How to avoid memory leaks when function inside parfor generates warning?

Hello!

I use matlab R2015b for Win Server 2012 OS with local parpool. I have memory leaks on local workers when function inside parfor-loop generate warning. Also memory is not released after parfor-loop ends. Example:

x = rand(78, 1);
for j = 1:100
    parfor i = 1:1000
        [~, z] = kpsstest(x);
    end
end

[~, z] = kpsstest(x) generates warning:

Warning: Test statistic #1 below tabulated critical values:
maximum p-value = 0.100 reported. 
> In kpsstest>getStat (line 632)
  In kpsstest (line 291)

Memory leaks on workers even when i use constructions:

warning('off', 'econ:kpsstest:StatTooSmall');
pctRunOnAll warning('off', 'econ:kpsstest:StatTooSmall')

Memory on start workers:

Memory on after executing code:

How to avoid memory leaks when function inside parfor generates warning?

Thank you.

Answer 1

[Edric Ellis]

This problem is the subject of this bug report. If you are able to upgrade to R2017a, that's the simplest workaround.

[Jonas]

Thank for your reply. I will test it soon as possible.

[Isabel]

Hello Jonas, were you able to fix your issue? I'm running my code on R2017b, and I'm facing the same issue. All the help will be most welcome.

[Jonas]

Hello Isabel.

As I remember, in my case the solution was to divide the calculations into batches and after calculating one data batch, the parpool stopped and then restarted again (see pseudocode below).

for i = 1:numel(batches)

      parpool('local', 'SpmdEnabled', false);

         curData = batches{i};

         parfor j = 1:numel(curData)
             % DO SEMETHING THAT LEADS TO MEMORY LEAKAGE 
         end

      delete(gcp('nocreate'));

end

It was a bad idea in point of performance, but it was quite stable. I hope it helps.

[xl z]

MATLAB 2023a/2023b also has such bug