Parpool thread-based pool size

I have a computer with 14 physical cores: 12th Gen Intel(R) Core(TM) i7-12800H, 2400 Mhz, 14 Core(s), 20 Logical Processor(s)).

I can set

p=parpool('processes',8)

and I get what I expect

p = 

ProcessPool with properties: 

Connected: true
NumWorkers: 8
Busy: false
Cluster: processes (Local Cluster)
AttachedFiles: {}
AutoAddClientPath: true
FileStore: [1x1 parallel.FileStore]
ValueStore: [1x1 parallel.ValueStore]
IdleTimeout: 30 minutes (30 minutes remaining)
SpmdEnabled: true

However, if I type

delete(p)
p=parpool('threads',8)

I get this error message:

Error using parpool (line 108)
A minimum pool size of 8 was requested. The maximum thread-based pool size is currently 6.

Is there a way to increase the maximum thread-based pool size above 6 (given that my Intel CPU has 14 physical cores)?

Thanks!

Answer 1

[Walter Roberson]

With that particular number of physical cores (14), chances are high that you have a mix of "performace" cores and "efficiency" cores. There is currently a restriction, that thread-based pools can only use performance cores.

[Georgi]

Walter, I have a problem that the regressionLearner tool from the (Statistics and) Machine Learning Toolbox starts a 'Threads' PArallel Profile, even when 'Processes' Parallel Environment has been chosen as Preferred (Default). I want to be able to force the regressionLearner tool to start a 'Processes' parallel pool with the max. number of Workers possible (16 logical cores here in a Gen9 Core i9 CPU).

[Walter Roberson]

(Checking, I find that Performance Cores vs Efficiency Cores was something introduced in Gen 12, so performance cores is not a relevant factor for @Georgi question about Gen9 Core i9.)

[Walter Roberson]

@Georgi

There does not appear to be any way to select process based instead of thread based.

[Georgi]

@Walter Roberson

Thank you. I found that when executing

maxNumCompThreads(16); % 16 Logical cores in a Gen9 Core i9 CPU

parpool('Processes',maxNumCompThreads);

before I start the regressionLearner model training, the tool uses the parallel pool already created with the 'Processes' profile.

But it turned out that numWorkers was not the bottleneck in my calculations, and even those 8 Workers were far from fully used (CPU load was barely 10-20%). Strangely, in R2024aU7, the regressionLearner tool uses at ~12-20% one of the DMA Engines (shown as 'Copy 1' in Windows Task Manager) of my nVidia Quadro RTX3000, and this at 0% 3D load on that GPU (well, most of the time). In R2025aU1 on the same laptop (128 GB or RAM), that GPU is not used at all, and the same model training (and iterations number) takes notably longer time. On a server with 36 logical cores per CPU and no GPU accelerator, in R2022a, the same model training takes 5.3 times longer (!!!).