CPU Frequency Scaling (cpufreq)

Introduction

CPU frequency scaling (cpufreq) is the Linux kernel subsystem that dynamically adjusts processor clock frequency and voltage to balance performance against power consumption. Modern processors support multiple Operating Performance Points (OPPs), also known as P-states in ACPI terminology, where each point represents a specific frequency-voltage pair. Higher frequencies provide more computational throughput but consume proportionally more power and generate more heat.

The cpufreq subsystem is part of the kernel’s power management framework. It provides a unified interface for different CPU architectures (x86, ARM, RISC-V) to implement frequency scaling through a set of “governors” that determine when and how to change the CPU frequency based on system load.

Architecture

flowchart TB
    subgraph Userspace["User Space"]
        UDEV["udev rules"]
        TOOLS["cpupower / turbostat"]
        GOVCONF["Governor configuration"]
    end
    
    subgraph Kernel["Kernel Space - cpufreq Subsystem"]
        subgraph Core["cpufreq Core"]
            CPUFREQ_CORE["cpufreq_driver"]
            POLICY["cpufreq_policy"]
            GOV["Governor Interface"]
        end
        
        subgraph Governors["Governors"]
            PERF["performance"]
            PS["powersave"]
            ODOC["ondemand"]
            CONS["conservative"]
            SCHED["schedutil"]
            US["userspace"]
        end
        
        subgraph Drivers["Hardware Drivers"]
            ACPI["acpi-cpufreq"]
            INTEL_PSTATE["intel_pstate"]
            AMD_PSTATE["amd-pstate"]
            CPPC["cppc_cpufreq"]
            DT["cpufreq-dt"]
        end
    end
    
    subgraph Hardware["Hardware"]
        CPU["CPU Cores"]
        HW_PSTATE["P-states / OPPs"]
        HWP["Hardware P-state Control"]
    end
    
    UDEV --> CPUFREQ_CORE
    TOOLS --> CPUFREQ_CORE
    GOVCONF --> GOV
    
    CPUFREQ_CORE --> POLICY
    POLICY --> GOV
    POLICY --> Drivers
    
    GOV --> Governors
    Governors --> GOV
    
    ACPI --> CPU
    INTEL_PSTATE --> HWP
    AMD_PSTATE --> HWP
    CPPC --> HWP
    DT --> HW_PSTATE
    
    CPU --> HW_PSTATE

Sysfs Interface

The cpufreq subsystem exposes its configuration through sysfs at /sys/devices/system/cpu/cpufreq/. Each CPU policy (which may cover multiple CPUs sharing the same clock) is represented as a policyN/ directory.

Global Interface

# List all cpufreq policies
ls /sys/devices/system/cpu/cpufreq/

# List available CPUs and their frequency info
cat /sys/devices/system/cpu/cpu*/cpufreq/scaling_cur_freq

# Kernel module parameters
cat /sys/module/cpufreq/parameters/off

Per-Policy Interface

# Current frequency (kHz)
cat /sys/devices/system/cpu/cpufreq/policy0/scaling_cur_freq

# Minimum and maximum frequency (kHz)
cat /sys/devices/system/cpu/cpufreq/policy0/scaling_min_freq
cat /sys/devices/system/cpu/cpufreq/policy0/scaling_max_freq

# Available frequencies (if hardware defines discrete steps)
cat /sys/devices/system/cpu/cpufreq/policy0/scaling_available_frequencies

# Available governors
cat /sys/devices/system/cpu/cpufreq/policy0/scaling_available_governors

# Current governor
cat /sys/devices/system/cpu/cpufreq/policy0/scaling_governor

# Governor-specific parameters
ls /sys/devices/system/cpu/cpufreq/policy0/

# Set minimum frequency (requires root)
echo 2000000 | sudo tee /sys/devices/system/cpu/cpufreq/policy0/scaling_min_freq

# Set maximum frequency
echo 3000000 | sudo tee /sys/devices/system/cpu/cpufreq/policy0/scaling_max_freq

# Change governor
echo performance | sudo tee /sys/devices/system/cpu/cpufreq/policy0/scaling_governor

# Hardware limits (read-only)
cat /sys/devices/system/cpu/cpufreq/policy0/cpuinfo_min_freq
cat /sys/devices/system/cpu/cpufreq/policy0/cpuinfo_max_freq
cat /sys/devices/system/cpu/cpufreq/policy0/cpuinfo_cur_freq
cat /sys/devices/system/cpu/cpufreq/policy0/cpuinfo_transition_latency

