The performance governor for cpufreq doesn't seem to behave as advertised

mgoldshteyn · 08-23-2012, 07:58 PM

Quote:

Originally Posted by byannoni

Try:

Code:

sudo rmmod speedstep-lib
sudo modprobe p4_clockmod

If that doesn't work, do this:

Code:

sudo /bin/echo "p4_clockmod" >> /etc/modules
sudo reboot

OK, the first set of commands failed with the same error when trying to add p4_clockmod. I can't modify /etc/modules and reboot at this time, because there is other stuff hapening on the box. However, I can asure you that speedstep-lib is not loaded by default on boot. If I want it loaded, I have to manually load it using modprobe. I don't know what is interfering with the loading of p4_clockmod, but it is not speedstep-lib.

Current contents of /etc/modules:

Code:

loop
lp
rtc

Seems like stock Ubuntu server...

abrinister · 08-23-2012, 07:58 PM

Oh sorry about that didn't notice.

I've read somewhere that cpufreq isn't that accurate. I would recommend i7z.

Quote:

I don't know what is interfering with the loading of p4_clockmod, but it is not speedstep-lib.

Did you read the link I gave you? It actually recommends using the speedstep modules (which is outdated). Thus, you should use the acpi-cpufreq module, which includes support for Speedstep.

Also, do you have the cpufreq_performance module loaded?

Alex Brinister

mgoldshteyn · 08-23-2012, 08:03 PM

Quote:

Originally Posted by abrinister

Oh sorry about that didn't notice.

I've read somewhere that cpufreq isn't that accurate. I would recommend i7z.

Just installed i7z. Wow, nifty tool! Yep, just as I suspected, the frequencies on both CPUs are jumping all over the place. They are not fixed at 3.6 GHz. So much for the performance governor...

abrinister · 08-23-2012, 08:06 PM

Quote:

ERROR: modinfo: could not find module acpi-cpufreq

What does lsmod say about this?

Code:

lsmod | grep cpufreq

Alex Brinister

mgoldshteyn · 08-23-2012, 08:31 PM

Quote:

Originally Posted by abrinister

What does lsmod say about this?

Code:

lsmod | grep cpufreq

Alex Brinister

Nothing matches. It's almost as if it's loaded as a non-module driver or perhaps built right into the kernel if that makes any sense.

With regard to cpufreq_performance, what's that? I can't seem to find that via a google search as a valid module.

Also with regard to the speedstep-lib, I don't have that loaded by default, and it's been removed now thanks to the rmmod command I issued in trying to get the p4-clockmod to load (albeit unsuccessfully).

mgoldshteyn · 08-23-2012, 08:46 PM

Gentlemen, I ran i7x for a while and I think I may have a clue as to what is happening. It may be the case that the reason the frequency is being lowered from its top speed of 3.6 GHz is because the CPU is entering various C* states. The i7x tool is showing that this is happening, but I didn't know that C states can alter the CPU frequency. Or perhaps I7x (and cpuinfo_cur_freq to some extent) is computing a logical CPU frequency based on the fact that C states are being entered by the CPU cores.

Confirmed! C1E does mess with CPU core multipliers (and voltages for that matter), see: http://techreport.com/articles.x/7998/2

The question now is why would a cpufreq performance governor put the CPU into sleep states as deep as C6? That's poor design if you ask me, especially for a server build of Linux (i.e., Ubuntu Server 12.04 LTS).

I am going to try to disable C1E and any C states I can in the BIOS tomorrow and see what happens. I cannot disable SpeedStep, because the hardware we are using loses Turbo ability if you disable EIST. I know that this is whacked, but I have to live with the BIOS features I'm given by our hardware vendor.

byannoni · 08-23-2012, 08:56 PM

Quote:

Originally Posted by mgoldshteyn

With regard to cpufreq_performance, what's that? I can't seem to find that via a google search as a valid module.

Each cpufreq governor has a module that is named cpufreq_<governor_name>.

mgoldshteyn · 08-23-2012, 08:58 PM

Quote:

Originally Posted by byannoni

Each cpufreq governor has a module that is named cpufreq_<governor_name>.

And, in fact:

Code:

sudo modprobe cpufreq_performance

..., does succeed. But, sadly there is no change in behavior.

abrinister · 08-23-2012, 09:15 PM

Have you tried powersave? I don't know if this will do anything or not...

Alex Brinister

mgoldshteyn · 08-23-2012, 09:46 PM

Quote:

Originally Posted by abrinister

Have you tried powersave? I don't know if this will do anything or not...

