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 Dec 30, 2025, at 07:23, Mark Millard <[email protected]> wrote:

> On Dec 30, 2025, at 00:59, Dave Cottlehuber <[email protected]> wrote:
> 
>> On Mon, 29 Dec 2025, at 22:37, Mark Millard wrote:
>>> Philip Paeps <philip_at_freebsd.org> wrote on
>>> Date: Mon, 29 Dec 2025 13:03:38 UTC :
>>> 
>>>> On 2025-12-29 18:14:11 (+0800), Ronald Klop wrote:
>>>>> Do you have some info about the new ampere4 and ampere5 servers? Maybe some tech specs and the expected use cases.
>>>> 
>>>> ampere4 and ampere5 are obscenely deep 2U machines that have been sitting in storage for a while because we didn't have racks that could accommodate them.
>>>> 
>>>> They're both Ampere Mt Collins machines.
>>> 
>>> So much for my earlier guessing.
>>> 
>>> QUOTE from Ampere_Mt._Collins_2U_PB_v0.50_20210223.pdf :
>>> Available in a dual socket configuration in a 2U form-factor with the 
>>> Ampere Altra®
>>> and Altra® Max processors, Mt. Collins provides 160 cores . . .
>>> END QUOTE
>>> 
>>> If I understand right, that means armv7 user space code likely is
>>> supported (Altra variants).
>>> 
>>> "8x 72-bit DDR4-3200 channels" and "Up to 16 DIMMs and 4 TB/socket".
>>> (2 DIMMs per channel are allowed.)
>>> 
>>>> Two CPUs each, for 160 cores per machine. There's 256G memory in ampere4 (16*16G) and 128G in ampere5 (8*32G).
>> 
>> Thanks Santa!
>> 
>> I’ll be interested to see how fast these eat ports for breakfast…
> 
> Both ampere4 and ampere5 got build/timeout in 48 hrs for building
> chromium. ampere4 has started the process of wasting 2 days doing
> that again, with nothing else building. Looks like either
> chromium should be blacklisted or given more time (given the
> build configuration). Wasting 2 days of cpu core time on 3 or
> so cores is the worst option.
> 
> But ampere4's configuration built 35501 port-packages in
> 78:25:36 to reach "done:" status in somewhat under 3.3 days.
> It was a build of 150arm64-quarterly .
> 
> On this scale, the next build was incremental instead of
> (nearly) from scratch: it has built 981 with 10 remaining
> (chromium and things depending on chromium). It had finished
> #981 in under 11 hrs (once started) and so would have been
> done at that point absent the chromium build being involved.
> (The chromium build is nearing 14 hrs at this point, about
> 34 hrs to go if it is let to run until build/timeout .)
> 
>> In my much smaller ports setup, I found I needed about 4Gb RAM per core at peak usage, but also these were using tmpfs for everything.
> 
> There are individual ports for which USE_TMPFS=all for
> poudriere uses 30 GiBytes+ of RAM+SWAP just for the tmpfs
> part of the overall RAM+SWAP use. (Such file system usage
> is not specific to poudriere style builds, tmpfs just
> uses RAM+SWAP instead of other media space.)

ampere5 completed but was using only 13 builders (like ampere[123])
and seemed to mostly stick to 3 cores or less per builder (also
like ampere[123]). Also it was using a different jail world, as
evidenced by:

Host OSVERSION: 1600007
vs. the:
Host OSVERSION: 1600000

that the rest are using. I do not know what all might be
different for the experiment they did.

Despite the difficulty in interpretation, given the above,
I'll note 3 150arm64-* build timings ("default" means:
latest):

ampere5 built 35997 for 150arm64-default in:   130:04:12 (started: 24 Dec 2025 21:05:27 GMT)
ampere2 built 35967 for 150arm64-default in:   188:58:54 (started: 20 Dec 2025 02:00:32 GMT)

(using 24 builders and so used lots more cores:)
ampere4 built 35501 for 150arm64-quarterly in:  78:25:36 (started: 25 Dec 2025 01:02:42 GMT)


===
Mark Millard
marklmi at yahoo.com