Related Power Management Files

# Intel P-state specific (if intel_pstate is active)
cat /sys/devices/system/cpu/intel_pstate/min_perf_pct
cat /sys/devices/system/cpu/intel_pstate/max_perf_pct
cat /sys/devices/system/cpu/intel_pstate/no_turbo
cat /sys/devices/system/cpu/intel_pstate/status  # active/passive/off

# AMD P-state specific (if amd-pstate is active)
cat /sys/devices/system/cpu/amd_pstate/status  # active/passive/off
cat /sys/devices/system/cpu/amd_pstate/epp     # energy performance preference

# Turbo boost (architecture-specific)
cat /sys/devices/system/cpu/intel_pstate/no_turbo  # 0=turbo on, 1=turbo off

Governors

Governors are the decision-making algorithms that determine the target CPU frequency based on system load and policy constraints.

performance

Always sets the CPU to the highest available frequency. Used for latency-sensitive workloads where maximum throughput is required.

# Set performance governor on all CPUs
for policy in /sys/devices/system/cpu/cpufreq/policy*/; do
    echo performance | sudo tee ${policy}scaling_governor
done

Characteristics:

• Zero latency for frequency transitions
• Maximum power consumption
• Best for benchmarks, real-time workloads, and HPC

powersave

Always sets the CPU to the lowest available frequency. Used to minimize power consumption at the cost of performance.

echo powersave | sudo tee /sys/devices/system/cpu/cpufreq/policy0/scaling_governor

Characteristics:

• Minimum power consumption
• Significant performance impact
• Best for idle or light-load servers

ondemand

Dynamically adjusts frequency based on CPU utilization. When load exceeds a threshold, the governor scales up to maximum frequency; when load drops, it gradually scales down.

echo ondemand | sudo tee /sys/devices/system/cpu/cpufreq/policy0/scaling_governor

# Configure ondemand parameters
cd /sys/devices/system/cpu/cpufreq/policy0/

# Sampling rate (microseconds) - how often to check load
echo 10000 | sudo tee sampling_rate

# Up threshold (%) - load percentage to trigger frequency increase
echo 80 | sudo tee up_threshold

# Down differential (%) - how much load must drop before scaling down
echo 5 | sudo tee down_differential

# Ignore nice load - don't count nice'd processes
echo 1 | sudo tee ignore_nice_load

# Sampling down factor - number of sampling intervals before scaling down
echo 10 | sudo tee sampling_down_factor

# Power save bias - bias toward lower frequencies on non-root CPUs
echo 0 | sudo tee powersave_bias

Characteristics:

• Good balance of performance and power savings
• Reacts quickly to load spikes
• May cause frequency oscillation under moderate load
• Legacy governor, replaced by schedutil in modern kernels

conservative

Similar to ondemand but uses more gradual frequency scaling. Instead of jumping to maximum frequency on high load, it increases frequency in steps.

echo conservative | sudo tee /sys/devices/system/cpu/cpufreq/policy0/scaling_governor

# Configure conservative parameters
cd /sys/devices/system/cpu/cpufreq/policy0/

echo 20 | sudo tee freq_step         # Percentage step for frequency increase
echo 40 | sudo tee down_threshold     # Load % to trigger frequency decrease
echo 80 | sudo tee up_threshold       # Load % to trigger frequency increase

Characteristics:

• Smoother frequency transitions than ondemand
• Better for battery-powered devices
• Slower to respond to load changes
• Less frequency oscillation

schedutil

The modern governor that integrates with the kernel’s CPU scheduler (CFS/EEVDF). It uses scheduler utilization signals to make frequency decisions, providing more accurate and timely frequency scaling.

echo schedutil | sudo tee /sys/devices/system/cpu/cpufreq/policy0/scaling_governor

# Parameters (usually not tunable, auto-configured based on scheduler)
# rate_limit_us - minimum time between frequency changes
cat /sys/devices/system/cpu/cpufreq/policy0/schedutil/rate_limit_us

Characteristics:

