Variable results in benchmarking making changes difficult to measure

[ydebilloez]

Dear all,

When running benchmarks, I see variable results on mac and linux in between each run. Can someone hint me to fix the execution speed in following conditions: system idle and system busy. I am thinking of real time linux or fixing processor speed. I can bear with short differences in runtime by false clock measurements in sub-second runtimes but it would be nice to fix this as well.

Code: Select all

$./comparebench.sh 
~/chess/belofte/dist/Linux-x86-64 ~/chess/belofte/test
...
bench Brute Force Time: 57.308 Nodes: 8470712 Total N.: 8870156 NPS: 154778
...
$ ./comparebench.sh 
...
bench Brute Force Time: 58.077 Nodes: 8470712 Total N.: 8870156 NPS: 152729

Subsequent runs are sometimes more than 1 second off.
Thanks,
Yves

[Ras]

I am thinking of real time linux or fixing processor speed.

I disable the CPU boost during testing so that the CPU will stay at the base clock.

Disable boost:

Code: Select all

echo "0" | sudo tee /sys/devices/system/cpu/cpufreq/boost

Enable boost:

Code: Select all

echo "1" | sudo tee /sys/devices/system/cpu/cpufreq/boost

Check boost status:

Code: Select all

cat /sys/devices/system/cpu/cpufreq/boost

If you're interested, I've made a shell script and a Zenity script around that so that I can easily access that from the desktop start menu.

[ydebilloez]

I disable the CPU boost during testing so that the CPU will stay at the base clock.

Disable boost:

Code: Select all

echo "0" | sudo tee /sys/devices/system/cpu/cpufreq/boost

Enable boost:

Code: Select all

echo "1" | sudo tee /sys/devices/system/cpu/cpufreq/boost

Hi Ras,

Thanks for your suggestion. On my pc, I get Permission denied when running this. But you gave me a clue. This is the solution I found working to some extend.

Code: Select all

# echo "1" | sudo tee /sys/devices/system/cpu/intel_pstate/no_turbo

And setting it back to "0" in to resume to normal. Turbo enabled, I get variations of 2-5%. With the turbo disabled, I get variations well below 0.5% on long runs (at the expense of 20% execution speed), both on AC and battery. So perfectly usable.

Thanks for your feedback,
Yves