Experimental comparisions

[Johhed15]

Hi! I am planning to use codecarbon to track the energy consumption from my code in a thesis where energy consumption from CPU one of the questions, And I later found this paper:

https://inria.hal.science/hal-04030223v2/file/_CCGrid23__An_experimental_comparison_of_software_based_power_meters__from_CPU_to_GPU.pdf
which said:
"""Code Carbon tends to report power consumption
values almost twice as high as other tools based on the same
interface for obtaining consumption values (Intel RAPL)."""
[Mathilde Jay, Vladimir Ostapenco, Laurent Lefèvre, Denis Trystram, Anne-Cécile Orgerie, et al.. An
experimental comparison of software-based power meters: focus on CPU and GPU. CCGrid 2023

• 23rd IEEE/ACM international symposium on cluster, cloud and internet computing, May 2023,
  Bangalore, India. pp.1-13, ff10.1109/CCGrid57682.2023.00020ff. ffhal-04030223v2f]

So I am wondering if this has been looked at before i start tracking my code.

Thanks!!

[benoit-cty]

Hello,

Thanks for letting us know about this. We will try to reproduce the paper and contact the authors.

Don't hesitate to report if you saw similar issue.

CodeCarbon works on multi-CPU architecture. I experienced this issue on an AMD CPU that has only one physical CPU but two die and the Linux kernel write two RAPL file, so we sum them.

Despite this specific case you will be fine.

[Johhed15]

Hi again!
I seem to get closer to the same results as them, but with a much lower variance in your estimations.

But I have one question regarding one of your parameters, is the measure_power_secs a sampling frequency from rapl? meaning that if iterations are quick, this needs to be low?

[benoit-cty]

measure_power_secs is used to wake up CodeCarbon to do the measure.
If you want to monitor short process you will prefer to use "task" mode as it doesn't use a scheduler.
See https://github.com/mlco2/codecarbon/blob/master/examples/task_inference.py

[Johhed15]

I am using "task" but sometimes I still get the message that the scheduler hasn't been running for x second if I initialized the EmissionsTracker with a low measure_power_secs . But maybe that does not even matter when i use 'task' then?

[benoit-cty]

Yes, the scheduler it stopped when task start :

codecarbon/codecarbon/emissions_tracker.py

Line 448 in 4987537

if self._scheduler:

We must prevent logging a warning when in task mode :

codecarbon/codecarbon/emissions_tracker.py

Line 743 in 4987537

if last_duration > warning_duration:

[Johhed15]

Thanks!

Last question, so when running with task, is the frequency then the frequency from RAPL(1000hz) and can be considered almost continous?

[benoit-cty]

I think so.

For the record, here is a paper for dealing with lower update frequency (20ms) of Nvidia GPU : https://www.tu-chemnitz.de/informatik/PI/forschung/publikationen/download/LR_europar13.pdf