• Integrated with the scheduler for accurate load measurement
• Accounts for per-task utilization (not just aggregate CPU usage)
• Better latency response than ondemand
• Default governor on most modern distributions (kernel 4.7+)
• Works correctly with CPU idle states and turbo boost
• Supports per-CPU frequency selection based on scheduler hints

flowchart LR
    subgraph Scheduler["CPU Scheduler (EEVDF)"]
        UTIL["util_avg signal"]
        UCLAMP["uclamp (min/max)"]
    end
    
    subgraph Schedutil["schedutil Governor"]
        NEXT["next_freq = util / max * max_freq"]
        RL["rate_limit_us"]
    end
    
    subgraph HW["Hardware"]
        FREQ["Set CPU frequency"]
    end
    
    UTIL --> NEXT
    UCLAMP --> NEXT
    NEXT --> RL
    RL --> FREQ

userspace

Allows a user-space program to set the CPU frequency directly. The governor itself does not make any decisions; it only provides the mechanism for user-space control.

echo userspace | sudo tee /sys/devices/system/cpu/cpufreq/policy0/scaling_governor

# Set specific frequency
echo 2400000 | sudo tee /sys/devices/system/cpu/cpufreq/policy0/scaling_setspeed

Use cases:

• Custom power management daemons
• Testing and benchmarking specific frequencies
• Embedded systems with specific power budgets

Governor Comparison

Hardware Drivers

intel_pstate

The Intel P-state driver is the primary driver for modern Intel CPUs (Sandy Bridge and later). It has two modes:

Active mode (default on supported hardware):

• Implements its own performance and powersave governors
• Does not use the generic cpufreq governors
• Directly controls Hardware P-states (HWP) on supported CPUs

# Check if intel_pstate is active
cat /sys/devices/system/cpu/intel_pstate/status
# Output: active, passive, or off

# Active mode parameters
cat /sys/devices/system/cpu/intel_pstate/min_perf_pct   # 0-100
cat /sys/devices/system/cpu/intel_pstate/max_perf_pct   # 0-100
cat /sys/devices/system/cpu/intel_pstate/no_turbo       # 0 or 1
cat /sys/devices/system/cpu/intel_pstate/turbo_pct      # Turbo ratio

# Disable turbo boost
echo 1 | sudo tee /sys/devices/system/cpu/intel_pstate/no_turbo

# Limit max performance to 50%
echo 50 | sudo tee /sys/devices/system/cpu/intel_pstate/max_perf_pct

Passive mode:

• Acts as a regular cpufreq driver
• Compatible with generic governors (ondemand, schedutil, etc.)
• Used when hardware P-state control is not desired

# Force passive mode (use generic governors)
# Boot parameter: intel_pstate=passive

# Or at runtime (if supported)
echo passive | sudo tee /sys/devices/system/cpu/intel_pstate/status

acpi-cpufreq

The ACPI-based cpufreq driver, used as a fallback when intel_pstate is not active, or on older Intel/AMD processors.

# Check if acpi-cpufreq is loaded
lsmod | grep acpi_cpufreq

# Available frequencies
cat /sys/devices/system/cpu/cpufreq/policy0/scaling_available_frequencies
# Output: 1200000 1400000 1600000 1800000 2000000 2200000 2400000

# Transition latency
cat /sys/devices/system/cpu/cpufreq/policy0/cpuinfo_transition_latency
# Output: 10000 (nanoseconds)

amd-pstate

The AMD P-state driver for Zen 2 and later processors. Supports multiple modes:

# Check status
cat /sys/devices/system/cpu/amd_pstate/status

# Modes:
# - active: Hardware-managed (like intel_pstate active)
# - passive: Software-managed (like acpi-cpufreq)
# - guided: Hybrid - software hints, hardware final decision

# Energy Performance Preference (EPP)
# 0 = performance, 128 = balance_performance, 192 = balance_power, 254 = power
cat /sys/devices/system/cpu/amd_pstate/epp

# Set EPP for specific policy
echo balance_performance | sudo tee /sys/devices/system/cpu/cpufreq/policy0/energy_performance_preference

cpufreq-dt (Device Tree)

Used on ARM and RISC-V platforms where frequency information is provided via Device Tree.

