Choice of the Core on mixed-CPU platform like Meteor-Lake

[dmpogo]

But, for amusement, choosing intel_pstate in kernel CONFIG, forces built-in performance and schedutil governors, and no other. What does that mean ?

It means: P-State can run in active mode ... or in passive mode. Yes you dont need the governors in active mode (because P-State has its own governors) ... but in passive mode ... so this kernel configuration is senseful (because the kernel configuration doesn't know which mode will be used in the running kernel later)

Indeed ! I forgot about the passive mode

[dmpogo]

[...] My only issue is that performance cores seem not just not preferred, but even avoided. [...]

TBH I dont know the internal scheduling policies for your MeteorLakeCPU ... so this is just a guess:

Maybe the kernel thinks that a Notebook-CPU (with different core-types) have more problems with energy than with performance and so prefers E-Cores ... and maybe this is different for a desktop-CPU. I see that first my (eight) P-Cores are used (but only ONE of the two logical cores; I have SMT), then all (16) E-cores and at last the 8 (second) logical cores of my P-Cores (I tested with "make -j8", "make -j12", "make -j16", "make -j24" and "make -j32").

I'll try to do more accurate testing. I also have SMT on my two P cores (which, btw, I did not find how to disable anymore in 6.12+ kernels, I would probably prefer just to have one thread on them, given that I have 10 other cores). So it is actually not obvious that -j4 on P-cores (using threads) is better than on 4 single threaded E-cores. But for a single threaded job I would much see it on P-core running that on E-core.

But how scheduler works in your case is exactly how I would expect it

[pietinger]

[...] However, your MeteorLake CPU has disabled [SMT] for the P-cores.

->

[...] I also have SMT on my two P cores [...]

Yes, that's right. I was mistaken. I had read that the new Intel CPUs no longer have SMT on the P-cores... but that only applies to the second (new) generation ArrowLake. MeteorLake P-cores still have SMT. (Sorry, I'm confusing things again).

[dmpogo]

[...] However, your MeteorLake CPU has disabled [SMT] for the P-cores.

->

[...] I also have SMT on my two P cores [...]

Yes, that's right. I was mistaken. I had read that the new Intel CPUs no longer have SMT on the P-cores... but that only applies to the second (new) generation ArrowLake. MeteorLake P-cores still have SMT. (Sorry, I'm confusing things again).

Interesting, my first though seeing these CPU's was why would I have SMT on P cores given all extra cores ? And now Intel agreed with me

[pingtoo]

From what I gather on the net it is recommended to setup intel-lpmd for Meteor Lake CPU.

As it was mentioned previous in pietinger post.

if you prefer use gentoo package the alternative sys-power/thermald could also work. however I am not sure if it need setup to enable it know how to use "hfi".

I'll have a look, though intel-lpmd by description is more doing the opposite - disactivate performance cores preferring low-power ones. I am more after more utilization of performance cores
There is still that choice of "performance" versus "balanced performance" that is still to be explored.

Sorry I think I lost track on the conversation on the "hfi".

Once you have intel-lpmd (modern standard for Meteor Lake and newer) or thermald (v2.5.4+) update support Meteor Lake. They can trigger HFI activation.

We never review your .config file do you mind share it? In my research there are too many different CONFIG_something need to be on/off, so instead we come back and forth about is this on and is that on too kind of questions may be I will just search myself and let you know if anything need to change.

Also if possible share the dmesg with initial boot sequence, same reason as above, many research note also mention that if you have this CONFIG_something you should see "something" in dmesg.

[dmpogo]

[

Sorry I think I lost track on the conversation on the "hfi".

Once you have intel-lpmd (modern standard for Meteor Lake and newer) or thermald (v2.5.4+) update support Meteor Lake. They can trigger HFI activation.

We never review your .config file do you mind share it? In my research there are too many different CONFIG_something need to be on/off, so instead we come back and forth about is this on and is that on too kind of questions may be I will just search myself and let you know if anything need to change.

Also if possible share the dmesg with initial boot sequence, same reason as above, many research note also mention that if you have this CONFIG_something you should see "something" in dmesg.

Yep, 'hfi' remained uninvestigated yet. Here is my dmesg (fresh after reboot)

https://pastebin.com/raw/p8nSr4xj
and .config

https://pastebin.com/raw/CnyLtudK

I see intel-lpmd is not in the main gentoo repo, I'll see if it is in some overlay

[pingtoo]

[

Sorry I think I lost track on the conversation on the "hfi".

Once you have intel-lpmd (modern standard for Meteor Lake and newer) or thermald (v2.5.4+) update support Meteor Lake. They can trigger HFI activation.

