Parfor number of workers, CPUs, how are they related? What are the limits?

Hi,

Using parfor specifying the number of parallel workers, how does it relate to the number of CPUs?

parfor (loopVar = initVal:endVal,M); statements; end

Where M specifies the Maximum number of workers running in parallel, as both a cap, and a limit.

It may seem like a very obvious question and answer: is the number of workers equal to the CPUs available? That is, on a quadcore computer, the maximum number of workers would be 4?

A separate question, I've been advised I can set some settings using this written in somewhere before the parfor coding:

tempdir = getenv('TMPDIR')
c = parcluster
c.JobStorageLocation = tempdir
c.NumWorkers=8
parpool(c.NumWorkers);

Would this specify the pool to 8 workers? In which case the code below for parfor would cap at 8 even if M is specifed as 10? What would happen if M was 4? Would it be capped at 4 or 8?

parfor (i=1:10,10)
...
end

parfor (i=1:10,4)
...
end

And where should theTMPDIR in tempdir = getenv('TMPDIR') be located?

Thanks a lot!

Answer 1

[Raymond Norris]

A quadcore computer should max at 4 workers. We don't suggest throttling it to include hyperthreads. With that said, let's say you have 4 additional HT and you start a pool of 8 workers, perhaps you get even a 10% increase (for example), then maybe it's worth it. You just shouldn't expect an 8x improvement.

Setting the JobStorageLocation is only a suggestion if you are running MATLAB jobs on an HPC cluster and in some sort of rapid batch mode. For example, you write a PBS job script to run MATLAB code that uses a local pool. And then you submit a slew of those at once. In that case, setting the JSL can be helpful (due to a race condition).

The number of workers used is the min(M,size-of-pool). In your example, that would be 8 and 4 respectfully.

[In-chan Kim]

Thanks very much Raymond!

If I understand the key points correctly:

• The number of CPUs equals max number of works possible without throttling it.
• JobStorageLocation only for HPC batch runs
• the coding of Parfor (i=1:10,M) means the parfor uses the lower value of M, or the actual number of workers set using NumWorkers. So that would mean it is all right to set M to be a ridiculously high number, and just let c.NumWorkers be the actual limit?

So, I am running this on a HPC with an array job submission.

Where should theTMPDIR in tempdir = getenv('TMPDIR') be?

I'm really not very good with the computing environment, and coding, and directories and all that.

[Raymond Norris]

Correct, you could set M to a high number. However, it's really designed more for setting it to be lower than the size of the pool. Otherwise, there's no need to set it. That is

parpool('local',8);
parfor (i = 1:10,100000)

Is the same as just

parpool('local',8);
parfor i = 1:10

I'd just assign JSL to tempname. For example

c = parcluster;
jsl = tempname;
[FAILED,emsg,eid] = mkdir(jsl);
if PASSED==false
    error(eid,emsg)
end
c.JobStorageLocation = jsl;
c.NumWorkers = 8;
c.parpool(c.NumWorkers);

tempname will generate a temporary name of the directory to use.

[In-chan Kim]

Brilliant, thanks very much Raymond! I really appreciate it!