Power Management GuideCapacity and lifetime of laptop batteries have improved much in the last years. Nevertheless modern processors consume much more energy than older ones and each laptop generation introduces more devices hungry for energy. That's why Power Management is more important than ever. Increasing battery run time doesn't necessarily mean buying another battery. Much can be achieved applying intelligent Power Management policies. Please notice that this guide describes Power Management for laptops. While some sections might also suite for servers, others do not and may even cause harm. Please do not apply anything from this guide to a server unless you really know what you are doing. As this guide has become rather long, here's a short overview helping you to find your way through it. The Prerequisites chapter talks about some requirements that should be met before any of the following device individual sections will work. This includes BIOS settings, kernel configuration and some simplifications in user land. The following three chapters focus on devices that typically consume most energy - processor, display and hard drive. Each can be configured seperately. CPU Power Management shows how to adjust the processor's frequency to save a maximum of energy without losing too much performance. A few different tricks prevent your hard drive from working unnecessarily often in Disk Power Management (decreasing noise level as a nice side effect). Some notes on graphics cards, Wireless LAN and USB finish the device section in Power Management For Other Devices while another chapter is dedicated to the (rather experimental) sleep states. Last not least Troubleshooting lists common pitfalls. Power Budget For Each Component
Nearly every component can operate in different states - off, sleep, idle, active to name a few - consuming a different amount of energy. Major parts are consumed by the LCD display, CPU, chipset and hard drives. Often one is able to activate OS-independent Power Management in the BIOS, but an intelligent setup in the operating system adapting to different situations can achieve much more. Before discussing the details of making individual devices Power Management aware, make sure certain requirements are met. After controlling BIOS settings, some kernel options want to be enabled - these are in short ACPI, sleep states and CPU frequency scaling. As power saving most of the time comes along with performance loss or increased latency, it should only be enabled when running on batteries. That's where a new runlevel battery comes in handy. First have a look into your BIOS Power Management settings. The best way is to combine BIOS and operating system policies, but for the moment it's better to disable most of the BIOS part. This makes sure it doesn't interfere with your policies. Don't forget to re-check BIOS settings after you configured everything else. Please check that the acpi USE flag is set in /etc/make.conf. Other USE flags that might be interesting for your system are apm, lm_sensors, nforce2, nvidia, pmu. See /usr/portage/profiles/use*.desc for details. If you forgot to set one of these flags, you can recompile affected packages using the --newuse flag in emerge, see man emerge. ACPI (Advanced Configuration and Power Interface) support in the kernel is still work in progress. Using a recent kernel will make sure you'll get the most out of it. There are different kernel sources in Portage. I'd recommend using gentoo-sources or tuxonice-sources. The latter contains patches for TuxOnIce, see the chapter about sleep states for more details. When configuring the kernel, activate at least these options:
Decide yourself whether you want to enable Software Suspend, and Sleep States (see below). If you own an ASUS, Medion, IBM Thinkpad or Toshiba laptop, enable the appropriate section. The kernel has to know how to enable CPU frequency scaling on your processor. As each type of CPU has a different interface, you've got to choose the right driver for your processor. Be careful here - enabling Intel Pentium 4 clock modulation on a Pentium M system will lead to strange results for example. Consult the kernel documentation if you're unsure which one to take. Compile your kernel, make sure the right modules get loaded at startup and boot into your new ACPI-enabled kernel. Next run emerge sys-power/acpid to get the acpi daemon. This one informs you about events like switching from AC to battery or closing the lid. Make sure the modules are loaded if you didn't compile them into the kernel and start acpid by executing /etc/init.d/acpid start. Run rc-update add acpid default to load it on startup. You'll soon see how to use it.
The default policy will be to enable Power Management only when needed - running on batteries. To make the switch between AC and battery convenient, create a runlevel battery that holds all the scripts starting and stopping Power Management.