We never review your .config file do you mind share it? In my research there are too many different CONFIG_something need to be on/off, so instead we come back and forth about is this on and is that on too kind of questions may be I will just search myself and let you know if anything need to change.

Also if possible share the dmesg with initial boot sequence, same reason as above, many research note also mention that if you have this CONFIG_something you should see "something" in dmesg.

Yep, 'hfi' remained uninvestigated yet. Here is my dmesg (fresh after reboot)

https://paste.ubuntu.com/p/BbKPZgMss9/

and .config

https://paste.ubuntu.com/p/3MJcPpCJyJ/

I see intel-lpmd is not in the main gentoo repo, I'll see if it is in some overlay

Is it possible to use a pastebin that not require login?

[dmpogo]

Is it possible to use a pastebin that not require login?

Oops, I never used pastebin before, and just went with 'pbput file' from pastebinit. Let me investigate

Ok, I have replaced the links in the original comment. Seems to work for me

[pietinger]

dmpogo,

I have rarely seen such a good configuration. It includes almost everything you need (and some unnecessary drivers as modules). I only have the following comments:

Code: Select all

1.
# CONFIG_HIGH_RES_TIMERS is not set
2.
CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL=y
3.
CONFIG_VMD=m
4.
# CONFIG_PCIE_DW_PLAT_HOST is not set
5.
# CONFIG_ATA is not set

1. I stongly recommend to enable it. Then you will be able to enable also CONFIG_SND_HRTIMER=y (I also suggest)

2. Intels P-State's internal governor choose POWERSAVE when you set SCHEDUTIL as default. Here you can change it to PERFOMANCE.

3. You dont have it .. but ...

4. You have a Designware PCIe controller. Maybe do: https://wiki.gentoo.org/wiki/User:Pieti ... controller

5. I was surprised that I couldn't find CONFIG_SENSORS_DRIVETEMP in your config. Then I looked into the dependencies:

Code: Select all

Depends on: HWMON [=y] && SCSI [=y] && ATA [=y]

Yes, you have no SATA device and wouldnt need it. But you can enable ATA without enabling any option in this submenu, then you wil be able to enable CONFIG_SENSORS_DRIVETEMP. I suggest it because also NVMEs have temperature sensors (here is mine):

Code: Select all

# sensors
...
nvme-pci-e100
Adapter: PCI adapter
Composite:    +30.9°C  (low  = -273.1°C, high = +81.8°C)
                       (crit = +84.8°C)
Sensor 1:     +30.9°C  (low  = -273.1°C, high = +65261.8°C)
Sensor 2:     +33.9°C  (low  = -273.1°C, high = +65261.8°C)

TBH I think the dependency in this option is wrong. IMHO it should be "HWMON and SCSI and (ATA or NVME)" ... but enabling ATA does not hurt.

(I have not checked NET, SND, and everything after DEVICE)


Internal note (because pastebin services delete it after some time):

Code: Select all

[    0.000000] DMI: LENOVO 21KC004AUS/21KC004AUS, BIOS N3YET80W (1.45 ) 08/14/2025

[    0.108146] hpet: HPET dysfunctional in PC10. Force disabled.

[    0.113921] CPU0: Thermal monitoring enabled (TM1)

[    0.117193] smpboot: CPU0: Intel(R) Core(TM) Ultra 7 165U (family: 0x6, model: 0xaa, stepping: 0x4)

[    0.127279] Memory: 32301112K/32990300K available (12347K kernel code, 2195K rwdata, 3512K rodata, 1304K init, 1020K bss, 667048K reserved, 0K cma-reserved)

[    0.913522]  nvme0n1: p1 p2 p3 p4

[    0.914294] proc_thermal_pci 0000:00:04.0: enabling device (0000 -> 0002)
[    0.914542] intel_rapl_common: Found RAPL domain package
[    0.916253] intel_pstate: Intel P-state driver initializing
[    0.917463] intel_pstate: HWP enabled

[    0.918708] pci 0000:00:08.0: Setting to D3hot
[    0.929604] pci 0000:00:0b.0: Setting to D3hot
[    0.943043] intel_pmc_ssram_telemetry 0000:00:14.2: enabling device (0000 -> 0002)
[    0.943314] intel_vsec 0000:00:0a.0: enabling device (0000 -> 0002)
[    0.943506] intel_rapl_msr: PL4 support detected.
[    0.943538] intel_rapl_common: Found RAPL domain package
[    0.943559] intel_rapl_common: Found RAPL domain core
[    0.943570] intel_rapl_common: Found RAPL domain uncore
[    0.943585] intel_rapl_common: Found RAPL domain psys

[    0.947388] microcode: Current revision: 0x00000025
[    0.947399] microcode: Updated early from: 0x00000024

[    0.963849] intel_pmc_core INT33A1:00:  initialized

[    1.603269] intel_vpu 0000:00:0b.0: enabling device (0000 -> 0002)

[    1.629597] Loading firmware: intel/vpu/vpu_37xx_v1.bin

[dmpogo]

Thanks ! I spent some time trimming down the configuration from the distribution kernel one. My first try was to take my old config from previous thinkpad, but it did not get me working microphone (the rest seemed fine)
So I installed distribution kernel, which had everything working (great job !) and trimmed it down. Still did not find what switched the microphone off I know that there are some modules which are still not needed,
but stopped for the moment.

