Avoiding CPU speed scaling in modern Linux distributions. Running CPU at full speed Tips.

Arty KSystem Administrator
CERTIFIED EXPERT
Do it right. Do it optimal.
Published:
If you have a server on collocation with the super-fast CPU, that doesn't mean that you get it running at full power.

Here is a preamble. When doing inventory of Linux servers, that I'm administering, I've found that some of them are running on lower CPU speed, then they could. This can be easily checked with this command:

grep -E '^model name|^cpu MHz' /proc/cpuinfo

Open in new window


What you can see:

model name      : Intel(R) Core(TM) i7 CPU         920  @ 2.67GHz
                      cpu MHz         : 1600.000
                      ...

Open in new window

or
model name      : Intel(R) Core(TM)2 CPU 6600 @ 2.40GHz
                      cpu MHz          : 1596.000
                      ...

Open in new window

or even
model name      : Intel(R) Core(TM)2 CPU          4400  @ 2.00GHz
                      cpu MHz         : 1000.000
                      model name      : Intel(R) Core(TM)2 CPU          4400  @ 2.00GHz
                      cpu MHz         : 600.000

Open in new window


Oops, we are paying for 2 Core 2Ghz CPU that runs on 600Mhz on one core and 1000Mhz on another!!!

There will be other lines for all the CPUs/cores/threads, probably with the same values.

This feature is nice, if we are running workstation, but what I've noticed, we do have the same CPU throttling on Ubuntu Server 10.04 builds and on CentOS 5.3, 5.4 and 5.5 builds (thus on RedHat too).

After hours of digging google, I've found that:
- this problem is very common
- there are several bug reports about this issue
- this is not BIOS settings problem, because on dual boot systems, CPU runs at full speed on Windows
- there are no 100% working solutions or they are too difficult to find
- this is not a bug, but a 'feature' of the new kernels, it is implemented differently on 2.6.18 (CentOS) and 2.6.32 (Ubuntu).

Here is a tip how to disable it on running system:

1) Check that 'kondemand' thread is running, run as root:  "pgrep -lf ondemand"

the output should be like:
[root@boston07 ~]# uname -a
                      Linux boston07 2.6.18-164.6.1.el5 #1 SMP Tue Nov 3 16:18:27 EST 2009 i686 i686 i386 GNU/Linux
                      [root@boston07 ~]# pgrep -lf ondemand
                      1444 kondemand/0
                      1445 kondemand/1

Open in new window


2) Check that current cpu speed differs from the maximum:

[root@boston07 ~]# grep -E '^model name|^cpu MHz' /proc/cpuinfo
                      model name      : Intel(R) Core(TM)2 CPU          6600  @ 2.40GHz
                      cpu MHz         : 1596.000
                      model name      : Intel(R) Core(TM)2 CPU          6600  @ 2.40GHz
                      cpu MHz         : 1596.000

Open in new window


3) Change CPU governor from 'ondemand' to 'performance' for all CPUs/cores, run as root:
for CPUFREQ in /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor; do [ -f $CPUFREQ ] || continue; echo -n performance > $CPUFREQ; done

Open in new window


4) Check that your changes have been applied:

[root@boston07 ~]# grep -E '^model name|^cpu MHz' /proc/cpuinfo
                      model name      : Intel(R) Core(TM)2 CPU          6600  @ 2.40GHz
                      cpu MHz         : 2394.000
                      model name      : Intel(R) Core(TM)2 CPU          6600  @ 2.40GHz
                      cpu MHz         : 2394.000

Open in new window


5) If you are running 'cpuspeed', 'cpufreqd', 'powerd' or other daemons, that can control CPU stepping, just stop them, if you really need to run your system on 100% of the CPU speed.
On CentOS:
# service cpuspeed stop

Open in new window


6) On Linux 2.6.32 (On RedHat 6, and Oracle Unbreakable Linux 6) remove CPU scaling kernel modules:
# lsmod | grep ondemand
                      cpufreq_ondemand        8764  0 
                      freq_table              3751  2 cpufreq_ondemand,acpi_cpufreq
                      # rmmod cpufreq_ondemand acpi_cpufreq freq_table

Open in new window


Ensure that no 'kondemand' kernel threads are running:
# pgrep -lf kondemand
                      #

Open in new window


7) To make changes permanent (on reboot):

