[gentoo-amd64] complete crash with ondemand scheduler

[gentoo-amd64] complete crash with ondemand scheduler

[Raffaele BELARDI]

When I enable the ondemand scheduler on my Athlon64 3200+ I get
immediate crash: video fills up with blinking colored lines, X cannot be
killed, the box does not ping, I need to hard reset. Other than that the
box is a stable mythtv station.

The motherboard is an Asus M2NPV-VM. Cool'n'Quiet is enabled in the
BIOS. Vanilla kernel with no binary drivers (actually, initially I was
using the proprietary nvidia graphics driver, which I though was the
culprit, but switching to open source driver did not solve), running
almost no ~amd64.
On my other box, an Athlon64 3000+/Asus K8VSE with same kernel, ondemand
works perfectly.

The only 'strange' hw on the 3200+ is a DVB tuner, the Hauppauge
HVR1300. Tonight I will try to remove it and see if it is related with
this problem.

Has anybody had similar problems with ondemand scheduler?
Any hints on how to debug the problem would be greatly appreciated.
Neither syslog nor xorg.log report anything unusual.

thanks,

raffaele
-- 
gentoo-amd64@gentoo.org mailing list

Re: [gentoo-amd64] complete crash with ondemand scheduler

[Beso]

[-- Attachment #1: Type: text/plain, Size: 1742 bytes --]

as a scheduler for athlons i reccomend userspace with powernowd enabled. on
my pc i start cpufrequtils with userspace at boot, and then i load powernowd
at boot with the option -n (for options on powernowd see the
/etc/conf.d/powernowd file and infos). that cpu governor gives the best freq
scaling with niced processes and good interoperability.
to come to the problem: i've never experienced such problems, if not when
using ati-drivers in the xorg conf, but not reinstalling the package after
kernel rebuild. after rebuilding or upgrading you'll always have to
reinstall the ati-drivers....

2007/8/2, Raffaele BELARDI <raffaele.belardi@st.com>:
>
> When I enable the ondemand scheduler on my Athlon64 3200+ I get
> immediate crash: video fills up with blinking colored lines, X cannot be
> killed, the box does not ping, I need to hard reset. Other than that the
> box is a stable mythtv station.
>
> The motherboard is an Asus M2NPV-VM. Cool'n'Quiet is enabled in the
> BIOS. Vanilla kernel with no binary drivers (actually, initially I was
> using the proprietary nvidia graphics driver, which I though was the
> culprit, but switching to open source driver did not solve), running
> almost no ~amd64.
> On my other box, an Athlon64 3000+/Asus K8VSE with same kernel, ondemand
> works perfectly.
>
> The only 'strange' hw on the 3200+ is a DVB tuner, the Hauppauge
> HVR1300. Tonight I will try to remove it and see if it is related with
> this problem.
>
> Has anybody had similar problems with ondemand scheduler?
> Any hints on how to debug the problem would be greatly appreciated.
> Neither syslog nor xorg.log report anything unusual.
>
> thanks,
>
> raffaele
> --
> gentoo-amd64@gentoo.org mailing list
>
>


-- 
dott. ing. beso

[-- Attachment #2: Type: text/html, Size: 2119 bytes --]

Re: [gentoo-amd64] complete crash with ondemand scheduler

[Joshua Hoblitt]

[-- Attachment #1: Type: text/plain, Size: 1483 bytes --]

This is clearly a BUG().  I recommend that you try to switch to the
ondemand governor while looking at a virtual terminal without X running
to see if you can see any of the bug output.  Then file a kernel bug in
bugs.gentoo.org bugzilla along with `emerge --info` and your running
kernels .config.

-J

--
On Thu, Aug 02, 2007 at 01:43:12PM +0200, Raffaele BELARDI wrote:
> When I enable the ondemand scheduler on my Athlon64 3200+ I get
> immediate crash: video fills up with blinking colored lines, X cannot be
> killed, the box does not ping, I need to hard reset. Other than that the
> box is a stable mythtv station.
> 
> The motherboard is an Asus M2NPV-VM. Cool'n'Quiet is enabled in the
> BIOS. Vanilla kernel with no binary drivers (actually, initially I was
> using the proprietary nvidia graphics driver, which I though was the
> culprit, but switching to open source driver did not solve), running
> almost no ~amd64.
> On my other box, an Athlon64 3000+/Asus K8VSE with same kernel, ondemand
> works perfectly.
> 
> The only 'strange' hw on the 3200+ is a DVB tuner, the Hauppauge
> HVR1300. Tonight I will try to remove it and see if it is related with
> this problem.
> 
> Has anybody had similar problems with ondemand scheduler?
> Any hints on how to debug the problem would be greatly appreciated.
> Neither syslog nor xorg.log report anything unusual.
> 
> thanks,
> 
> raffaele
> -- 
> gentoo-amd64@gentoo.org mailing list
> 

