Segmentation error during parfor loop (remote linux server)

I am using a Linux server with 64 GB RAM and 16 CPUs to run multiple repetitions of a MATLAB function in parallel (by using the "parfor" command); let's define this function as [Output]=myfunc(a,b,c). In particular, for "n" iterations, I have to calculate the Output of a "population" of 16 "agents", in which each agent has slightly different values of a,b,c.

Specific agents (i.e values of a,b,c) cause a worker to crash due to segmentation violation (I attach the related crash dump file).

If I run only one of these agents in a parallel pool with just one worker I get the same outcome, whereas if I don't use "parfor" the code works without any problem. Moreover, if I perform the same simulation on my Windows PC I get no error even when using the parallel pool.

Answer 1

[Daniel Duarte]

Hello Stefano,

The main requirement of parfor is to allow Matlab to work with sections of the problem, in a vectorised fashion. Given a population of agents pop(i), we need to vectorise the output as output(i). This way we allow the parfor loop to work in parallel on each section of the problem.

A possible parfor application would be:

% initialize inputs and output
Output=zeros(16,1);
for i=1:16
    pop(i).a = rand;
    pop(i).b = rand;
    pop(i).c = rand;
end

% implement parfor
parfor i=1:16
   Output(i)=sumABC(pop(i).a,pop(i).b,pop(i).c) 
end

Output

function sum = sumABC(A,B,C)
    sum = A+B+C;
end

See more details at:

https://www.mathworks.com/help/parallel-computing/parfor.html

Answer 2

[Edric Ellis]

The crash dump suggests the workers are crashing inside ipqpdense. This problem should be fixed if you upgrade to R2020a or later.