12 Power management #

Modern processors have several power saving modes called C-states. They reflect the capability of an idle processor to turn off unused components to save power.

When a processor is in the C0 state, it is executing instructions. A processor running in any other C-state is idle. The higher the C number, the deeper the CPU sleep mode: more components are shut down to save power. Deeper sleep states can save large amounts of energy. Their downside is that they introduce latency. This means, it takes more time for the CPU to go back to C0. Depending on workload (threads waking up, triggering CPU usage and then going back to sleep again for a short period of time) and hardware (for example, interrupt activity of a network device), disabling the deepest sleep states can significantly increase overall performance. For details on how to do so, refer to Section 12.3.2, “Viewing kernel idle statistics with cpupower”.

Some states also have submodes with different power saving latency levels. Which C-states and submodes are supported depends on the respective processor. However, C1 is always available.

Table 12.1, “C-states” gives an overview of the most common C-states.

To avoid needless power consumption, it is recommended to test your workloads with deep sleep states enabled versus deep sleep states disabled. For more information, refer to Section 12.3.2, “Viewing kernel idle statistics with cpupower” or the cpupower-idle-set(1) man page.

While a processor operates (in C0 state), it can be in one of several CPU performance states (P-states). Whereas C-states are idle states (all but C0), P-states are operational states that relate to CPU frequency and voltage.

The higher the P-state, the lower the frequency and voltage at which the processor runs. The number of P-states is processor-specific and the implementation differs across the various types. However, P0 is always the highest-performance state (except for Section 12.1.3, “Turbo features”). Higher P-state numbers represent slower processor speeds and lower power consumption. For example, a processor in P3 state runs more slowly and uses less power than a processor running in the P1 state. To operate at any P-state, the processor must be in the C0 state, which means that it is working and not idling. The CPU P-states are also defined in the ACPI specification, see https://uefi.org/sites/default/files/resources/ACPI_Spec_6_5_Aug29.pdf.

C-states and P-states can vary independently of one another.

Turbo features allow to dynamically overtick active CPU cores while other cores are in deep sleep states. This increases the performance of active threads while still complying with Thermal Design Power (TDP) limits.

However, the conditions under which a CPU core can use turbo frequencies are architecture-specific. Learn how to evaluate the efficiency of those new features in Section 12.3, “The cpupower tools”.

The cpupower frequency-info command shows the statistics of the cpufreq driver used in the kernel. Additionally, it shows if turbo (boost) states are supported and enabled in the BIOS. Run without any options, it shows an output similar to the following:

# cpupower frequency-info
analyzing CPU 0:
  driver: intel_pstate
  CPUs which run at the same hardware frequency: 0
  CPUs which need to have their frequency coordinated by software: 0
  maximum transition latency: 0.97 ms.
  hardware limits: 1.20 GHz - 3.80 GHz
  available cpufreq governors: performance, powersave
  current policy: frequency should be within 1.20 GHz and 3.80 GHz.
                  The governor "powersave" may decide which speed to use
                  within this range.
  current CPU frequency is 3.40 GHz (asserted by call to hardware).
  boost state support:
    Supported: yes
    Active: yes
    3500 MHz max turbo 4 active cores
    3600 MHz max turbo 3 active cores
    3600 MHz max turbo 2 active cores
    3800 MHz max turbo 1 active cores

To get the current values for all CPUs, use cpupower -c all frequency-info.

The idle-info subcommand shows the statistics of the cpuidle driver used in the kernel. It works on all architectures that use the cpuidle kernel framework.

# cpupower idle-info
CPUidle driver: intel_idle
CPUidle governor: menu

Analyzing CPU 0:
Number of idle states: 6
Available idle states: POLL C1-SNB C1E-SNB C3-SNB C6-SNB C7-SNB
POLL:
Flags/Description: CPUIDLE CORE POLL IDLE
Latency: 0
Usage: 163128
Duration: 17585669
C1-SNB:
Flags/Description: MWAIT 0x00
Latency: 2
Usage: 16170005
Duration: 697658910
C1E-SNB:
Flags/Description: MWAIT 0x01
Latency: 10
Usage: 4421617
Duration: 757797385
C3-SNB:
Flags/Description: MWAIT 0x10
Latency: 80
Usage: 2135929
Duration: 735042875
C6-SNB:
Flags/Description: MWAIT 0x20
Latency: 104
Usage: 53268
Duration: 229366052
C7-SNB:
Flags/Description: MWAIT 0x30
Latency: 109
Usage: 62593595
Duration: 324631233978

After finding out which processor idle states are supported with cpupower idle-info, individual states can be disabled using the cpupower idle-set command. Typically one wants to disable the deepest sleep state, for example:

# cpupower idle-set -d 5

Or, for disabling all CPUs with latencies equal to or higher than 80:

# cpupower idle-set -D 80

Use the monitor subcommand to report processor topology, and monitor frequency and idle power state statistics over a certain period of time. The default interval is 1 second, but it can be changed with the -i. Independent processor sleep states and frequency counters are implemented in the tool—some retrieved from kernel statistics, others reading out hardware registers. The available monitors depend on the underlying hardware and the system. List them with cpupower monitor -l. For a description of the individual monitors, refer to the cpupower-monitor man page.

The monitor subcommand allows you to execute performance benchmarks. To compare kernel statistics with hardware statistics for specific workloads, concatenate the respective command, for example:

cpupower monitor db_test.sh

# cpupower monitor
|Mperf                   || Idle_Stats
 1                         2 
CPU | C0   | Cx   | Freq || POLL | C1   | C2   | C3
   0|  3.71| 96.29|  2833||  0.00|  0.00|  0.02| 96.32
   1| 100.0| -0.00|  2833||  0.00|  0.00|  0.00|  0.00
   2|  9.06| 90.94|  1983||  0.00|  7.69|  6.98| 76.45
   3|  7.43| 92.57|  2039||  0.00|  2.60| 12.62| 77.52

By comparing the values of the individual monitors, you can find correlations and dependencies and evaluate how well the power saving mechanism works for a certain workload. In Example 12.3 you can see that CPU 0 is idle (the value of Cx is near 100%), but runs at a very high frequency. This is because the CPUs 0 and 1 have the same frequency values which means that there is a dependency between them.

You can use cpupower frequency-set command as root to modify current settings. It allows you to set values for the minimum or maximum CPU frequency the governor may select or to create a new governor. With the -c option, you can also specify for which of the processors the settings should be modified. That makes it easy to use a consistent policy across all processors without adjusting the settings for each processor individually. For more details and the available options, see the man page cpupower-frequency-set or run cpupower frequency-set --help.

The CPUfreq subsystem offers several tuning options for P-states: You can switch between the different governors, influence minimum or maximum CPU frequency to be used or change individual governor parameters.

To switch to another governor at runtime, use cpupower frequency-set with the -g option. For example, running the following command (as root) will activate the performance governor:

# cpupower frequency-set -g performance

To set values for the minimum or maximum CPU frequency the governor may select, use the -d or -u option, respectively.