[-- Attachment #2: Type: application/pgp-signature, Size: 189 bytes --]

[gentoo-amd64] Re: complete crash with ondemand scheduler

[Raffaele BELARDI]

It was not a hw problem, I tried to remove the Hauppauge and the crash
persisted.

Updating the motherboard BIOS solved it. According to the Asus
changelog, there was a "Patch DMI rebuilding issue." between my original
BIOS an the updated one.

Thanks to those who responded.

raffaele

Raffaele Belardi wrote:
> When I enable the ondemand scheduler on my Athlon64 3200+ I get
> immediate crash: video fills up with blinking colored lines, X cannot be
> killed, the box does not ping, I need to hard reset. Other than that the
> box is a stable mythtv station.
> 
> The motherboard is an Asus M2NPV-VM. Cool'n'Quiet is enabled in the
> BIOS. Vanilla kernel with no binary drivers (actually, initially I was
> using the proprietary nvidia graphics driver, which I though was the
> culprit, but switching to open source driver did not solve), running
> almost no ~amd64.
> On my other box, an Athlon64 3000+/Asus K8VSE with same kernel, ondemand
> works perfectly.
> 
> The only 'strange' hw on the 3200+ is a DVB tuner, the Hauppauge
> HVR1300. Tonight I will try to remove it and see if it is related with
> this problem.
> 
> Has anybody had similar problems with ondemand scheduler?
> Any hints on how to debug the problem would be greatly appreciated.
> Neither syslog nor xorg.log report anything unusual.
> 
> thanks,
> 
> raffaele
> 
-- 
gentoo-amd64@gentoo.org mailing list

Re: [gentoo-amd64] complete crash with ondemand scheduler

[Beso]

[-- Attachment #1: Type: text/plain, Size: 1100 bytes --]

don't start thinking that everything that happen and breaks your pc is a
bug... first you have to be sure about you hw compatibility, about your type
of packages that have installed, about the toolchain that you're using ecc
ecc... first i was thinking that that problem could be due to wrong
ati-driver compilation, but as you can see that was a acpi-bios problem.
so if you experience problems with acpi you should check if acpi is ok
following this wiki:
http://gentoo-wiki.com/HOWTO_Fix_Common_ACPI_Problems

and for amd users, i suggest not using the ondemand governor cause it uses
the cpu in a very bad way. use conservative one or userspace with powernowd
or powerthend (the last one on my systems gave a better efficiency).

2007/8/3, Joshua Hoblitt <jhoblitt@ifa.hawaii.edu>:
>
> This is clearly a BUG().  I recommend that you try to switch to the
> ondemand governor while looking at a virtual terminal without X running
> to see if you can see any of the bug output.  Then file a kernel bug in
> bugs.gentoo.org bugzilla along with `emerge --info` and your running
> kernels .config.
>
>

[-- Attachment #2: Type: text/html, Size: 1496 bytes --]

Re: [gentoo-amd64] complete crash with ondemand scheduler

[Raffaele BELARDI]

Beso,

that is an interesting observation, can you elaborate a bit?

thanks,

raffaele

Beso wrote:
> 
> and for amd users, i suggest not using the ondemand governor cause it
> uses the cpu in a very bad way. use conservative one or userspace with
> powernowd or powerthend (the last one on my systems gave a better
> efficiency).
> 

-- 
gentoo-amd64@gentoo.org mailing list

Re: [gentoo-amd64] complete crash with ondemand scheduler

[Beso]