BTW, sensors from lm_sensors give me

Code: Select all

ucsi_source_psy_USBC000:001-isa-0000
Adapter: ISA adapter
in0:           0.00 V  (min =  +0.00 V, max =  +0.00 V)
curr1:         0.00 A  (max =  +0.00 A)

thinkpad-isa-0000
Adapter: ISA adapter
fan1:           0 RPM
fan2:           0 RPM
CPU:          +42.0°C  
GPU:              N/A  
temp3:         +0.0°C  
temp4:         +0.0°C  
temp5:         +0.0°C  
temp6:         +0.0°C  
temp7:         +0.0°C  
temp8:            N/A  
pwm1:            128%

coretemp-isa-0000
Adapter: ISA adapter
Package id 0:  +46.0°C  (high = +110.0°C, crit = +110.0°C)
Core 0:        +39.0°C  (high = +110.0°C, crit = +110.0°C)
Core 1:        +39.0°C  (high = +110.0°C, crit = +110.0°C)
Core 2:        +39.0°C  (high = +110.0°C, crit = +110.0°C)
Core 3:        +38.0°C  (high = +110.0°C, crit = +110.0°C)
Core 4:        +40.0°C  (high = +110.0°C, crit = +110.0°C)
Core 5:        +40.0°C  (high = +110.0°C, crit = +110.0°C)
Core 6:        +40.0°C  (high = +110.0°C, crit = +110.0°C)
Core 7:        +40.0°C  (high = +110.0°C, crit = +110.0°C)
Core 8:        +43.0°C  (high = +110.0°C, crit = +110.0°C)
Core 12:       +36.0°C  (high = +110.0°C, crit = +110.0°C)
Core 32:       +42.0°C  (high = +110.0°C, crit = +110.0°C)
Core 33:       +42.0°C  (high = +110.0°C, crit = +110.0°C)

acpitz-acpi-0
Adapter: ACPI interface
temp1:        +42.0°C  

ucsi_source_psy_USBC000:002-isa-0000
Adapter: ISA adapter
in0:           5.00 V  (min =  +5.00 V, max = +13.20 V)
curr1:         3.00 A  (max =  +3.21 A)

iwlwifi_1-virtual-0
Adapter: Virtual device
temp1:        +28.0°C  

nvme-pci-0400
Adapter: PCI adapter
Composite:    +30.9°C  (low  = -20.1°C, high = +77.8°C)
                       (crit = +81.8°C)
Sensor 1:     +30.9°C  (low  = -273.1°C, high = +65261.8°C)

BAT0-acpi-0
Adapter: ACPI interface
in0:          13.34 V  
power1:        0.00 W  

acpi_fan-isa-0000
Adapter: ISA adapter
fan1:           0 RPM

so sensors on nvme are picked up, even apparently without CONFIG_SENSORS_DRIVETEMP


1. I'll do that. I am always confused between hpet and high res timerrs, and given that dmesg resported disfunction hpet I probably switched them off.

2. I guessed as much following our yesterday discussion ! It is a bit fun that intel_pstate powersave is schedutil, while powersave in config has different meaning

3. Did not pay attention to that

4. Good point, that Designerware was appearing here and there, and I was not sure what part it is.

[pietinger]

[...] so sensors on nvme are picked up, even apparently without CONFIG_SENSORS_DRIVETEMP

Great ... I didnt knew that this works without SENSORS_DRIVETEMP (so, maybe it is really only necessary for SATA harddisks/SSDs and not for M.2 SSDs; then the dependency is correct). Thank you!

A word to the option CONFIG_INTEL_HFI_THERMAL.

1. You must know this:

Starting with Linux 6.10, the kernel is designed to automatically disable HFI if no userspace application is "listening" for its updates. This change was made because keeping HFI active without a userspace consumer was found to waste a significant number of CPU cycles.

2. So, you will need also the Intel Low Power Mode Daemon (LPMD). If you read the manapge of it, you maybe see that you dont want it (because you want performance and not energy saving) -> https://www.mankier.com/8/intel_lpmd

3. There must be a good reason that this package is not in our Gentoo repository ... maybe because of some open problems? -> https://github.com/intel/intel-lpmd/issues

[pingtoo]

if intel-lpmd does not work will you can try sys-power/thermald it is at 2.5.9 so it should have the necessary update to support Meteor Lake.

for thermald you might need "--adaptive" for daemon start up argument. however this setting use manufacturer's thermal limit (from DPTF table) so it is possible that being laptop it might limit performance.

