Skip site navigation (1)Skip section navigation (2)
Date:      Wed, 2 Aug 2017 03:45:35 +1000 (EST)
From:      Ian Smith <smithi@nimnet.asn.au>
To:        Kevin Oberman <rkoberman@gmail.com>
Cc:        Daniel Braniss <danny@cs.huji.ac.il>, Karl Denninger <karl@denninger.net>,  stable@freebsd.org
Subject:   Re: issues with powerd/freq_levels
Message-ID:  <20170802004343.T6737@sola.nimnet.asn.au>
In-Reply-To: <CAN6yY1t5xYXMig0PSGn=Gdr=fOZJUBUYEGtXYxxJ9Q7R=z_kWw@mail.gmail.com>
References:  <8AEC9DBC-BADD-4FB2-8358-DA43F7EF5E68@cs.huji.ac.il> <20170731201323.A6737@sola.nimnet.asn.au> <CAN6yY1t5xYXMig0PSGn=Gdr=fOZJUBUYEGtXYxxJ9Q7R=z_kWw@mail.gmail.com>

next in thread | previous in thread | raw e-mail | index | archive | help
On Mon, 31 Jul 2017 12:03:27 -0700, Kevin Oberman wrote:
 > On Mon, Jul 31, 2017 at 3:48 AM, Ian Smith <smithi@nimnet.asn.au> wrote:
 > 
 > > On Mon, 31 Jul 2017 10:09:11 +0300, Daniel Braniss wrote:
 > >
 > >  > I am trying out PCengines latest apu2 boards, and I just noticed that
 > > with different Freebsd versions I get
 > >  > different freq_levels, and so when idling, each box (have 5) has a
 > > different freq/temperature value, ranging
 > >  > from 125/69.1C, 600/59.0C to 75/56.0C
 > >  >
 > >  > FreeBSD apu-4 11.1-STABLE FreeBSD 11.1-STABLE #5 f565b5a06ab3 (11) tip:
 > > Mon Jul 31 09:36:33 IDT 2017
 > >  > apu-4# sysctl dev.cpu.0.freq_levels
 > >  > dev.cpu.0.freq_levels: 1000/980 800/807 600/609
 > >
 > > That looks about right.  On a Core2Duo (still on 9.3) I get:
 > > dev.est.1.freq_settings: 2401/35000 2400/35000 1600/15000 800/12000
 > > dev.est.0.freq_settings: 2401/35000 2400/35000 1600/15000 800/12000
 > > dev.cpu.0.freq_levels: 2401/35000 2400/35000 1600/15000 800/12000
 > > dev.cpu.0.freq: 800
 > >
 > > But only because I'd added to /boot/loader.conf:
 > >
 > > hint.p4tcc.0.disabled=1
 > > hint.acpi_throttle.0.disabled=1
 > >
 > > which became the defaults sometime, maybe not before 11.0?  Otherwise
 > > mine would look more similar to the one below, with all 12.5% increments
 > > in frequency enabled, which doesn't actually save any power at all.
 > >
 > >  > FreeBSD apu-5 11.1-PRERELEASE FreeBSD 11.1-PRERELEASE #0 21e9d1ca9b80
 > > (11) tip: Tue May 30 11:51:48 IDT 2017
 > >  > apu-5# sysctl dev.cpu.0.freq_levels
 > >  > dev.cpu.0.freq_levels: 1000/966 875/845 800/795 700/695 600/600 525/525
 > > 450/450 375/375 300/300 225/225 150/150 75/75
 > >
 > > Looks like either p4tcc or acpi_throttle is enabled?  See cpufreq(4).
 > > As above, these don't buy you anything but extra busyness for powerd.
 > >
 > > Also noticed that the (nice, low!) milliwatt figures for 1000/800/600
 > > freqs are a bit different to the -stable one.  Slightly Different model?
 > >
 > >  > FreeBSD apu-1 10.3-STABLE FreeBSD 10.3-STABLE #4 267788fd852c (10) tip:
 > > Tue Jan 10 09:09:00 IST 2017
 > >  > apu-1# sysctl dev.cpu.0.freq_levels
 > >  > dev.cpu.0.freq_levels: 1000/-1 875/-1 750/-1 625/-1 500/-1 375/-1
 > > 250/-1 125/-1
 > >
 > > And that looks like est(4) isn't enabled/attaching at all .. see dmesg
 > > on all of these for clues.
 > >
 > >  > so, any ideas as to what is going on?
 > >
 > > Pure guesswork on experience with older versions, I'm not up to date.
 > >
 > 
 > Very odd. Are all systems running identical CPUs and BIOSes? Identical
 > loader and sysctl configurations? Look at /var/rn/dmesg.boot for CPU
 > information. Is EST being detected? It used to be early in the boot
 > process, but is now fairly late. (In my case, about 2/3 through the
 > dmesg.boot file.