[-- Attachment #1: Type: text/plain, Size: 3087 bytes --]

the ondemand governor steps the processor between the least step to the most
one. for example, i have a turion 64 with steps from 800mhz to 2ghz. the
ondemand governor would step from 800mhz directly to 2ghz when the cpu is
under load and then return to 800 mhz when the load drops down.
the conservative governor instead steps to the most intelligent frequency
when it is needed. on my processor it steps to 1200mhz, or to 1,6ghz or
1,8ghz based on the load of the processor, but sometimes it lags a bit
before going to upper steps. of course you can set it according to the wiki
and forum information in the best way for your system.
then there are powernowd and powerthend, which are based on userspace
governor.
i'm actually using the first one (which is in the gentoo main repository)
and i find quite good with it, that steps quite good and in a quite smooth
way. i'll look for some tweaking for it because there are still some things
i don't like (mainly when copying to a truecrypted volume).
the last one is another daemon which is outside the gentoo repository, so
you'll have to go to their site ( http://powerthend.scheissname.de/ ) and
add the overlay to portage or you can download the source and compile it
manualy. i reccomend adding the overlay so that you can have it in the world
db. this one is quite a good daemon but for my standards it jumps too much
the processor current speed but it is really good cause i don't even know of
processor lags when transitioning from one step to another if i don't take a
look at kima (a little kmenu applet that let you see thermal temp, core
temp, cpufreq speed, uptime, and proc speed, that normally is equal to the
cpufreqspeed). if you'd like to try it, try it first and if you think that
is steps too much turn to powernowd.
the only thing that you need to know is that this 2 daemons need the
userspace governor built in the kernel or loaded and set when they start.
for that reason i use cpufrequtils to set the governor to userspace and
powernowd/powerthend to actually set the cpu speed of my processor.
i've read that you had a new processor with cool&quiet so i think that you
should really use this tools to get the powernowd! technology working. maybe
on the wiki you could find something also about cool&quiet and on how to
tune it.
the last thing i would advise to use is the klaptopdaemon. it is a fantastic
kde daemon that is capable of controlling cpu steps, standby, suspend which
are quite great, at least for my laptop. i don't know if it works also on a
desktop, but if not there is kpowersaved which is quite good also.

2007/8/3, Raffaele BELARDI <raffaele.belardi@st.com>:
>
> Beso,
>
> that is an interesting observation, can you elaborate a bit?
>
> thanks,
>
> raffaele
>
> Beso wrote:
> >
> > and for amd users, i suggest not using the ondemand governor cause it
> > uses the cpu in a very bad way. use conservative one or userspace with
> > powernowd or powerthend (the last one on my systems gave a better
> > efficiency).
> >
>
> --
> gentoo-amd64@gentoo.org mailing list
>
>


-- 
dott. ing. beso

[-- Attachment #2: Type: text/html, Size: 3565 bytes --]

Re: [gentoo-amd64] complete crash with ondemand scheduler

[Joshua Hoblitt]

[-- Attachment #1: Type: text/plain, Size: 3978 bytes --]

Your statement was that the on-demand governor is "very bad" for amd
CPUs.  I don't see how you have qualified that statement in any way.
Yes, it behaves differently then the conservative governor but that
doesn't make it "bad", "harmful", or even inefficient with power usage.

I use the ondemand govern on a large number of production Opteron
machines with great results.  According to my empirical measurements
with an ammeter it saves an average of ~35w per socket in systems with
dual core CPUs (not per core).  As you can imagine, this is a pretty use
full savings for the 8-socket system that I have.

