On Feb 24, 2025, at 00:36, Mark Millard <[email protected]> wrote:
> On Feb 23, 2025, at 22:14, Yuri <[email protected]> wrote:
>
>> On 2/23/25 21:46, Mark Millard wrote:
>>> What port/package?
>>
>> finance/hyperswitch in arm64.
>>
>> It builds in a very high memory (20GB) on amd64, which is probably what
>> causes runaway builds on arm64.
>
> For the bulk -ca with 14 builders for 14 VM cores on the M4 Max, each
> builder allowed: 14 processes :
>
>
> build of finance/hyperswitch | hyperswitch-2024.12.23.0_1 ended at
> 2025-02-17T11:36:22-08:00
> build time: 01:14:38
>
>
> It likely had 13 other builders competing for cores over much of that time.
> By itself it would have taken less time. (High load averages.)
>
> I could potentially force a rebuild by itself and monitor its RAM+SWAP use,
> which would be for the core count involved in my context. Also, the time
> scale factor would give a clue about the load average ratio.
Note: Context is UFS, not ZFS. So no ARC RAM use. USE_TMPFS=all
NOTE: hyperswitch is not in my TMPFS_BLACKLIST (yet?).
[00:00:49] [02] [00:00:00] Building finance/hyperswitch |
hyperswitch-2024.12.23.0_1
[00:24:44] [02] [00:23:55] Finished finance/hyperswitch |
hyperswitch-2024.12.23.0_1: Success ending TMPFS: 12.65 GiB
So: a little under 24 min when run by itself.
Various "MaxObs(erved)" figures are:
load averages: . . . MaxObs: 9.86, 3.17, 1.76
Note: It spends lots of time under 1.5. It is not a
heavily parallel type of processing.
Mem: . . . 43057Mi MaxObsActive, 3103Mi MaxObsWired, 45479Mi
MaxObs(Act+Wir+Lndry)
Swap: . . . 43057Mi MaxObs(Act+Lndry+SwapUsed), 45479Mi
MaxObs(A+Wir+L+SU), 60897Mi (A+W+L+SU+InAct)
Note: The (A+W+L+SU+InAct) figure is from the MaxObs(A+Wir+L+SU) time frame
instead of being a MaxObs figure.
Note: Lndry (L) always observed as 0; SwapUsed (SU) always observed as 0.
(Plenty of RAM.)
Note: it is unclear what the dirty vs. clean pages counts are for
the difference 60897Mi - 45479Mi. Clean pages could have been freed
based on RAM pressure otherwise.
So: up to possibly near 59.47+12.65 == 72.12 GiBytes of RAM+SWAP required.
But likely more than 44.41+12.65 == 57.06 GiByes of RAM+SWAP required.
That is for: USE_TMPFS=all
That might not match the official builders.
If I ran it with RAM at 40 GiBytes in the VM, then the RAM pressure would
make the RAM+SWAP figure have a much narrower range. But that should be
with a realistic USE_TMPFS status as well if such a test is made.
===
Mark Millard
marklmi at yahoo.com