Finished. Your new runlevel battery contains everything like default, but there is no automatic switch between both yet. Time to change it. Typical ACPI events are closing the lid, changing the power source or pressing the sleep button. An important event is changing the power source, which should cause a runlevel switch. A small script will take care of it. First you need a script which changes the runlevel to default respectively battery depending on the power source. The script uses the on_ac_power command from sys-power/powermgmt-base - make sure the package is installed on your system.
You are now able to determine the power source by executing on_ac_power && echo AC available || echo Running on batteries in a shell. The script below is responsible for changing runlevels. Save it as /etc/acpi/actions/pmg_switch_runlevel.sh.
Dont forget to run chmod +x /etc/acpi/actions/pmg_switch_runlevel.sh to make the script executable. The last thing that needs to be done is calling the script whenever the power source changes. That's done by catching ACPI events with the help of acpid. First you need to know which events are generated when the power source changes. The events are called ac_adapter and battery on most laptops, but it might be different on yours.
Run the command above and pull the power cable. You should see something like this:
The interesting part is the quoted string after received event. It will be matched by the event line in the files you are going to create below. Don't worry if your system generates multiple events or always the same. As long as any event is generated, runlevel changing will work.
Finally acpid has to be restarted to recognize the changes.
Give it a try: Plug AC in and out and watch syslog for the "Switching to AC mode" or "Switching to battery mode" messages. See the Troubleshooting section if the script is not able to detect the power source correctly. Due to the nature of the event mechanism, your laptop will boot into runlevel default regardless of the AC/battery state. This is fine when running from AC, but we'd like to boot into the battery runlevel otherwise. One solution would be to add another entry to the boot loader with the parameter softlevel=battery, but it's likely to forget choosing it. A better way is faking an ACPI event in the end of the boot process and letting pmg_switch_runlevel.sh script decide whether a runlevel change is necessary. Open /etc/conf.d/local.start in your favourite editor and add these lines:
Prepared like this you can activate Power Management policies for individual devices. Mobile processors can operate at different frequencies. Some allow changing voltage as well. Most of the time your CPU doesn't need to run at full speed and scaling it down will save much energy - often without any performance decrease. CPU frequency scaling brings up some technical terms that might be unknown to you. Here's a quick introduction. First of all, the kernel has to be able to change the processor's frequency. The CPUfreq processor driver knows the commands to do it on your CPU. Thus it's important to choose the right one in your kernel. You should already have done it above. Once the kernel knows how to change frequencies, it has to know which frequency it should set. This is done according to the policy which consists of a CPUfreq policy and a governor. A CPUfreq policy are just two numbers which define a range the frequency has to stay between - minimal and maximal frequency. The governor now decides which of the available frequencies in between minimal and maximal frequency to choose. For example, the powersave governor always chooses the lowest frequency available, the performance governor the highest one. The userspace governor makes no decision but chooses whatever the user (or a program in userspace) wants - which means it reads the frequency from /sys/devices/system/cpu/cpu0/cpufreq/scaling_setspeed. This doesn't sound like dynamic frequency changes yet and in fact it isn't. Dynamics however can be accomplished with various approaches. For example, the ondemand governor makes its decisions depending on the current CPU load. The same is done by various userland tools like cpudyn, cpufreqd, powernowd and many more. ACPI events can be used to enable or disable dynamic frequency changes depending on power source. Setting The Frequency Manually Decreasing CPU speed and voltage has two advantages: On the one hand less energy is consumed, on the other hand there is thermal improvement as your system doesn't get as hot as running on full speed. The main disadvantage is obviously the loss of performance. Decreasing processor speed is a trade off between performance loss and energy saving.
It's time to test whether CPU frequency changing works. Let's install another tool which is very handy for debugging purposes: sys-power/cpufrequtils
Here is an example output:
Now play around with cpufreq-set to make sure frequency switching works. Run cpufreq-set -g ondemand for example to activate the ondemand governor and verify the change with cpufreq-info. If it doesn't work as expected, you might find help in the Troubleshooting section in the end of this guide. The above is quite nice, but not doable in daily life. Better let your system set the appropriate frequency automatically. There are many different approaches to do this. The following table gives a quick overview to help you decide on one of them. It's roughly separated in three categories kernel for approaches that only need kernel support, daemon for programs that run in the background and graphical for programs that provide a GUI for easy configuration and changes.