Cheers,

-J

--
On Fri, Aug 03, 2007 at 02:50:48PM +0200, Beso wrote:
> the ondemand governor steps the processor between the least step to the most
> one. for example, i have a turion 64 with steps from 800mhz to 2ghz. the
> ondemand governor would step from 800mhz directly to 2ghz when the cpu is
> under load and then return to 800 mhz when the load drops down.
> the conservative governor instead steps to the most intelligent frequency
> when it is needed. on my processor it steps to 1200mhz, or to 1,6ghz or
> 1,8ghz based on the load of the processor, but sometimes it lags a bit
> before going to upper steps. of course you can set it according to the wiki
> and forum information in the best way for your system.
> then there are powernowd and powerthend, which are based on userspace
> governor.
> i'm actually using the first one (which is in the gentoo main repository)
> and i find quite good with it, that steps quite good and in a quite smooth
> way. i'll look for some tweaking for it because there are still some things
> i don't like (mainly when copying to a truecrypted volume).
> the last one is another daemon which is outside the gentoo repository, so
> you'll have to go to their site ( http://powerthend.scheissname.de/ ) and
> add the overlay to portage or you can download the source and compile it
> manualy. i reccomend adding the overlay so that you can have it in the world
> db. this one is quite a good daemon but for my standards it jumps too much
> the processor current speed but it is really good cause i don't even know of
> processor lags when transitioning from one step to another if i don't take a
> look at kima (a little kmenu applet that let you see thermal temp, core
> temp, cpufreq speed, uptime, and proc speed, that normally is equal to the
> cpufreqspeed). if you'd like to try it, try it first and if you think that
> is steps too much turn to powernowd.
> the only thing that you need to know is that this 2 daemons need the
> userspace governor built in the kernel or loaded and set when they start.
> for that reason i use cpufrequtils to set the governor to userspace and
> powernowd/powerthend to actually set the cpu speed of my processor.
> i've read that you had a new processor with cool&quiet so i think that you
> should really use this tools to get the powernowd! technology working. maybe
> on the wiki you could find something also about cool&quiet and on how to
> tune it.
> the last thing i would advise to use is the klaptopdaemon. it is a fantastic
> kde daemon that is capable of controlling cpu steps, standby, suspend which
> are quite great, at least for my laptop. i don't know if it works also on a
> desktop, but if not there is kpowersaved which is quite good also.
> 
> 2007/8/3, Raffaele BELARDI <raffaele.belardi@st.com>:
> >
> > Beso,
> >
> > that is an interesting observation, can you elaborate a bit?
> >
> > thanks,
> >
> > raffaele
> >
> > Beso wrote:
> > >
> > > and for amd users, i suggest not using the ondemand governor cause it
> > > uses the cpu in a very bad way. use conservative one or userspace with
> > > powernowd or powerthend (the last one on my systems gave a better
> > > efficiency).
> > >
> >
> > --
> > gentoo-amd64@gentoo.org mailing list
> >
> >
> 
> 
> -- 
> dott. ing. beso

[-- Attachment #2: Type: application/pgp-signature, Size: 189 bytes --]

Re: [gentoo-amd64] complete crash with ondemand scheduler

[Bernhard Auzinger]

> the ondemand governor steps the processor between the least step to the
> most one. for example, i have a turion 64 with steps from 800mhz to 2ghz.
> the ondemand governor would step from 800mhz directly to 2ghz when the cpu
> is under load and then return to 800 mhz when the load drops down.

I think you are wrong, because the ondemand governor does also switch between 
all available frequencies. For example my amd64X2 is capable of 
1000,1800,2000,2200,2400MHz and the ondemand governor does switch very well 
between these frequencies if needed.

Rgds
Bernhard
-- 
gentoo-amd64@gentoo.org mailing list

Re: [gentoo-amd64] complete crash with ondemand scheduler

[Joshua Hoblitt]

[-- Attachment #1: Type: text/plain, Size: 780 bytes --]

On Sun, Aug 05, 2007 at 09:24:17AM +0200, Bernhard Auzinger wrote:
> > the ondemand governor steps the processor between the least step to the
> > most one. for example, i have a turion 64 with steps from 800mhz to 2ghz.
> > the ondemand governor would step from 800mhz directly to 2ghz when the cpu
> > is under load and then return to 800 mhz when the load drops down.
> 
> I think you are wrong, because the ondemand governor does also switch between 
> all available frequencies. For example my amd64X2 is capable of 
> 1000,1800,2000,2200,2400MHz and the ondemand governor does switch very well 
> between these frequencies if needed.

That is correct, it moves one "step" per polling cycle the rate of which
can be adjusted via the sysfs interface.

-J

--

[-- Attachment #2: Type: application/pgp-signature, Size: 189 bytes --]

Re: [gentoo-amd64] complete crash with ondemand scheduler

[Beso]

[-- Attachment #1: Type: text/plain, Size: 1328 bytes --]

2007/8/5, Joshua Hoblitt <jhoblitt@ifa.hawaii.edu>:
>
> Your statement was that the on-demand governor is "very bad" for amd
> CPUs.  I don't see how you have qualified that statement in any way.
> Yes, it behaves differently then the conservative governor but that
> doesn't make it "bad", "harmful", or even inefficient with power usage.


i  never intended that it was harmful or bad in the way you think. i was
only saying that it is a bad governor for an amd processor. or at least it
is for my turion 64 and for my athlon x2. when i used it on these processors
i had a bad feeling, cpu lags about it. so i switched to the conservative
wich is smoother.

