Theo de Raadt - Wed, 11 May 2022 at 18:08:53
> f.holop <min...@obiit.org> wrote:
> 
> > Stuart Henderson - Mon, 09 May 2022 at 17:17:57
> > > Currently, you can either set it manually to low speed
> > > (hw.perfpolicy=manual, hw.setperf=0), modify the kernel (e.g. with the
> > > diff below), or use obsdfreqd from packages. The latter is only in
> > > -current packages not 7.1, but it could be built from ports.
> > 
> > I think the elephant in the room is:
> > will this change be reverted?
> > 
> > What is the rationale of not letting wall powered servers
> > throttle down?
> 
> As it is today the scheduler-based algorithm seriously sucks, and after
> the change it was discovered many machines were running 10-20% less than
> peak performance even under load.  This was discovered during
> suspend/hibernate/resume events but it affects all workloads.

a 3rd party package popping up immediately to "solve" this issue is a
good sign that it's not only about me me me. i don't know your workload,
you don't know mine.  some do not need "peak performance" workloads all
the time and actaully care more about power consumption on the long run
than a couple of percentage points lost in performance.

> It may get fixed after work in the scheduler which is not so damned
> pessimistic to give people 20% less machine than they bought, and
> this may happen as a resultt of pivoting to the opposite side of the problem
> space, so that the people who actually performance adjustments ensure the
> ramping up/down actually works without hurting PERFORMANCE.

this mailing list is littered with performance issues regarding openbsd
and it was always a clear message by the developers that it is something
important but clearly secondary after correctness, readability, etc.

it's great that this was discovered but openbsd could simply come with
those default settings and recommendations. let me choose my tradeoffs
please, that was the entire point of that sysctl.  the exact same effect
the hardcoding does was solvable with 2 lines of sysctl.conf for _those
who wanted it that way_.

this whole discussion is really bizarre, do you want to hardcode some
of my other sysctls as well in the next release?
apparently we are not adults who should have a choice.

i am really curious how this will affect your own power bill, you have
lots of iron at home, some quite power hungry i imagine.  summer is
coming maybe you will need more cooling too.  please publish some
numbers in a month or two if you dont mind the transparency.

-f
-- 

Reply via email to