Alex Brinister

I am unable to get the package for Ubuntu, for some reason (e.g., sudo apt-get install powersaved and also tried sudo apt-get install powersave, both errored with E: Unable to locate package).

So, I tried downloading and building powersave 0.15.20, but after installing all of the dependencies that it needs and finally successfully running configure, I get build errors that I cannot resolve:

Code:

...
make[2]: Entering directory `/home/mgoldshteyn/powersave-0.15.20/daemon'
g++ -DHAVE_CONFIG_H -I. -I..  -I../libpower -I/usr/include/dbus-1.0 -I/usr/lib/x86_64-linux-gnu/dbus-1.0/include   -I/usr/include/dbus-1.0 -I/usr/lib/x86_64-linux-gnu/dbus-1.0/include -I/usr/include/glib-2.0 -I/usr/lib/x86_64-linux-gnu/glib-2.0/include   -DBIN_DIR=\""/usr/local/bin"\" -DCONFIG_DIR=\""/usr/local/etc/powersave"\" -DSCRIPT_DIR=\""/usr/local/libexec/powersave"\" -DPUB_SCRIPT_DIR=\""/usr/local/libexec/powersave/scripts"\" -DKDE_BINDIR=\""/opt/kde3/bin"\" -DGNOME_BINDIR=\""/opt/gnome/bin"\"  -Wall  -DPOWERSAVE_DEBUG -DUSE_SYSLOG -g -O2 -MT acpi.o -MD -MP -MF .deps/acpi.Tpo -c -o acpi.o acpi.cpp
acpi.cpp: In member function âvirtual int ACPI_Interface::openHWEventFD()â:
acpi.cpp:85:46: error: no matching function for call to âstat::stat(const char*, stat*)â
acpi.cpp:85:46: note: candidates are:
/usr/include/x86_64-linux-gnu/bits/stat.h:46:8: note: stat::stat()
/usr/include/x86_64-linux-gnu/bits/stat.h:46:8: note:   candidate expects 0 arguments, 2 provided
/usr/include/x86_64-linux-gnu/bits/stat.h:46:8: note: stat::stat(const stat&)
/usr/include/x86_64-linux-gnu/bits/stat.h:46:8: note:   candidate expects 1 argument, 2 provided
acpi.cpp:89:33: error: âS_ISSOCKâ was not declared in this scope
acpi.cpp:95:39: error: âS_ISREGâ was not declared in this scope
acpi.cpp: In member function âvirtual int ACPI_Interface::handleHWEventRequest(int)â:
acpi.cpp:191:82: warning: deprecated conversion from string constant to âchar*â [-Wwrite-strings]
acpi.cpp:191:82: warning: deprecated conversion from string constant to âchar*â [-Wwrite-strings]
acpi.cpp:197:75: warning: deprecated conversion from string constant to âchar*â [-Wwrite-strings]
make[2]: *** [acpi.o] Error 1
make[2]: Leaving directory `/home/mgoldshteyn/powersave-0.15.20/daemon'
make[1]: *** [all-recursive] Error 1
make[1]: Leaving directory `/home/mgoldshteyn/powersave-0.15.20'
make: *** [all] Error 2

abrinister · 08-23-2012, 09:53 PM

Here's the deb file....

Alex Brinister

markseger · 08-24-2012, 08:49 AM

or look at cpu load and frequency together like this:

mjs@blkjack:~/hpcs$ collectl --import misc -sc
waiting for 1 second sample...
#<--------CPU--------><------Misc------>
#cpu sys inter ctxsw UTim MHz MT Log
3 1 2600 3358 0 1596 0 1
2 1 2443 3305 0 1596 0 1

-mark

mgoldshteyn · 08-24-2012, 08:55 AM

Quote:

Originally Posted by markseger

or look at cpu load and frequency together like this:

mjs@blkjack:~/hpcs$ collectl --import misc -sc
waiting for 1 second sample...
#<--------CPU--------><------Misc------>
#cpu sys inter ctxsw UTim MHz MT Log
3 1 2600 3358 0 1596 0 1
2 1 2443 3305 0 1596 0 1

-mark

On my hardware, collectl misreports the frequency (MHz) as 3601, even under powerschemes where the frequency is clearly moving between the low and high (with nothing running on the box).

markseger · 08-25-2012, 08:11 AM

Quote:

Originally Posted by mgoldshteyn

On my hardware, collectl misreports the frequency (MHz) as 3601, even under powerschemes where the frequency is clearly moving between the low and high (with nothing running on the box).

bummer. that means the contents of /proc/cpuinfo must be wrong. I've certainly seen/reported kernel problems in the past as they're not unknown.
-mark