Re: Someone installed an aarch64 Christmas gift: There is now a ampere4 with 150 or so FreeBSD CPUs and 24 builders running in parallel [an ampere5 exists too]

On 12/28/25 09:13, Mark Millard wrote:
> On Dec 27, 2025, at 08:08, Mark Millard <[email protected]> wrote:
> 
>> On Dec 25, 2025, at 18:18, Mark Millard <[email protected]> wrote:
>>
>>> On Dec 25, 2025, at 16:06, Mark Millard <[email protected]> wrote:
>>>
>>>> On Dec 25, 2025, at 15:10, Mark Millard <[email protected]> wrote:
>>>>
>>>>> ampere4 was showing 52.27 for the short term load
>>>>> average and that that was 33%. So, something like
>>>>> 150+ FreeBSD CPUs (1 per core I assume), if I
>>>>> understand right.
>>>>>
>>>>> The Jobs Id's were 01 .. 24 , so 24 poudriere builders.
>>>>>
>>>>> It was also showing: SwapInfo 100.00% when I looked
>>>>> at that. So it may currently be RAM+SWAP limited for
>>>>> its current configuration.
>>>>>
>>>>> ampere4 was building 150arm64-quarterly. In 21:45:36
>>>>> it had built 10802 port-packages, including 9 that
>>>>> had taken over 10 hours to build and 10 more that
>>>>> had taken over 5 hrs to build (top 15 completed
>>>>> time takers as of when I looked).
>>>>>
>>>>> Hmm. It just showed a 105.?? load average with a 66%
>>>>> figure. (I do not remember the ?? digits.) That is
>>>>> a mean of well over 3 per builder. (For all I know
>>>>> ampere4 could be doing other things as well as the
>>>>> port-package builds.)
>>>>>
>>>>> Having ampere4 in use ought to help with the aarch64
>>>>> port package build-times issues, even if ampere1
>>>>> continues to fail fairly frequently.
>>>>
>>>> Looking at the Ampere site, it looks like ampere4 is
>>>> a 160 core AmpereOne or AmpereOne M system. The "M"
>>>> as more RAM channels but only up to 1.5 TiBytes: 1 DIMM
>>>> per channel, 12 channels. The plain AmpereOne can have
>>>> 4 TiBytes: 1 or 2 DIMMs per channel, 8 channels.
>>>>
>>>> I've no clue how much RAM ampere4 actually has, or what
>>>> its RAM+SWAP figure is as configured.
>>>>
>>>
>>> While I've not found a definitive source of information,
>>> it does appear that ampere4 likely does not support armv7
>>> or aarch32 code, so no native lib32, armv7 chroot, or
>>> armv7 jail support. Native use of the armv7 build jails
>>> would need to run on ampere[123] instead.
>>
>>
>> There is another  new ampere* : ampere5 looks to be
>> another 160 core ampere ["(21%) 32.92" for a short
>> term load average]. It looks to be doing some
>> sort of experimental build of 150arm64:
>>
>> Master: 150arm64-default-foo
>> Build : 2025-12-24_21h05m27s
>> . . .
>> Set:    foo
>> . . .
>>
>> It is using 13 poudriere builders (Jobs Id's 01 .. 13).
>> It has been building for a little under 67 hr and has
>> 18278 built 17839 remaining.
>>
>> Note: "foo" is a poudriere Set name, which for normal
>> Package Builds is "default". And Build names that are
>> encoded date and time information is also not normal.
> 
> 
> ampere 4 finished 150arm64-quarterly. For as the build
> was configured:
> 
> Queued : 37202
> Built  : 35501
> Failed :   154
> Skipped:   554
> Ignored:   993
> 
> Elapsed: 78:25:36
> Pkg/Hr : 455
> 
> Example large build times:
> 
> 48:08:19 chromium-143.0.7499.169 (build/timeout)
> 44:28:50 electron37-37.6.0       (success)
> 39:41:27 qt6-webengine-6.9.3     (success)
> 
> 
> ===
> Mark Millard
> marklmi at yahoo.com
> 
> 
> 

ampere4 has now also finished building 143arm64-quarterly from scratch:

Queued : 37202
Built  : 35544
Failed :   118
Skipped:   548
Ignored:   992

Elasped: 60:06:37
Pkg/Hr : 594

Example large build-time port packages:

48:08:04 chromium-143.0.7499.169 (build/timeout)
43:28:43 electron37-37.6.0       (success)
25:43:38 qt6-webengine-6.9.3     (success)

So, on the same machine, 150arm64-quarterly took notably longer to build 
everything:

78:25:36 / 60:06:37 approx.= 1.30

That ratio is similar to other past comparisons of official 14.3 and 
before vs. 15.0 and later builds of (such builds). A personal build test 
got such a ratio for 14.3-RELEASE vs. the 14.3-STABLE at the time, 
indicating that stable/14 has the change(s) that make the difference in 
the build times.


-- 
===
Mark Millard
marklmi at yahoo.com