I see that your machine does have DPTF table.

Code: Select all

[    0.000516] ACPI: SSDT 0x000000004E0A4000 003C93 (v02 LENOVO DptfTabl 00001000 INTL 20210930)

So for experimenting you could also try to remove the "--adaptive", However this might cause hardware damage so be careful and remember to turn it back on once finish the experiment.

Once you got thermald running you can search dmesg for "intel_hfi" which is the driver that gave feedback to the intel-pstate driver.

P.S. thanks for the update. I am able to get the config and dmesg.

[dmpogo]

[...] so sensors on nvme are picked up, even apparently without CONFIG_SENSORS_DRIVETEMP

Great ... I didnt knew that this works without SENSORS_DRIVETEMP (so, maybe it is really only necessary for SATA harddisks/SSDs and not for M.2 SSDs; then the dependency is correct). Thank you!

I suspect nvme sits directly on PCI bus and that's where "PCI adapter" reports the sensors. While for SCSI and ATA sensors are on the harddrive itself, and you need a separate driver to talk to them via SCSI/SATA connection ?

A word to the option CONFIG_INTEL_HFI_THERMAL.

1. You must know this:

Starting with Linux 6.10, the kernel is designed to automatically disable HFI if no userspace application is "listening" for its updates. This change was made because keeping HFI active without a userspace consumer was found to waste a significant number of CPU cycles.

2. So, you will need also the Intel Low Power Mode Daemon (LPMD). If you read the manapge of it, you maybe see that you dont want it (because you want performance and not energy saving) -> https://www.mankier.com/8/intel_lpmd

That in all probability answers why I do not see corresponding /sys entries. I'll play with lpmd and thermald to see if they are useful.

[pingtoo]

Note: You have multiple thermal governor build into kernel, many of them could throttle power which might impact on the performance experiment.

to check which is your current default

Code: Select all

cat /sys/class/thermal/thermal_zone0/policy

to see all zones at once

Code: Select all

paste <(cat /sys/class/thermal/thermal_zone*/type) <(cat /sys/class/thermal/thermal_zone*/policy)

to temporary manual change

Code: Select all

grep "x86_pkg_temp" /sys/class/thermal/thermal_zone*/type
# so change the X to what every numver return by above
echo "user_space" | sudo tee /sys/class/thermal/thermal_zoneX/policy

[pietinger]

dmpogo,

I checked my notices (before throwing away) ... and saw that I had forgotten something to write here ... please enable also:

Code: Select all

# CONFIG_INTEL_IOMMU_DEFAULT_ON is not set

(See also: https://wiki.gentoo.org/wiki/User:Pieti ... rnel#IOMMU )

[dmpogo]

dmpogo,

I checked my notices (before throwing away) ... and saw that I had forgotten something to write here ... please enable also:

Code: Select all

# CONFIG_INTEL_IOMMU_DEFAULT_ON is not set

(See also: https://wiki.gentoo.org/wiki/User:Pieti ... rnel#IOMMU )

Right, thanks

[dmpogo]

Note: You have multiple thermal governor build into kernel, many of them could throttle power which might impact on the performance experiment.

to check which is your current default

Code: Select all

cat /sys/class/thermal/thermal_zone0/policy

to see all zones at once

Code: Select all

paste <(cat /sys/class/thermal/thermal_zone*/type) <(cat /sys/class/thermal/thermal_zone*/policy)

to temporary manual change

Code: Select all

grep "x86_pkg_temp" /sys/class/thermal/thermal_zone*/type
# so change the X to what every numver return by above
echo "user_space" | sudo tee /sys/class/thermal/thermal_zoneX/policy

Currently, it is "step-wise"

[bagas]

I still don't understand how the new processors with P-cores, E-cores, LP-E-cores will work on Linux.
After purchasing a processor, will I need to configure the task scheduler for the cores?

[pietinger]

I was surprised (and annoyed) when I read this:
https://www.phoronix.com/news/Intel-P-S ... ify-Energy

It seems that Intel prioritizes energy efficiency over performance.

[dmpogo]

I was surprised (and annoyed) when I read this:
https://www.phoronix.com/news/Intel-P-S ... ify-Energy

It seems that Intel prioritizes energy efficiency over performance.

It is simply wow, basically let us make Celeron like machine with 1Ghz CPU's out of expensive laptops. And my current experience seems to confirm this tendency already in the currrent intel p_states.