CPU频率/电压优化及OpenMPI CPU密集型任务防降频性能优化咨询

Hi there, let's break down your questions step by step based on your Core i7-3770 (4 cores with HT) setup and the thermal throttling issue you're facing with OpenMPI CPU-bound workloads.

Is optimizing CPU frequency to avoid throttling reasonable? Which mode delivers better performance? #

Short answer: Yes, optimizing frequency to avoid thermal throttling is absolutely reasonable, and in your case, it will likely deliver better consistent performance than relying solely on the ondemand governor. Here's why:

• The ondemand governor works by ramping up CPU frequency to the maximum when workload is high, and scaling down when idle. But it doesn't account for CPU temperature—once your CPU hits its thermal threshold (TJ Max, which is 77℃ for the i7-3770), the hardware will force throttling regardless of the governor setting. This leads to frequent frequency spikes and drops, which disrupts your workload's performance and lowers average throughput.
• By manually or automatically setting a stable frequency that keeps your CPU temperature below the throttling threshold, you can maintain 100% CPU utilization continuously. This consistent performance will almost always outperform the fluctuating speeds caused by ondemand + thermal throttling.

That said, hardware fixes (cleaning dust, re-applying thermal paste, upgrading the CPU fan) are still the most impactful long-term solution—software tuning is a complementary measure, not a replacement.

How to optimize CPU frequency (manual and automatic methods) #

1. First, diagnose your current state #

Before tuning, get a clear picture of your CPU's behavior:

• Check current CPU frequencies: cat /proc/cpuinfo | grep "cpu MHz"
• Monitor CPU temperatures: Install lm-sensors with sudo apt install lm-sensors, run sudo sensors-detect to configure, then use sensors to view temps.
• Check if throttling is happening: dmesg | grep throttling or cat /sys/devices/system/cpu/cpu*/cpufreq/throttle_state

2. Manual frequency tuning #

The simplest approach is to lock your CPU to a frequency that stays cool enough to avoid throttling:

1. Install cpufrequtils to manage frequencies: sudo apt install cpufrequtils
2. Check available governors: cpufreq-info | grep available
3. Switch all CPUs to the userspace governor (lets you set custom frequencies):

   for i in /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor; do echo userspace | sudo tee $i; done
   

4. Set a test frequency (start with your CPU's base clock, 3400000 = 3.4GHz for i7-3770):

   for i in /sys/devices/system/cpu/cpu*/cpufreq/scaling_setspeed; do echo 3400000 | sudo tee $i; done
   

5. Run your OpenMPI workload, monitor temperatures. If temps stay below ~70℃ (well below the 77℃ TJ Max), try increasing frequency in 200MHz increments. If temps spike close to 77℃, drop the frequency until it stabilizes.

3. Automatic frequency tuning script #

If you want dynamic adjustment based on real-time temperature, here's a simple bash script that scales frequencies up/down to avoid throttling:

#!/bin/bash

# Configuration parameters
MAX_ALLOWED_TEMP=70  # Keep this 5-7℃ below your CPU's TJ Max (77℃ for i7-3770)
MIN_FREQ=2000000     # Minimum frequency (2GHz, adjust as needed)
MAX_FREQ=3900000     # Maximum frequency (3.9GHz, i7-3770's max turbo)
CHECK_INTERVAL=5     # Check temperature every 5 seconds

# Switch all CPUs to userspace governor
for i in /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor; do
    echo userspace | sudo tee $i > /dev/null
done

# Start at maximum frequency
current_freq=$MAX_FREQ
for i in /sys/devices/system/cpu/cpu*/cpufreq/scaling_setspeed; do
    echo $current_freq | sudo tee $i > /dev/null
done

echo "Starting automatic CPU frequency tuning. Press Ctrl+C to stop."

while true; do
    # Calculate average core temperature
    avg_temp=$(sensors | grep -A 0 "Core" | awk '{print $3}' | sed 's/[^0-9.]//g' | awk '{sum+=$1} END {print sum/NR}')
    avg_temp=$(printf "%.0f" $avg_temp)

    echo "Current average temp: ${avg_temp}℃ | Current freq: $((current_freq/1000000))GHz"

    # Adjust frequency based on temperature
    if [ $avg_temp -gt $MAX_ALLOWED_TEMP ]; then
        # Cool down: reduce frequency by 200MHz
        current_freq=$((current_freq - 200000))
        [ $current_freq -lt $MIN_FREQ ] && current_freq=$MIN_FREQ
        echo "Temp too high! Dropping freq to $((current_freq/1000000))GHz"
        for i in /sys/devices/system/cpu/cpu*/cpufreq/scaling_setspeed; do
            echo $current_freq | sudo tee $i > /dev/null
        done
    elif [ $avg_temp -lt $((MAX_ALLOWED_TEMP - 5)) ]; then
        # Safe temp: increase frequency by 200MHz
        current_freq=$((current_freq + 200000))
        [ $current_freq -gt $MAX_FREQ ] && current_freq=$MAX_FREQ
        echo "Temp is low! Raising freq to $((current_freq/1000000))GHz"
        for i in /sys/devices/system/cpu/cpu*/cpufreq/scaling_setspeed; do
            echo $current_freq | sudo tee $i > /dev/null
        done
    fi

    sleep $CHECK_INTERVAL
done

How to use the script:

• Save it as cpufreq_auto_tune.sh
• Make it executable: chmod +x cpufreq_auto_tune.sh
• Run with sudo (required for modifying CPU settings): sudo ./cpufreq_auto_tune.sh
• Adjust the configuration parameters at the top to match your CPU's capabilities and cooling setup.

Additional tips for OpenMPI workloads #

• Bind tasks to cores: Use mpirun --bind-to core or --bind-to socket to avoid thread migration, which reduces overhead and helps distribute heat more evenly across cores.
• Permanent governor setup: If you want to keep the userspace governor after reboot, edit /etc/default/cpufrequtils and set GOVERNOR="userspace", along with MIN_SPEED and MAX_SPEED values.

备注：内容来源于stack exchange，提问作者roentgenium