cpu feature where user can pass what cores will be disabled

[nwildner]

In this feature request idea, the user would be allowed to power down some of the CPU cores when on battery mode for some extra juice.

On my current model(ASUS TUF Dash F15) keeping only core0 and core1 and the "efficiency cores"(E-Cores) cpu12, cpu13, cpu14 and cpu15 last four cores provide me roughly and extra of 20 to 35% of battery time. I think that the fact you have less cores turned on and the 4 E-Cores have lower max mhz helps a bit with some indirect power saving:

lscpu -e 
CPU NODE SOCKET CORE L1d:L1i:L2:L3 ONLINE    MAXMHZ   MINMHZ       MHZ
  0    0      0    0 0:0:0:0          yes 4600.0000 400.0000  400.0000
  1    0      0    0 0:0:0:0          yes 4600.0000 400.0000  400.0000
  2    0      0    1 4:4:1:0          yes 4600.0000 400.0000  728.3510
  3    0      0    1 4:4:1:0          yes 4600.0000 400.0000  400.0000
  4    0      0    2 8:8:2:0          yes 4700.0000 400.0000  598.7270
  5    0      0    2 8:8:2:0          yes 4700.0000 400.0000  400.0000
  6    0      0    3 12:12:3:0        yes 4700.0000 400.0000  843.1310
  7    0      0    3 12:12:3:0        yes 4700.0000 400.0000  400.0000
  8    0      0    4 16:16:4:0        yes 4600.0000 400.0000  400.0000
  9    0      0    4 16:16:4:0        yes 4600.0000 400.0000  491.3900
 10    0      0    5 20:20:5:0        yes 4600.0000 400.0000  400.0000
 11    0      0    5 20:20:5:0        yes 4600.0000 400.0000  400.0000
 12    0      0    6 28:28:7:0        yes 3500.0000 400.0000  400.0000
 13    0      0    7 29:29:7:0        yes 3500.0000 400.0000  615.0990
 14    0      0    8 30:30:7:0        yes 3500.0000 400.0000 1102.9550
 15    0      0    9 31:31:7:0        yes 3500.0000 400.0000  400.0000

The initial solution would be: A parameter where the user can pass down which CPUs can be shutdown through hotplug event or if the laptop is starting on battery only.

Example:

CPU_DISABLE_CORES_BAT="2 3 4 5 6 7 8 9 10 11"

Some insights on how this would work:

• A script would send echo 0 > /sys/devices/system/cpu/cpuX/online where X is the CPU index provided on CPU_DISABLE_CORES_BAT
• cpu0 could be ignored since it can't be hotplugged to off. On my current setup, I ignore the first 2 cores just in case (cpu0 an cpu1)
• Implicitly turn on all cores back again when power is plugged if there is something declared on CPU_DISABLE_CORES_BAT

That is a feature I miss most from laptop-mode-tools and there were some predefined values like "turn off half of the cores" were available but, TLP could start by simply allowing to hotplug CPUs by index on this first implementation.

I'm currently controlling this through udev since laptop-mode is dead in the water and I've added a patch specifically to manage e-cores which I think the owner lost interest

• https://github.com/rickysarraf/laptop-mode-tools/pull/206
• https://github.com/rickysarraf/laptop-mode-tools/blob/lmt-upstream/usr/share/laptop-mode-tools/modules/cpuhotplug

Since I've migrated recently to TLP and didn't had the time to learn the code, I'm opening a Feature Request here :)

Related:

• https://docs.kernel.org/core-api/cpu_hotplug.html

[linrunner]

I think this is a feature for a very small niche of users. Nevertheless, I would accept a PR from you.

First, to make this usable and user-friendly, tlp-stat -p must be able to output the type (P, E, LP-E) for each core. So that users get to know what they should enter into the configuration. There was already an attempt to write something like this, but the creator lost interest: https://github.com/linrunner/TLP/issues/736#issuecomment-2036295981

The configuration must be symmetrical, i.e. CPU_DISABLE_CORES_ON_AC and CPU_DISABLE_CORES_ON_BAT, so that it is defined what happens when switching the power source. Both defaults would be empty of course.

[nwildner]

Thanks for the heads up @linrunner

Wouldn't it be better to just categorize the CPUs into 2 groups, the E-Core and the Performance ones? As quoted on the other case, even Intel itself through their software only really categorize them into these two groups simplifying the nomenclature a little bit:

• https://github.com/linrunner/TLP/issues/736#issuecomment-2065437644
• https://industrialpc.com/blog/intel-cpu-ecore-pcore/

Also, it is fairly easy to probe that using only 2 files and no complex calculations. Here:

[nwildner@arrakis ~]$ cat /sys/devices/cpu_atom/cpus 
12-15
[nwildner@arrakis ~]$ cat /sys/devices/cpu_core/cpus 
0-11

Cores listed inside /sys/devices/cpu_atom/cpus are E-Cores and cores listed at /sys/devices/cpu_core/cpus are the performance ones.

Another way to compare would be to match what CPU lines from the lscpu -e output have more than one CORE column associated with it, indicating a SMT cpu, thus, performance. E-Cores are by design non-SMT. I still prefer relying on the information that is already provided on the /sys interface on the files referenced above.

What do you think?

[linrunner]

@nwildner I think a distinction in P- and E/LP-E cores based on /sys/devices/cpu_{atom,core}/cpus is the most pragmatic. I always try to avoid dependencies on external tools if there is a way in sysfs. However, I do not have a computer with these Intel CPUs.

[linrunner]

@nwildner isn't intel-lpmd just what you want to achieve?

See https://github.com/linrunner/TLP/issues/810

[nwildner]

Neat, thanks for the highlight @linrunner

I've activated this service and indeed, while it will not turn those cores down, tasks are being placed mainly on the E-Cores, which is kinda cool. Definitely, intel_lpmd is a nice additional software to pair with tlp for energy optimization.