Fix for M3 Max - Using average of cores instead of relying on cluster idle ratio

[gavi]

Seems like the idle ratio of Cluster is reporting as zero in powermetrics which might be causing the CPU gauges to show 100%. This code fix simply averages the individual core idle ratios.

[gavi]

Before fix on M3 Max (40 core GPU)

After fix

[ivanfioravanti]

please @tlkh merge this one 🙏

[samarthm314]

@gavi Thank you so much! Your fix worked for me.

[chiltalk]

I tried the patch. It helped a little bit but seems wrong still. I am using M3 Max 16C/40GPU with just chrome browser running to test.

M3 max 16c/40gpu still showing 90% usage on e-Cores and 42-100% usage on P-cores.

Comparing this to

istatmenu (4 cpus - probably E-cores pegged near 90%, and 12 other CPU (I assume P-cores) showing nothing 1% here and there. Activity monitor is matching the istatsmenu

At least it's no longer pegging both e-cores and p-cores at 100% but something is still wrong.

I wonder how the youtubers are checking the stats these days with asitop. :)

[gavi]

I am not sure what to trust though. If you look at the plist file that is generated by powermetrics , the p-cores were showing usage but like you said iStats and the activity monitor were not. I believe this will be a temporary fix until they actually fix powermetrics itself.

[phgrosjean]

Using a M3 Max here. The readings with this PR seem OK when the P-cores are actives, but they are overestimated when they are not. I think it is because the calculations do not take into account the P-Clusters and the P-Cores can also be down a part of the time, and that must be added to the "idle" time. I got something more plausible when I replace line 97

total_idle_ratio += cpu["idle_ratio"]

with this:

total_idle_ratio += cluster["down_ratio"] + (1 - cluster["down_ratio"]) * (cpu["idle_ratio"] + cpu["down_ratio"])

[gavi]

@phgrosjean I think your change is much better. I am updating my PR

[chiltalk]

Oh yes awesome. I noticed the down CPUs counter last night when I ran powermetrics manually to understand the output. Didn’t know how to account for them yet. I did notice that the graph looks more accurate when I stress the computer to actually use the P-cores and that could explain what you just wrote. Will try soon. Thank you!

…

On Wed, Jan 10, 2024 at 12:03 PM Philippe Grosjean ***@***.***> wrote: Using a M3 Max here. The readings with this PR seem OK when the P-cores are actives, but they are overestimated when they are not. I think it is because the calculations do not take into account the P-Clusters and the P-Cores *can also be down* a part of the time, and that must be added to the "idle" time. I got something more plausible when I replace line 97 total_idle_ratio += cpu["idle_ratio"] with this: total_idle_ratio += cluster["down_ratio"] + (1 - cluster["down_ratio"]) * (cpu["idle_ratio"] + cpu["down_ratio"]) — Reply to this email directly, view it on GitHub <#68 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADLJ5PU5N7PDC26IYKVG7H3YN3CUPAVCNFSM6AAAAABBGYINLCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQOBVGI2DMNJQGI> . You are receiving this because you commented.Message ID: ***@***.***>

[gavi]

I updated the PR

[chiltalk]

Looks a lot better to me! thank you both.

[iamaziz]

@gavi thanks for the fix! would be great to see this merged.

[Ender-Wang]

For anyone who are not sure how to port this changes to your own machine, just replace the parsers.py file with the changed parsers.py file in this PR, for reference, I installed asitop via pip and it's located at /opt/homebrew/lib/python3.11/site-packages/asitop for me. Then run asitop again in the terminal, you are good to go.

[iamaziz]

pip install git+https://github.com/gavi/asitop.git should work too.

[Proryanator]

Looking forward to this being merged and released, thanks @gavi !

[Ender-Wang]

Looking forward to this being merged and released, thanks @gavi !

Not likely, seems the repo author is off the planet to Mars.