- On Ubuntu, modify /etc/init.d/ondemand script:

change this
echo -n ondemand > $CPUFREQ

Open in new window

to this:
echo -n performance > $CPUFREQ

Open in new window


OR ALTERNATIVELY just remove all references to ondemand from /etc/rc?.d/
rm -f /etc/rc?.d/S99ondemand

Open in new window

- On CentOS, just create a new script /etc/init.d/ondemand:
#! /bin/bash
                      #
                      # ondemand sets cpu govermor
                      #
                      # chkconfig: 2345 10 90
                      #
                      # description: Set the CPU Frequency Scaling governor to "performance"
                      #
                      ### BEGIN INIT INFO
                      # Provides: $ondemand
                      ### END INIT INFO
                      
                      PATH=/sbin:/usr/sbin:/bin:/usr/bin
                      
                      case "$1" in
                          start)
                              for CPUFREQ in /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
                              do
                                      [ -f $CPUFREQ ] || continue
                                      echo -n performance > $CPUFREQ
                              done
                              ;;
                          restart|reload|force-reload)
                              echo "Error: argument '$1' not supported" >&2
                              exit 3
                              ;;
                          stop)
                              ;;
                          *)
                              echo "Usage: $0 start|stop" >&2
                              exit 3
                              ;;
                      esac

Open in new window

then enable it:
chmod +x /etc/init.d/ondemand
                      chkconfig --add ondemand
                      service ondemand restart

Open in new window


I'm using 'ondemand' name of the script, this may be a little bit misleading (because really it is a 'performance'), but you may change it.

Here are some useful links (just FYI, don't try to install cpufreq-set, cpufreq-get or other utilities, they will also install themselves as daemons, that will control your cpu speed regardless of your cpu governor settings):

1.  http://www.redhat.com/docs/wp/performancetuning/powermanagement/cpufreq_governors.html

2. http://ego.randomwalk.in/blog/2008/04/04/cause-and-effect/


Regards,
Arty
5
35,633 Views
Arty KSystem Administrator
CERTIFIED EXPERT
Do it right. Do it optimal.

Comments (3)

Commented:
I'm not sure I understand the use case where you wish to maximize your electric bill.

With Ondemand, the CPU runs faster when it needs to, slower when it doesn't need to.  This results in lower electrical use and temperatures, hence longer CPU life and lower electric and cooling bills.

If you do the "grep ... cpuinfo" enough times you will see several different results (presuming that your system actually has some sort of load).  Better yet, add the cpu speed monitor to the panel so you can see it change as you do stuff.  Of course, if your system isn't doing anything then usually it will show the lowest speed, while if it is fully loaded, it will almost always show the maximum speed.  Your  "grep ... cpuinfo" simply takes a snapshot, which may or may not be representative ... it's simply a roll of the dice.

For most workloads, Ondemand represents a nice balance between performance and energy use..  Performance supposedly leans towards performance a little more, but I have never been able to see a meaningful difference.

Simply locking the CPU speed at the max only means that you run the idle loop as fast as possible, while increasing CPU power consumption, CPU temperature and fan speed, while reducing the life of your CPU.

Personally, I've never been that concerned about seeing the idle loop run as fast as possible.
Arty KSystem Administrator
CERTIFIED EXPERT
Top Expert 2007

Author

Commented:
When you rent a server in a datacenter, that doesn't matter how much it comsumes - you pay fixed monthly fee.

Sometimes this matters how FAST your EVERY request is processed regardless of CPU load average.

Here is an example.
There are ping lowering services, like smoothping.com,where every milisecond of packet processing matters (users run ssh, that connects to server, then they use socks tunnel, provided by openssh, that helps to reduce ping, say from 300ms to 150ms, that is very important in online games).
If your load average is below 0.2 (even Atom 1.6Ghz can manage 100 ssh users with 0.2 LA), then your CPU runs on, say 600Mhz instead of 1.6Ghz, that makes every user connection a bit slower (if you don't beleave me - you can run tests). Does that make sense?

Regards,
Arty

Arty,

Thank you for writing this article. We just bought a bunch of servers with the new Sandybridge chipset and dual E5-2670's. We were expecting much more then they were performing and were not sure why. I researched the KONDEMAND process that was constantly at the top of the task list and found this page. We found that under heavy load the CPU's were only running at half speed.

After following the directions the CPU's are now running at full speed and performing much better. To comment on jjmcd... Well there is a big difference between a home (desktop) user and a data center manager. Most customers pay a fixed rate for power in any data center.

Have a question about something in this article? You can receive help directly from the article author. Sign up for a free trial to get started.