Hi Kevin, it's been a while ..

Danny, can you put up a verbose boot dmesg.boot of one(?) for a browse? 
Or maybe apu-4 and -1, if not all.  I'd expect error msgs on -1 anyway.

 > I have p4tcc and throttling explicitly turned off (which should now be the
 > default), but my Sandy Bridge Core i5 still shows:
 > dev.cpu.0.freq_levels: 2501/35000 2500/35000 2000/26426 1800/23233
 > 1600/20164 1400/17226 1200/14408 1000/11713 800/9140

All truly available I see on more recent processors.  Certainly not 1/8 
duty-cycle multipliers as p4tcc and maybe? acpi_throttle (not seen here)

 > The first is really bogus to indicate "turbo" mode.

Usefully bogus, in that you can flag powerd to (in your case) -M 2500 to 
prevent it engaging "turbo" mode, as I do on my old Core2Duo, as advised 
by Warner years ago to avoid overheating on buildworlds and such - but 
more recent incarnations of "turbo" are supposedly far more functional.

Admittedly a digression .. mostly coming from wondering about data Karl
posted in response, indicating different Cx levels available and so used 
by the latter 3 AP cores, which was news to me.  I'd like to know more, 
if only for gratuitous curiosity.  Others can tick their TL;DR box :)

 > Temperature is a totally separate issue. It is VERY sensitive to external
 > issue like airflow and position of the CPU in relation to other components
 > in the chassis Also, unless you have a lot of cores, you probably should
 > set both economy_cx_lowest and performance_cx_lowest to Cmax. Economy
 > should default to that, but  performance will not as that can cause issues
 > on systems with large numbers of cores, so is set to C2. Many such system
 > used to disable deeper sleep modes in BIOS, but I am way behind the times
 > and don't know about the current state of affairs. Certainly for systems
 > with 32 or fewer cores, this should not be an issue. In any case, Cx state
 > can sharply impact temperature.

Indeed.  But as these are low-power devices already, it's likely less of 
a concern, but maximising efficiency and minimising stress never hurts.

 > Finally, the last case with power levels of -1 for all frequencies is
 > probably because the CPU manufacturer (Intel?) has not published this
 > information. For a while they were treating this as "proprietary"
 > information. Very annoying! It's always something that is not readily
 > available. Thi is one reason I suspect your CPUs are not identical.

Hmm, bought as a batch, that sounds unlikely, though their BIOSes (ono) 
may vary, and would be worth checking on each - and BIOS settings, too.

Danny, is powerd running on all these?  I doubt it would load on apu-1 
as it stands.  Note these are 'pure' 1/8 factors of 1000, p4tcc-alike, 
and I think quite likely indicate that cpufreq(4) failed to initialise? 
debug.cpufreq.verbose=1 in /boot/loader.conf might show a clue, with a 
verbose dmesg.boot anyway.

Later: oops, just reread Karl's message, where I was unfaniliar with 
different CPUs showing different C-states, and noticing that despite 
cpu0 showing C2(io) available, and cx_lowest as C2, yet it used 100% C1 
state, which was all that was available to cpu1 to 3.

But then I twigged to Karl's hwpstate errors, so with 'apropos hwpstate' 
still showing nothing after all these years, along with other cpufreq(4) 
drivers, I used the list search via duckduckgo to finally find one (1) 
message, which lead to one detailed thread (that I even bought into!)

 https://lists.freebsd.org/pipermail/freebsd-stable/2012-May/subject.html
 https://lists.freebsd.org/pipermail/freebsd-stable/2012-June/thread.html

/hwpstate  Note the May one needs following by Subject, else it splits 
into 5 separate threads (?)

Which may be interesting to cpufreq nerds, but had me remember that 
hwpstate(0) is for AMD not Intel CPUs.  So now I'm totally confused :)

Danny, do your results from Karl's sysctl listings agree with his?

cheers, Ian



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20170802004343.T6737>