Re: Load 0.20 on a freshly installed idle system
Date: Tue, 24 Jan 2023 19:10:14 UTC
On 23 January 2023 6:31:27 am AEDT, Stavros Filargyropoulos <stav@stav.cc> wrote: > On Fri, Jan 20, 2023 at 8:44 PM Ian Smith <smithi@nimnet.asn.au> > wrote: > > This has been bugging quite a few people for quite a few years: > > > > https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=173541 > Bug is now marked as fixed and closed but no reference to the commit > that fixed it or what the root cause was. > > I still see the problem in STABLE where it was supposedly fixed. Yes; I updated to 12.4-R yesterday (due anyway) assuming 'months ago' should have been before 12.4 left 12-STABLE, but see no change. I emailed gbe@ as well, in case he's buried in PR mail. > $ uname -a > FreeBSD nas 13.1-STABLE FreeBSD 13.1-STABLE #0 > stable/13-n253535-4a8af507ebe5: Fri Jan 20 03:21:10 UTC 2023 > root@releng3.nyi.freebsd.org:/usr/obj/usr/src/amd64.amd64/sys/GENERIC > amd64 > > last pid: 1545; load averages: 0.31, 0.11, 0.04 > Where do we go from here? I put quite some time exploring this back in 2016, as did Jeremy Chadwick, so I'd pretty much given up since 0.1 or 0.2 nowadays is better than 0.6 then, and noone could explain how my magic fix for 9.2-STABLE worked, let alone me. I now understand a bit better what Alexander Motin (mav@) was explaining in posts #9, #27 and now #46 and #48, and accept the trade-off between idle accuracy and performance (particularly minimal power consumption on laptops) that the current scheme provides. That said, it would be great if someone wanted to implement a parallel set of event callouts specifically for load measurement that added very little to interrupt rate, and thus power use. For example, using periodic rather than one-shot event timers shows low, even 0.0 LAs at idle, but by costing 2000 interrupts per second per CPU!, 8000/s on my i5, adding ~25% idle power consumption. cheers, Ian