while another build is still going.
> 
> Our aarch64 builders spend most of their time building about 19 different versions of LLVM (at last count).  This really slows down the pipeline.


There is also the processor time wasted on the 4
ports that almost always fail by build/timeout
for going over 48 hrs or so.
(#, Package, Origin, Skipped, Log, Time):

22 ungoogled-chromium-129.0.6668.100 www/ungoogled-chromium build/timeout 0 ??? 49:06:45
21 iridium-browser-2024.08.128.3 www/iridium build/timeout 0 ??? 49:00:07
23 chromium-129.0.6668.100 www/chromium build/timeout 9 ??? 48:33:54
45 electron30-30.5.1 devel/electron30 build/timeout 2 ??? 48:21:11

(So over 4*48 == 192 that could have been
spent producing successful builds instead,
per arm64 "pouriere bulk -a" across the 3
machines, or that could have been spent not
using as much power if more cores would
have been idle.)

It would seem that two better alternatives would be:

A) prevent those 4 builds from even being attempted
    on any of ampere1 .. ampere3

or:

B) change the configuration settings (such as
   timeouts or processes allowed) enough that
   the builds usually complete so that folks
   get an advantage from the build attempts

(A) would cut the elapsed time to the following
     armv7 bulk attempt (probably by at least
     something like 48 hrs)

(B) would instead increase that time but would
     produce more arm64 packages

armv7 already does not attempt to build those 4
packages.


Note: The devel/llvm* builds look to total to
something like 190 hrs as well.

===
Mark Millard
marklmi at yahoo.com