Parfor becomes serial half-way through computation

I am running a parfor loop with 9 iterations. My laptop has 6 physical cores.

At first it runs 6 jobs in parallel, as expected. Once that is done, it begins to serially run the remaining 3 jobs 1-by-1 rather than running them on 3 cores in parallel (while idling the other 3).

Why is it not running the last 3 jobs in parallel on 3 cores?

Thank you.

Answer 1

[Edric Ellis]

parfor uses heuristics to try and divide up loop execution into sub-ranges to maximise worker utilisation, but it sounds like the default strategy isn't working well for you. This can happen when you have a smallish number of loop iterations. Here's what I tried in R2019a on my 6-core machine, both in default mode and using parforOptions (new in R2019a) to explicitly control the subrange selection:

%% Get or create a parallel pool
pool = gcp();
if isempty(pool)
    pool = parpool('local', 6);
end

%% Run a |parfor| loop, and time it.
% Here we're simply using the default loop division strategy.
t = tic();
parfor idx = 1:9
    pause(1);
end
elapsedDefault = toc(t)

%% Use |parforOptions|
% Here we're forcing the loop to be divided into exactly 9 subranges
% each of 1 iteration
t = tic();
opts = parforOptions(pool', 'RangePartitionMethod', 'fixed', ...
    'SubrangeSize', 1);
parfor (idx = 1:9, opts)
    pause(1);
end
elapsedFixed = toc(t)

For me, I got the following results showing that in fact the default strategy seems to work OK...

elapsedDefault =
    2.0460
elapsedFixed =
    2.0275