While adjusting the frequency to the current load looks simple at a first glance, it's not such a trivial task. A bad algorithm can cause switching between two frequencies all the time or wasting energy when setting frequency to an unnecessary high level. Which one to choose? If you have no idea about it, try cpufreqd:
cpufreqd can be configured by editing /etc/cpufreqd.conf. The default one that ships with cpufreqd may look a bit confusing. I recommend replacing it with the one from former Gentoo developer Henrik Brix Andersen (see below). Please notice that you need cpufreqd-2.0.0 or later. Earlier versions have a different syntax for the config file.
Now you can start the cpufreqd daemon. Add it to the default and battery runlevel as well.
Sometimes it can be desirable to select another policy than the daemon chooses, for example when battery power is low, but you know that AC will be available soon. In that case you can turn on cpufreqd's manual mode with cpufreqd-set manual and select one of your configured policies (as listed by cpufreqd-get). You can leave manual mode by executing cpufreqd-set dynamic.
The last thing to check is that your new policies do a good job. An easy way to do so is monitoring CPU speed while working with your laptop:
If /proc/cpuinfo doesn't get updated (see Troubleshooting), monitor the CPU frequency with sys-apps/x86info:
Depending on your setup, CPU speed should increase on heavy load, decrease on no activity or just stay at the same level. When using cpufreqd and verbosity set to 5 or higher in cpufreqd.conf you'll get additional information about what's happening reported to syslog. As you can see in figure 1.1, the LCD display consumes the biggest part of energy (might not be the case for non-mobile CPU's). Thus it's quite important not only to shut the display off when not needed, but also to reduce it's backlight if possible. Most laptops offer the possibility to control the backlight dimming. The first thing to check is the standby/suspend/off timings of the display. As this depends heavily on your windowmanager, I'll let you figure it out yourself. Just two common places: Blanking the terminal can be done with setterm -blank <number-of-minutesM>, setterm -powersave on and setterm -powerdown <number-of-minutesM>. For X.org, modify /etc/X11/xorg.conf similar to this:
Probably more important is the backlight dimming. If you have access to the dimming settings via a tool, write a small script that dims the backlight in battery mode and place it in your battery runlevel. The following script should work on most IBM Thinkpads and Toshiba laptops. You've got to enable the appropriate option in your kernel (IBM Thinkpads only). For Toshiba laptops, install sys-power/acpitool and skip configuration of thinkpad_acpi (formerly called ibm_acpi) as described below.
To be able to set the brightness level, the thinkpad_acpi module has to be loaded with the experimental parameter.
This should work without error messages and a file /proc/acpi/ibm/brightness should be created after loading the module. An init script will take care of choosing the brightness according to the power source.
When done, make sure brightness is adjusted automatically by adding it to the battery runlevel.
Hard disks consume less energy in sleep mode. Therefore it makes sense to activate power saving features whenever the hard disk is not used for a certain amount of time. I'll show you two alternative possibilities to do it. First, laptop-mode will save most energy due to several measures which prevent or at least delay write accesses. The drawback is that due to the delayed write accesses a power outage or kernel crash will be more dangerous for data loss. If you don't like this, you have to make sure that there are no processes which write to your hard disk frequently. Afterwards you can enable power saving features of your hard disk with hdparm as the second alternative. Increasing idle time - laptop-mode Recent 2.6 kernels include the so-called laptop-mode. When activated, dirty buffers are written to disk on read calls or after 10 minutes (instead of 30 seconds). This minimizes the time the hard disk needs to be spun up.
laptop-mode-tools has its configuration file in /etc/laptop-mode/laptop-mode.conf. Adjust it the way you like it, it's well commented. Run rc-update add laptop_mode battery to start it automatically. Recent versions (1.11 and later) of laptop-mode-tools include a new tool lm-profiler. It will monitor your system's disk usage and running network services and suggests to disable unneeded ones. You can either disable them through laptop-mode-tools builtin runlevel support (which will be reverted by Gentoo's /sbin/rc) or use your default/battery runlevels (recommended).