# Device Tree frequency table (example)
cat /sys/devices/system/cpu/cpufreq/policy0/scaling_available_frequencies
# Output: 408000 600000 816000 1008000 1200000 1368000 1512000 1608000

# Transition latency varies by platform
cat /sys/devices/system/cpu/cpufreq/policy0/cpuinfo_transition_latency

Turbo Boost and Frequency Limits

Intel Turbo Boost / Turbo Boost Max 3.0

# Check if turbo boost is available
cat /sys/devices/system/cpu/intel_pstate/no_turbo
# 0 = turbo enabled, 1 = turbo disabled

# Disable turbo boost
echo 1 | sudo tee /sys/devices/system/cpu/intel_pstate/no_turbo

# Check turbo ratio limits (per-core)
# MSR 0x1AD through 0x1AE (use rdmsr from msr-tools)
sudo rdmsr 0x1AD  # TURBO_RATIO_LIMIT

# Turbo Boost Max 3.0 (preferred cores)
cat /sys/devices/system/cpu/intel_pstate/turbo_pct

AMD Core Performance Boost

# AMD boost control (if available)
cat /sys/devices/system/cpu/cpufreq/policy0/boost
# 0 = boost disabled, 1 = boost enabled

# Disable boost
echo 0 | sudo tee /sys/devices/system/cpu/cpufreq/policy0/boost

CPU Power Domains and Policies

Modern CPUs organize cores into power domains (packages, clusters, CCX/CCD on AMD). A cpufreq_policy groups CPUs that share the same clock domain and must scale frequency together.

# List which CPUs share a policy
cat /sys/devices/system/cpu/cpufreq/policy0/related_cpus
# Output: 0 1 2 3 4 5 6 7  (all share same clock)

# On multi-socket or hybrid systems:
cat /sys/devices/system/cpu/cpufreq/policy0/related_cpus  # P-cores
cat /sys/devices/system/cpu/cpufreq/policy8/related_cpus  # E-cores

# Affected CPUs (can be influenced by this policy)
cat /sys/devices/system/cpu/cpufreq/policy0/affected_cpus

Hybrid Architectures (Intel Alder Lake+, AMD Zen 5)

flowchart TB
    subgraph P_Core["P-cores (Performance)"]
        P0["CPU 0-7"]
        P_POL["policy0: 800MHz - 5200MHz"]
    end
    
    subgraph E_Core["E-cores (Efficiency)"]
        E0["CPU 8-15"]
        E_POL["policy8: 600MHz - 3800MHz"]
    end
    
    P0 --> P_POL
    E0 --> E_POL
    
    P_POL --> PERF_GOV["performance / schedutil"]
    E_POL --> SAVE_GOV["powersave / schedutil"]

Monitoring CPU Frequency

Real-Time Monitoring

# Watch current frequency of all CPUs
watch -n 0.5 'cat /sys/devices/system/cpu/cpu*/cpufreq/scaling_cur_freq'

# Using cpupower tool
sudo cpupower frequency-info
sudo cpupower monitor

# Turbostat (Intel) - shows actual frequencies, C-states, power
sudo turbostat --interval 1
# Output includes:
# PkgWatt  CorWatt  GFXWatt  Avg_MHz  Busy%  Bzy_MHz  TSC_MHz
# 15.02    12.34    0.00     2456     45.2   5432     3600

# Using perf for frequency statistics
sudo perf stat -e power/energy-pkg/ -a sleep 5

# Per-CPU frequency from /proc
grep "cpu MHz" /proc/cpuinfo

Frequency Statistics via cpufreq_stats

# Time spent at each frequency (kernel cpufreq_stats module)
cat /sys/devices/system/cpu/cpufreq/policy0/stats/time_in_state
# Output:
# 1200000 4523
# 1400000 2341
# 1600000 15234
# 1800000 34521
# 2000000 12345
# 2200000 6789
# 2400000 89012

# Total transitions
cat /sys/devices/system/cpu/cpufreq/policy0/stats/total_trans

# Transition table (how often transitions between specific frequencies occur)
cat /sys/devices/system/cpu/cpufreq/policy0/stats/trans_table

Energy Performance Preference (EPP) and EPB

Intel Energy Performance Bias (EPB)

# Read current EPB (0=performance, 15=powersave)
sudo rdmsr -a 0x1B0  # IA32_ENERGY_PERF_BIAS

# Set EPB via x86_energy_perf_policy tool
sudo x86_energy_perf_policy performance
sudo x86_energy_perf_policy balance-performance
sudo x86_energy_perf_policy balance-power
sudo x86_energy_perf_policy power

Intel Hardware P-States (HWP)

HWP (also known as Speed Shift) allows the hardware to autonomously control frequency within software-defined limits, providing faster frequency transitions than software-based scaling.