I use the ondemand govern on a large number of production Opteron
> machines with great results.  According to my empirical measurements
> with an ammeter it saves an average of ~35w per socket in systems with
> dual core CPUs (not per core).  As you can imagine, this is a pretty use
> full savings for the 8-socket system that I have.
>
> Cheers,
>
> -J
>
> on my turion 64 it does not behave as you say, and just step from 800mhz
to 2ghz, and it doesn't save anything when compared to conservative and
userspace with powernowd/powerthend. and i assure you that the overall
system energy consumption is of very great importance to me, specially when
i'm on battery.

[-- Attachment #2: Type: text/html, Size: 1826 bytes --]

Re: [gentoo-amd64] complete crash with ondemand scheduler

[Beso]

[-- Attachment #1: Type: text/plain, Size: 1008 bytes --]

i don't want to have to adjust it. i want it to step by default as i need,
and if there are daemons and better governors (the conservative one is
better in my opionion) i don't really understand why i should something that
will bother me with tunings when there are things that don't need them and
which perform better?!

2007/8/5, Joshua Hoblitt <jhoblitt@ifa.hawaii.edu>:
>
> On Sun, Aug 05, 2007 at 09:24:17AM +0200, Bernhard Auzinger wrote:
>
> > I think you are wrong, because the ondemand governor does also switch
> between
> > all available frequencies. For example my amd64X2 is capable of
> > 1000,1800,2000,2200,2400MHz and the ondemand governor does switch very
> well
> > between these frequencies if needed.


i repeat: on my 2 systems it didn't worked in that way, so i don't want to
spend time bothering about tunings when there is really no need for that.

That is correct, it moves one "step" per polling cycle the rate of which
> can be adjusted via the sysfs interface.
>
> -J
>
> --
>
>

[-- Attachment #2: Type: text/html, Size: 1466 bytes --]

Re: [gentoo-amd64] complete crash with ondemand scheduler

[Joshua Hoblitt]

[-- Attachment #1: Type: text/plain, Size: 2476 bytes --]

On Sun, Aug 05, 2007 at 04:29:43PM +0200, Beso wrote:
> i don't want to have to adjust it. i want it to step by default as i need,
> and if there are daemons and better governors (the conservative one is
> better in my opionion) i don't really understand why i should something that
> will bother me with tunings when there are things that don't need them and
> which perform better?!

So you'd rather install and have to configure an additional userspace
daemon then have to adjusting a single parameter in sysfs?  I don't
understand the logic of that.   This is not to say that there aren't good
reasons for using a userspace governor but in the simple case of only
scaling the CPUs frequency in reason to load it is not required, not
necessary, and just adds an extra piece of software to your system that
will wake your CPU up from a deep C state (wasting power).

You have in no way explained what your criteria for "perform better",
quantified the claim of "very bad" performance or why this discussion is
even specific to amd64 CPUs.

I'm not interested in arguing with you about this.  I just don't want to
leave your false claims unchallenged for the others reading this list as
power management is becoming a serious issue that all system
administrators need to deal with.  That last thing we need is FUD
floating around.

> 
> 2007/8/5, Joshua Hoblitt <jhoblitt@ifa.hawaii.edu>:
> >
> > On Sun, Aug 05, 2007 at 09:24:17AM +0200, Bernhard Auzinger wrote:
> >
> > > I think you are wrong, because the ondemand governor does also switch
> > between
> > > all available frequencies. For example my amd64X2 is capable of
> > > 1000,1800,2000,2200,2400MHz and the ondemand governor does switch very
> > well
> > > between these frequencies if needed.
> 
> 
> i repeat: on my 2 systems it didn't worked in that way, so i don't want to
> spend time bothering about tunings when there is really no need for that.

I suggest that you try reading the documentation for the ondemand
governor.  You can find it in the kernel sources as:

	Documentation/cpu-freq/governors.txt 

If you don't believe the documentation (or me), why don't you just try
adjusting the ondemand governors polling interval?

	echo ondemand > /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor
	cat /sys/devices/system/cpu/cpu0/cpufreq/ondemand/sampling_rate_max >
/sys/devices/system/cpu/cpu0/cpufreq/ondemand/sampling_rate

Cheers,

-J

--

[-- Attachment #2: Type: application/pgp-signature, Size: 189 bytes --]