After profiling your system for ten minutes, lm-profiler will present a list of services which might have caused disk accesses during that time.
To disable atd as suggested in the example above, you would run rc-update del atd battery. Be careful not to disable services that are needed for your system to run properly - lm-profiler is likely to generate some false positives. Do not disable a service if you are unsure whether it's needed. If you don't want to use laptop-mode, you must take special care to disable services that write to your disk frequently - syslogd is a good candidate, for example. You probably don't want to shut it down completely, but it's possible to modify the config file so that "unnecessary" things don't get logged and thus don't create disk traffic. Cups writes to disk periodically, so consider shutting it down and only enable it manually when needed.
You can also use lm-profiler from laptop-mode-tools (see above) to find services to disable. Once you eliminated all of them, go on with configuring hdparm. The second possibility is using hdparm. Skip this if you are using laptop-mode. Otherwise, edit /etc/conf.d/hdparm and add the following values to your drive entries. This example assumes your hard drive is called hda:
This will activate power management for your hard drive. If you ever want to deactivate power management, you can edit /etc/conf.d/hdparm and change the values to -q -S0, or just run hdparm -q -S0 /dev/hda. See man hdparm for the options. Though you can always start hdparm manually when you are on battery power by running /etc/init.d/hdparm start, it's much easier to automate its startup and shutdown. To do so, add hdparm to the battery runlevel so that it will automatically enable power management.
Another possibility is to deactivate swap in battery mode. Before writing a swapon/swapoff switcher, make sure there is enough RAM and swap isn't used heavily, otherwise you'll be in big problems. If you don't want to use laptop-mode, it's still possible to minimize disk access by mounting certain directories as tmpfs - write accesses are not stored on a disk, but in main memory and get lost with unmounting. Often it's useful to mount /tmp like this - you don't have to pay special attention as it gets cleared on every reboot regardless whether it was mounted on disk or in RAM. Just make sure you have enough RAM and no program (like a download client or compress utility) needs extraordinary much space in /tmp. To activate this, enable tmpfs support in your kernel and add a line to /etc/fstab like this:
6. Power Management For Other Devices In case you own an ATI graphics card supporting PowerPlay (dynamic clock scaling for the graphics processing unit GPU), you can activate this feature in X.org. Open /etc/X11/xorg.conf and add (or enable) the DynamicClocks option in the Device section. Please notice that this feature will lead to crashes on some systems.
Wireless LAN cards consume quite a bit of energy. Put them in Power Management mode just like your hard drives.
Add the following option to /etc/conf.d/net to automatically enable power management for your wireless card:
See man iwconfig for details and more options like the period between wakeups or timeout settings. If your driver and access point support changing the beacon time, this is a good starting point to save even more energy. There are two problems with USB devices regarding energy consumption: First, devices like USB mice, digital cameras or USB sticks consume energy while plugged in. You cannot avoid this (nevertheless remove them in case they're not needed). Second, when there are USB devices plugged in, the USB host controller periodically accesses the bus which in turn prevents the CPU from going into sleep mode. The kernel offers an experimental option to enable suspension of USB devices through driver calls or one of the power/state files in /sys.
7. Sleep States: sleep, standby, and suspend to disk ACPI defines different sleep states. The more important ones are
They can be called whenever the system is not in use, but a shutdown is not wanted due to the long boot time. The ACPI support for these sleep states is marked experimental for good reason. APM sleep states seem to be more stable, however you can't use APM and ACPI together.
Once your kernel is properly configured, you can use the hibernate-script to activate suspend or sleep mode. Let's install that first.
