For Arne by ribalba · Pull Request #4 · green-kernel/procpower

ribalba · 2026-02-13T10:25:26Z

So you can see what I have changed

ribalba · 2026-02-13T10:27:51Z

@ArneTR can you see how this performs. I create the data with running energy-logger and then benchmark --rounds 10 Also I added --trim-upper-quantile 99 as a parameter to the ./model to remove outliers. Also I now sample randomly for the test set.

ArneTR

Code works. Comments inline

ArneTR · 2026-03-18T09:18:30Z

src/energy_proc.c

    rcu_read_lock();
-    struct task_struct *p = pid_task(find_vpid(pid), PIDTYPE_PID);
-    alive = p && pid_alive(p);
+    task = pid_task(find_vpid(tgid), PIDTYPE_TGID);


Why this change?

I removed all the alive saving as this was quite the overhead.

ArneTR · 2026-03-18T09:18:44Z

src/energy_proc.c

+    return task;
+}
+
+static int ensure_pid_entry(u32 tgid)


Can you explain what this does and why it is necessary and was not necessary before?

It is the same thing as below but only in a function and changed to work on tasks and not threads/processes

ArneTR · 2026-03-18T12:59:54Z

src/energy_proc.c

-
-        rcu_read_lock();
-        u32 tgid = task_tgid_nr(t);
+    /* discover tasks that do not have a pid entry yet */


Can you summarize for the changes from here until the end of the file why you changed this and what is the expected gain?

I mainly changed it from processes to tasks and now remove tasks that are not alive anymore.

ArneTR · 2026-03-18T13:00:50Z

tools/model.py

+    q = train_df["target_uj"].quantile(quantile)
+    return (
+        train_df[train_df["target_uj"] <= q],
+        test_df[test_df["target_uj"] <= q],


I would not trim the test set also as this skews the information about its prediction power

ArneTR · 2026-03-18T13:01:29Z

tools/model.py

+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(


At the moment the code cannot compare to a new out of sample file. I opened up a PR to do so #5

ArneTR · 2026-03-18T13:13:53Z

tools/model.py

+    )
+
+
+def fit_nonnegative_ridge(X: np.ndarray, y: np.ndarray, alpha: float) -> np.ndarray:


Can this model be used in the kernel? Or is it just to showcase?

To my understanding this uses no intercept. But I believe we should have one

ArneTR · 2026-03-18T13:23:16Z

Just for documentation purposes these are the results I am getting when running this code out of the box:

(venv) arne@framebook:~/Sites/procpower/tools$ python3 model.py /tmp/energy-gKpkxu81.log --mode delta --alpha 10
# mode=delta alpha=10.0
# rows_total=1246 rows_train=995 rows_test=250

# --- test metrics (current module defaults) ---
R2     : -3.00767
RMSE   : 3340332.139 uJ
MAE    : 2897199.311 uJ
MAPE   : 99.824 %

# --- test metrics (linear non-negative ridge) ---
R2     : 0.87867
RMSE   : 581195.727 uJ
MAE    : 502103.984 uJ
MAPE   : 27.267 %

# --- test metrics (nonlinear reference) ---
R2     : 0.99567
RMSE   : 109762.387 uJ
MAE    : 72772.653 uJ
MAPE   : 3.640 %

# --- test metrics (distilled kernel model) ---
R2     : 0.89540
RMSE   : 539636.194 uJ
MAE    : 459136.968 uJ
MAPE   : 24.124 %

# --- train metrics (distilled kernel model) ---
R2     : 0.90925
RMSE   : 517380.926 uJ
MAE    : 453785.128 uJ
MAPE   : 24.045 %

I have also seen better runs on my system though where test metrics (distilled kernel model) goes down to 16% error rate. The reason for that is unknown.

ribalba

As already mentioned I didn't spend to much time on the procpower refactoring. I even forgot I had done that, as already mentioned. I will have to work on this again at some stage to validate everything.

ribalba · 2026-03-19T14:03:03Z

src/energy_proc.c

    rcu_read_lock();
-    struct task_struct *p = pid_task(find_vpid(pid), PIDTYPE_PID);
-    alive = p && pid_alive(p);
+    task = pid_task(find_vpid(tgid), PIDTYPE_TGID);


I removed all the alive saving as this was quite the overhead.

ribalba · 2026-03-19T14:13:35Z

src/energy_proc.c

+    return task;
+}
+
+static int ensure_pid_entry(u32 tgid)


It is the same thing as below but only in a function and changed to work on tasks and not threads/processes

ribalba · 2026-03-19T14:14:30Z

src/energy_proc.c

-
-        rcu_read_lock();
-        u32 tgid = task_tgid_nr(t);
+    /* discover tasks that do not have a pid entry yet */


I mainly changed it from processes to tasks and now remove tasks that are not alive anymore.

ribalba · 2026-03-19T14:30:30Z

(venv) didi@fedora:~/code/procpower/tools$ python3 model.py /tmp/energy-kIyAfwhh.log --mode delta --alpha 10
# mode=delta alpha=10.0
# rows_total=2678 rows_train=2139 rows_test=536

# --- test metrics (current module defaults) ---
R2     : -1.12845
RMSE   : 6359584.911 uJ
MAE    : 4649711.760 uJ
MAPE   : 99.898 %

# --- test metrics (linear non-negative ridge) ---
R2     : 0.91347
RMSE   : 1282310.178 uJ
MAE    : 1077131.597 uJ
MAPE   : 27.082 %

# --- test metrics (nonlinear reference) ---
R2     : 0.77467
RMSE   : 2069211.622 uJ
MAE    : 421845.440 uJ
MAPE   : 6.913 %

# --- test metrics (distilled kernel model) ---
R2     : 0.90722
RMSE   : 1327807.047 uJ
MAE    : 1077579.672 uJ
MAPE   : 27.704 %

# --- train metrics (distilled kernel model) ---
R2     : 0.96312
RMSE   : 1714277.637 uJ
MAE    : 1106326.862 uJ
MAPE   : 26.251 %

* main: Added more ignores

ribalba added 3 commits February 12, 2026 16:28

Small fixes

c055816

non linear models

70e739d

Adds the benchmark and smaller fixes

2c22c02

ArneTR mentioned this pull request Mar 18, 2026

Out of sample testing #5

Open

ArneTR reviewed Mar 18, 2026

View reviewed changes

Fixes the hetzner download

a55eda1

ribalba commented Mar 19, 2026

View reviewed changes

ArneTR added 3 commits March 21, 2026 13:24

Moved argparse to main block and calling main from there

114230e

Allowing model to fit on different target cols

ff3f026

Merge branch 'main' into fixes

682b42b

* main: Added more ignores

		)


		def fit_nonnegative_ridge(X: np.ndarray, y: np.ndarray, alpha: float) -> np.ndarray:

Conversation

ribalba commented Feb 13, 2026

Uh oh!

ribalba commented Feb 13, 2026

Uh oh!

ArneTR left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ArneTR commented Mar 18, 2026

Uh oh!

ribalba left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ribalba commented Mar 19, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants