cpu_percent_% still higher than 100% in Alumet v0.9.3

[thealanjason]

Describe the bug
We observe cpu_percentage still higher than 100% even though PR #341 says that it was fixed in v0.9.3.

The hardware on which we ran the tests has 96 CPU Cores.

Refer the following two issues where we the bug surfaces:

• Identify Polling Frequency Threshold for Spikes thealanjason/alumet-viz#7 (comment)
• cpu_percent is still above 100% thealanjason/topographic_uq_energy#11
• Monte carlo update thealanjason/topographic_uq_energy#9 (comment)

Are we missing something? Do we need to use the cgroups plugin?

Environment information
OS: Ubuntu Linux 24.04LTS
Alumet version²: v0.9.3.
Enabled plugins:

    🧩 5 plugins started:
        - csv v0.2.0
        - rapl v0.3.1
        - procfs v0.2.1
        - nvml v0.5.0
        - energy-attribution v0.1.1

To Reproduce
Steps to reproduce the behavior:
Use the config here: https://github.com/thealanjason/energy_measurement/blob/main/measurement_tools/alumet/experiments/08_cpu_percentage_test/alumet-config-cpu-percent-test.toml

to exec the program here: https://github.com/thealanjason/energy_measurement/blob/86dcdd60bf89a0996294a7200fd82230c2658b3a/case_studies/gemm/gemm.py

The program would need a conda/micromamba to be created with the following environment.yml file

name: test-case-gemm
channels:
  - conda-forge
dependencies:
  - numpy

Expected behavior
cpu_percent less than 100% when the program uses all 96 cores.

Logs / Output

• Log file: https://github.com/thealanjason/energy_measurement/blob/86dcdd60bf89a0996294a7200fd82230c2658b3a/measurement_tools/alumet/experiments/08_cpu_percentage_test/alumet-agent-cpu-percent-test.log
• CSV Output file: https://github.com/thealanjason/energy_measurement/blob/86dcdd60bf89a0996294a7200fd82230c2658b3a/measurement_tools/alumet/experiments/08_cpu_percentage_test/alumet-output-rapl%2Benergy-attribution.csv

[TheElectronWill]

Thanks for creating an issue. You are using the procfs plugin to get process-related metrics, right? So, #341 does not impact your usage, it was only for cgroups-related metrics. They both use cpu_percent, because it means the exact same thing in both cases, the only difference is the consumer and… the scale (!)

• For the cgroups plugins, cpu_percent is in [0, 100], 100% means that all core are used at 100%.
• For the procfs plugin, cpu_percent is in [0, ?], 100% means that one core is used at 100%. See process.rs.

The two should be aligned, we should fix this.

In the meantime, you can divide by the number of (logical) cores (96?).