Some configuration has to be done in /etc/hibernate. The default package introduces a few configuration files for each sleep state. Options that are common to all suspend methods are placed in common.conf; make sure this file is properly set up for your system. To configure sleep, edit sysfs-ram.conf in /etc/hibernate. UseSysfsPowerState mem is already setup correctly, but if you need to make further changes to this particular sleep state (or any other sleep state) you should add them to /etc/hibernate/hibernate.conf. The comments and option names will guide you. If you use nfs or samba shares over the network, make sure to shutdown the appropriate init scripts to avoid timeouts.
Ready? Now is the last chance to backup any data you want to keep after executing the next command. Notice that you probably have to hit a special key like Fn to resume from sleep.
If you're still reading, it seems to work. You can also setup standby (S1) in a similar way by editing sysfs-ram.conf and changing "UseSysfsPowerState mem" to "UseSysfsPowerState standby". S3 and S4 are the more interesting sleep states due to greater energy savings however. This section introduces hibernation, where a snapshot of the running system is written to disk before powering off. On resume, the snapshot is loaded and you can go on working at exactly the point you called hibernate before.
There are two different implementations for S4. The original one is swsusp, then there is the newer tuxonice (former suspend2) with a nicer interface (including fbsplash support). A feature comparison is available at the tuxonice Homepage. There used to be Suspend-to-Disk (pmdisk), a fork of swsusp, but it has been merged back. TuxOnIce is not included in the mainline kernel yet, therefore you either have to patch your kernel sources with the patches provided by tuxonice.net or use sys-kernel/tuxonice-sources. The kernel part for both swusp and TuxOnIce is as follows:
The configuration for swsusp is rather easy. If you didn't store the location of your swap partition in the kernel config, you can also pass it as a parameter with the resume=/dev/SWAP directive. If booting is not possible due to a broken image, use the noresume kernel parameter. The hibernate-cleanup init script invalidates swsusp images during the boot process.
To activate hibernate with swsusp, use the hibernate script and set UseSysfsPowerState disk in /etc/hibernate/sysfs-disk.
If you experience kernel panics due to uhci or similar, try to compile USB support as module and unload the modules before sending your laptop to sleep mode. There are configuration options for this in common.conf
The following section discusses the setup of TuxOnIce including fbsplash support for a nice graphical progress bar during suspend and resume. The first part of the configuration is similar to the configuration of swsusp. In case you didn't store the location of your swap partition in the kernel config, you have to pass it as a kernel parameter with the resume=swap:/dev/SWAP directive. If booting is not possible due to a broken image, append the noresume parameter. Additionally, the hibernate-cleanup init script invalidates TuxOnIce images during the boot process.
Now edit /etc/hibernate/suspend2.conf, enable the TuxOnIce options you need. Do not enable the fbsplash options in common.conf just yet.
Please configure fbsplash now if you didn't do already. To enable fbsplash support during hibernation, the sys-apps/tuxonice-userui package is needed. Additionally, you've got to enable the fbsplash USE flag.
The ebuild tells you to make a symlink to the theme you want to use. For example, to use the livecd-2005.1 theme, run the following command:
If you don't want a black screen in the first part of the resume process, you have to add the tuxoniceui_fbsplash tool to your initrd image. Assuming you created the initrd image with splash_geninitramfs and saved it as /boot/fbsplash-emergence-1024x768, here's how to do that.
Afterwards adjust grub.conf (or lilo.conf) so that your TuxOnIce kernel uses /boot/fbsplash-tuxonice-emergence-1024x768 as initrd image. You can now test a dry run to see if everything is setup correctly.
