Reflex CPUFreq Governor

A Linux kernel CPUFreq governor that combines the instant responsiveness of the interactive governor with the efficiency of modern schedutil PELT-based scaling.

Overview

Reflex is a hybrid CPUFreq governor designed to deliver the best of both worlds:

• Instant high-speed jumps on idle-to-busy transitions, based on real CPU busy% measured via kcpustat idle-time counters.
• PELT-proportional scaling for sustained workloads, identical to schedutil with 1.25x DVFS headroom.
• Smooth blending via a PELT-complementary exponential decay algorithm (32 ms half-life), ensuring high-speed + PELT coverage stays at ~100% at all times.

After approximately 10 half-lives (~320 ms), PELT takes full control with no residual frequency floor lock-in.

How It Works

Blending Algorithm

When a CPU wakes from idle:

1. An immediate frequency floor is set based on measured CPU busy%.
2. This high-speed contribution decays exponentially: blended = pelt + (hispeed - pelt) >> half_lives.
3. As high-speed decays, PELT's utilization signal naturally rises to fill the gap.

Asymmetric EWMA Filtering

The busy% measurement uses asymmetric filtering:

• Ramp-up: Instant tracking (filtered = measured).
• Ramp-down: Exponential decay (filtered -= (filtered - measured) >> shift), tunable via hispeed_filter_shift.

Two-Phase Idle-Time Measurement

Uses kcpustat kernel idle-time counters (not PELT) for raw CPU busy ratio, with a two-phase measurement protocol that prevents transient busy_pct=0 readings from being mistakenly treated as genuine idle signals.

Grace Period for Decay Timer

The high-speed decay timer is only reset after two consecutive idle windows, preventing rapid idle/busy cycling from defeating the decay mechanism.

I/O Wait Boost

Schedutil-compatible I/O wait boosting: starts at 1/8 of max capacity, doubles on each consecutive I/O wait event, and decays by half each period when not waiting.

Tunables

Exposed under /sys/devices/system/cpu/cpufreq/reflex/:

Tunable	Default	Description
rate_limit_us	Driver default	Minimum interval between frequency updates (typically 1000-2000 us)
hispeed_window_us	4000	Observation window duration for busy% calculation (us)
hispeed_filter_shift	3	EWMA down-ramp shift for filtered busy% (0 = off, higher = slower decay)
version	(read-only)	Reports the governor version string

Installation

Reflex is delivered as a kernel patch. Apply it to your Linux kernel source tree:

1. Apply the patch

cd /path/to/linux
patch -p1 < /path/to/reflex/patches/0001-Reflex-CPUFreq-Governor-v0.2.1.patch

2. Enable in kernel config

CONFIG_CPU_FREQ_GOV_REFLEX=m

Or to set as the default governor:

CONFIG_CPU_FREQ_DEFAULT_GOV_REFLEX=y

3. Build

make modules

4. Load the module

sudo modprobe cpufreq_reflex

5. Activate

Note: If your system uses intel_pstate or amd_pstate in active mode, external governors like Reflex cannot be activated. Check the current mode:

cat /sys/devices/system/cpu/intel_pstate/status
# or
cat /sys/devices/system/cpu/amd_pstate/status

If it reports active, switch to passive or guided mode:

echo passive | sudo tee /sys/devices/system/cpu/intel_pstate/status
# or
echo passive | sudo tee /sys/devices/system/cpu/amd_pstate/status

To make this persistent across reboots, add a kernel boot parameter:

intel_pstate=passive
# or
amd_pstate=passive

Once the driver is in passive/guided mode, activate Reflex:

echo reflex | sudo tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor

Dependencies

• Linux kernel with CPU_FREQ, SMP, and CPU_FREQ_GOV_SCHEDUTIL enabled
• x86 or other architecture with a CPUFreq driver supporting the update_util callback (intel_pstate/amd_pstate in passive or guided mode, acpi-cpufreq, etc.)
• Standard kernel build toolchain (gcc or clang)