# Check if HWP is enabled
dmesg | grep -i hwp

# With intel_pstate in active mode and HWP:
# - Hardware manages frequency autonomously
# - Software sets hints via min/max performance percentages
# - Transitions occur in < 1ms (vs ~10ms for software scaling)

# Set HWP hints
echo 50 | sudo tee /sys/devices/system/cpu/intel_pstate/min_perf_pct
echo 100 | sudo tee /sys/devices/system/cpu/intel_pstate/max_perf_pct

Kernel Boot Parameters

# Force specific driver
intel_pstate=disable       # Disable intel_pstate, use acpi-cpufreq
intel_pstate=passive       # intel_pstate in passive mode
amd_pstate=active          # AMD p-state in active mode
cpufreq.off=1              # Disable cpufreq entirely

# Force specific governor
cpufreq.default_governor=performance

# Disable turbo boost at boot
intel_pstate=no_turbo

# Example GRUB configuration
# GRUB_CMDLINE_LINUX="intel_pstate=passive"

Practical Examples

Setting Up Power-Saving Profile

#!/bin/bash
# power-save.sh - Configure CPU for power saving

for policy in /sys/devices/system/cpu/cpufreq/policy*/; do
    echo powersave | sudo tee ${policy}scaling_governor
done

# Limit maximum frequency to 50% of capability
for policy in /sys/devices/system/cpu/cpufreq/policy*/; do
    max=$(cat ${policy}cpuinfo_max_freq)
    half=$((max / 2))
    echo $half | sudo tee ${policy}scaling_max_freq
done

# Disable turbo boost (Intel)
if [ -f /sys/devices/system/cpu/intel_pstate/no_turbo ]; then
    echo 1 | sudo tee /sys/devices/system/cpu/intel_pstate/no_turbo
fi

Setting Up Performance Profile

#!/bin/bash
# performance.sh - Configure CPU for maximum performance

for policy in /sys/devices/system/cpu/cpufreq/policy*/; do
    echo performance | sudo tee ${policy}scaling_governor
    # Set min to max
    max=$(cat ${policy}cpuinfo_max_freq)
    echo $max | sudo tee ${policy}scaling_min_freq
done

# Enable turbo boost (Intel)
if [ -f /sys/devices/system/cpu/intel_pstate/no_turbo ]; then
    echo 0 | sudo tee /sys/devices/system/cpu/intel_pstate/no_turbo
fi

Monitoring Script

#!/bin/bash
# freq-monitor.sh - Real-time CPU frequency monitoring

while true; do
    clear
    echo "=== CPU Frequency Monitor ==="
    echo "Time: $(date)"
    echo ""
    for cpu in /sys/devices/system/cpu/cpu[0-9]*/cpufreq/; do
        cpuid=$(basename $(dirname $cpu))
        cur=$(cat ${cpu}scaling_cur_freq 2>/dev/null || echo "N/A")
        gov=$(cat ${cpu}scaling_governor 2>/dev/null || echo "N/A")
        printf "%-6s: %8s kHz  [%s]\n" "$cpuid" "$cur" "$gov"
    done
    sleep 1
done

Troubleshooting

Common Issues

Diagnostic Commands

# Check loaded cpufreq modules
lsmod | grep -i cpufreq
lsmod | grep -i pstate

# Check active driver
ls -la /sys/devices/system/cpu/cpufreq/policy0/scaling_driver

# Check kernel power management configuration
zcat /proc/config.gz | grep -i cpufreq
# CONFIG_CPU_FREQ=y
# CONFIG_CPU_FREQ_GOV_SCHEDUTIL=y

# Check for thermal throttling
cat /sys/class/thermal/thermal_zone*/temp
dmesg | grep -i "thermal\|throttl"

References

• Linux kernel documentation, “CPU Performance Scaling,” https://docs.kernel.org/admin-guide/pm/cpufreq.html
• Linux kernel documentation, “CPUFreq Governors,” https://docs.kernel.org/admin-guide/pm/cpufreq-drivers.html
• Linux cpufreq subsystem, drivers/cpufreq/ in kernel source
• Intel 64 and IA-32 Architectures Software Developer’s Manual, Volume 3, Chapter 14 (Power Management)
• turbostat tool documentation, tools/power/x86/turbostat/ in kernel source
• cpupower tool documentation: https://manpages.ubuntu.com/manpages/noble/man1/cpupower.1.html
• Rafael J. Wysocki, “CPU Performance Scaling in Linux,” Intel Corporation, 2017