Afterwards open /etc/hibernate/common.conf and activate the fbsplash options. Execute hibernate and enjoy. Q: I'm trying to change the CPU frequency, but /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor does not exist. A: Make sure your processor supports CPU frequency scaling and you chose the right CPUFreq driver for your processor. Here is a list of processors that are supported by cpufreq (kernel 2.6.7): ARM Integrator, ARM-SA1100, ARM-SA1110, AMD Elan - SC400, SC410, AMD mobile K6-2+, AMD mobile K6-3+, AMD mobile Duron, AMD mobile Athlon, AMD Opteron, AMD Athlon 64, Cyrix Media GXm, Intel mobile PIII and Intel mobile PIII-M on certain chipsets, Intel Pentium 4, Intel Xeon, Intel Pentium M (Centrino), National Semiconductors Geode GX, Transmeta Crusoe, VIA Cyrix 3 / C3, UltraSPARC-III, SuperH SH-3, SH-4, several "PowerBook" and "iBook2" and various processors on some ACPI 2.0-compatible systems (only if "ACPI Processor Performance States" are available to the ACPI/BIOS interface). Q: My laptop supports frequency scaling, but /sys/devices/system/cpu/cpu0/cpufreq/ is empty. A: Look for ACPI related error messages with dmesg | grep ACPI. Try to update the BIOS, especially if a broken DSDT is reported. You can also try to fix it yourself (which is beyond the scope of this guide). Q: My laptop supports frequency scaling, but according to /proc/cpuinfo the speed never changes. A: Probably you have activated symmetric multiprocessing support (CONFIG_SMP) in your kernel. Deactivate it and it should work. Some older kernels had a bug causing this. In that case, run emerge x86info, update your kernel as asked and check the current frequency with x86info -mhz. Q: I can change the CPU frequency, but the range is not as wide as in another OS. A: You can combine frequency scaling with ACPI throttling to get a lower minimum frequency. Notice that throttling doesn't save much energy and is mainly used for thermal management (keeping your laptop cool and quiet). You can read the current throttling state with cat /proc/acpi/processor/CPU/throttling and change it with echo -n "0:x" > /proc/acpi/processor/CPU/limit, where x is one of the Tx states listed in /proc/acpi/processor/CPU/throttling. Q: When configuring the kernel, powersave, performance and userspace governors show up, but that ondemand thing is missing. Where do I get it? A: The ondemand governor is only included in recent kernel sources. Try updating them. Q: Battery life time seems to be worse than before. A: Check your BIOS settings. Maybe you forgot to re-enable some of the settings. Q: My battery is charged, but KDE reports there would be 0% left and immediately shuts down. A: Check that battery support is compiled into your kernel. If you use it as a module, make sure the module is loaded. Q: My system logger reports things like "logger: ACPI group battery / action battery is not defined". A: This message is generated by the /etc/acpi/default.sh script that is shipped with acpid. You can safely ignore it. If you like to get rid of it, you can comment the appropriate line in /etc/acpi/default.sh as shown below:
Q: I have a Dell Inspiron 51XX and I don't get any ACPI events. A: This seems to be a kernel bug. Read on here. Q: I activated the DynamicClocks option in xorg.conf and now X.org crashes / the screen stays black / my laptop doesn't shutdown properly. A: This happens on some systems. You have to disable DynamicClocks. Q: I want to use TuxOnIce, but it tells me my swap partition is too small. Resizing is not an option. A: If there is enough free space on your system, you can use the filewriter instead of the swapwriter. The hibernate-script supports it as well. More information can be found in /usr/src/linux/Documentation/power/tuxonice.txt. Q: I just bought a brand new battery, but it only lasts for some minutes! What am I doing wrong? A: First follow your manufacturer's advice on how to charge the battery correctly. Q: The above didn't help. What should I do then? A: Some batteries sold as "new" are in fact old ones. Try the following:
If the "last full capacity" differs significantly from the design capacity, your battery is probably broken. Try to claim your warranty. Q: My problem is not listed above. Where should I go next? A: Don't fear to contact me, Dennis Nienhüser, directly. The Gentoo Forums are a good place to get help as well. If you prefer IRC, try the #gentoo-laptop channel at irc.freenode.net. The contents of this document are licensed under the Creative Commons - Attribution / Share Alike license. |
|
