parpool sorting out the default number of workers

I'm a little bit lost in the default number of workers when parpool() is invoked. My PC is standalone with Intel Processor and Windows-based with MATLAB R200B.

• In the Preferences Parallel Computing Toolbox window, both Porcesses and Threads profiles indicate the default value of number of workers is 12
• My processor has 14 cores; 6 performance, 6 Efficient, and 20 (hyper) threading according to Intel link

The command maxNumCompThreads returns 6

>> maxNumCompThreads('automatic')

ans =

6

The command parpool returns 14

>> delete(gcp('nocreate'))
>> p = parpool('Threads')
Starting parallel pool (parpool) using the 'Threads' profile ...
    Connected to the parallel pool (number of workers: 14).

p = 

ThreadPool with properties: 

NumWorkers: 14
Busy: false

The questions are

• Why 14 and not 12 as setup by Preferences?
• How parpool and Preferences setup those numbers?

In general it's all over the place to me, if someone can shed the light it would be great.

Answer 1

[Alison Eele]

Thanks for the follow on questions Bruno!

Preferred Number of Workers in a Pool

For parpool('processes') the number of workers you will get will be the minimum of: the parallel preference; or the NumWorkers field in the 'processes' cluster profile. In your case I would expect it to open a parallel pool of 12 workers [min(12,14)].

The default value of 12 in the Parallel Preferences will only apply to non-thread pools. This stems from Thread pools historically following the value of maxNumCompThreads() thus this preference never applied. From R2022b you can now specify a size of thread pool and parpool('threads') defers to the same default number of workers as the 'processes' profile i.e. the number of physical cores detected.

maxNumCompThreads behaviour

The value returned by maxNumCompThreads('automatic') is counting just your performance cores (6) rather than performance + efficiency (14). This is because maxNumCompThreads primarily focuses on the implicit multi-threading of mathematical operations which should take place on the Performance cores of your processor. Adding additional threads would either cause resource contention on the Performance cores or cause them to run on some of the efficiency cores which could incur some slowdown on the overall operation.

Standalone Deployed Application

When you deploy onto another machine using the 'processes' or 'local' profile it by default will use a number of cores = to the number of physical cores detected on the machine. I would recommend finding out how many workers the cluster profile thinks it has by instantiating a cluster object then querying the number of workers:

c = parcluster('processes');
numcores = c.NumWorkers

This would be safer than using an undocumented feature command.

[Bruno Luong]

Many thanks @Alison Eele for clearing hings out.

Answer 2

[Raymond Norris]

Hey @Bruno Luong, per your link, you have 8 Efficient cores, hence 14 (6+8) cores. For R2022b (which I think is the version you're using), the Threads profile doesn't allow for less than the number of cores you have. The Process profile uses the standard min(12,#cores).

For Performance vs Efficient, can you run the following and post the answer?

feature numcores

[Edric Ellis]

There was a typo in @Raymond Norris's code (which I'll correct) - feature numcores is the thing.

[Bruno Luong]

Thanks.

>> feature numcores
MATLAB detected: 14 physical cores.
MATLAB detected: 20 logical cores.
MATLAB was assigned: 20 logical cores by the OS.
MATLAB is using: 14 logical cores.
MATLAB is not using all logical cores because hyper-threading is enabled.

ans =

14

So can I conclude

• the default value (12) in the preference GUI of the Threads/Proceeses profile is ignored by parpool?
• What is the value returned by maxNumCompThreads('automatic') (6)?
• If I deploy a standalone application via MATLAB compiler toolbox on another platform what command should I use then to get a right number of cores for proper configuring parpool (I don't want the default value since there are other apps runing in concurrent with MATLAB)? I tend to use this:

    numcores = feature('numcores')