fix(bpf): exclude swapper process bpf_cpu_time

[vimalk78]

when cpu is idle, swapper process is scheduled, and kepler adds its bpf_cpu_time to kernel processes

adding a configuration to exclude swapper process

with-swapper

without-swapper

[github-actions]

🤖 SeineSailor

Here's a concise summary of the pull request changes:

Summary: This pull request, "fix(bpf): exclude swapper process bpf_cpu_time", introduces a configuration option to exclude the swapper process's bpf_cpu_time from kernel processes when the CPU is idle.

Key Modifications:

• Added ExcludeSwapperProcess boolean field to KeplerConfig struct
• Modified getKeplerConfig() function to include the new field
• Added ExcludeSwapperProcess() retrieval function
• Set default value of defaultExcludeSwapperProcess to true

Impact: These changes affect the internal implementation of the code, but do not alter the external interface or behavior.

Observations:

• The addition of a configuration option provides flexibility for users to customize the behavior of the bpf_cpu_time collection.
• The default value of true for defaultExcludeSwapperProcess suggests that excluding the swapper process is the recommended behavior.

Suggestions for Improvement:

• Consider adding a comment or documentation to explain the reasoning behind excluding the swapper process and the implications of enabling/disabling this feature.
• It would be helpful to include tests to verify the correct behavior of the ExcludeSwapperProcess feature.

[rootfs]

Instead of disabling idle time accounting, I am thinking of using the idle time to calculate the idle power as the next step.
cc @marceloamaral @sunya-ch

[rootfs]

how about add a new metrics called idle cpu time?

[sthaha]

Yes, this is a good idea and aligns with user / system / idle time in https://man7.org/linux/man-pages/man5/proc_stat.5.html

[sthaha]

@vimalk78 with this patch are you able to get sum(rate(kepler_process_bpf_cpu_time_ms_total[20s]))) == sum(rate(node_cpu_seconds_total{mode=~"user|system"}[20s]))

[vimalk78]

with this change, under load condition

[rootfs]

I am merging this now and will come up with a new idle time metrics PR.

[rootfs]

@vimalk78 @sthaha @KaiyiLiu1234 @sunya-ch

I tested this idle time exclusion in metal CI. There is improvement on SGD trainer but not much on exponential or logarithmic trainer.

Please find the new config parameter in kepler metal and VM systemd.

In the trained model validation report, those in Nov 01 are without this change, and those in Nov 04 are with this new config param. The SGD trainer MAPE sees improvement from 51% to 21%

[sthaha]

We should see improvement in all models in theory. @sunya-ch is there a way to plot or export all metrics used for training the model ?