Re: What happen with vm.v_* on stable?

On 2025-03-08, Rozhuk Ivan <rozhuk.im@gmail.com> wrote:
> Hi!
>
> "vm.v_free_min", "vm.v_free_reserved", and "vm.v_free_target" sysctls describeb in i386 notes and some FBSD docs.
> I try to tune it - see no effects.
> I look into code of 14/STABLE and fail to find code that read these values.
>
> It seems it only show some initial on boot values and do nothink while OS work.

Yes. I noticed that when upgrading from R.12 to R.13, my tunings on
these values no longer work.

The matter is probably related to this (but I don't think you can tune
these easily - doing so would be somehow ambitioned anyway): 

admin@edge:515:1~$ sysctl vm.domain | egrep "free_(reserved|min|target)"
vm.domain.3.stats.free_min: 25825
vm.domain.3.stats.free_reserved: 5388
vm.domain.3.stats.free_target: 87136
vm.domain.2.stats.free_min: 25827
vm.domain.2.stats.free_reserved: 5388
vm.domain.2.stats.free_target: 87144
vm.domain.1.stats.free_min: 25827
vm.domain.1.stats.free_reserved: 5388
vm.domain.1.stats.free_target: 87144
vm.domain.0.stats.free_min: 51281
vm.domain.0.stats.free_reserved: 10647
vm.domain.0.stats.free_target: 173183

Which brings me to another question: when a domain runs full, why
does the system start to pageout instead of utilizing QPI?

I for my part solved most of the problem by hacking ZFS UMA to do
round-robin. UMA can perfectly do that, it's just not documented
anywhere. But there is a patch suggestion from I think Mark Johnston?
that does practically the same.

cheerio,
PMc