Why parfor loop memory keep increasing?

I am using R2018b on Windows 10. My code have 3 for loops and I want to use parfor above them. In each parfor loop, it is expected to save a 2D array of size N. I want to keep N around 500. The code is given below

clear; clc;
N = 500;

Ts = 0.1:0.1:10; %range of outter most "for" loop

xmin = -pi; xmax = pi; %range of first "for" loop
ymin = -pi; ymax = pi; %range of second "for" loop
wmin = 3;   wmax = 15; %range of "parfor" loop

% building x,y and w arrays
dx = (xmax-xmin)/(N-1);
dy = (ymax-ymin)/(N-1);
dw = (wmax-wmin)/(2*N-1);

xs = xmin : dx : xmax;
ys = ymin : dy : ymax;
ws = wmin : dw : wmax;
L = length(ws); %length of parfor loop

%allocating memory for final results:
D3_I1 = zeros(1,length(Ts));
D3_I2 = zeros(1,length(Ts));
D3_I3 = zeros(1,length(Ts));
D3_I4 = zeros(1,length(Ts));


nT = 1;
for T = Ts
    T
    %allocating memory for parfor loop results:
    D2_I1 = zeros(1,L);
    D2_I2 = zeros(1,L);
    D2_I3 = zeros(1,L);
    D2_I4 = zeros(1,L);
    
    parfor nw = 1:L
        w = ws(nw);
        
        % allocating memory for "for" loop results:
        R1 = zeros(N,N);
        R2 = zeros(N,N);
        R3 = zeros(N,N);
        R4 = zeros(N,N);
        
        ny = 1;
        for ky = ys
            nx = 1;
            for kx = xs
                
                %following are the main calculations A1,2,3,4.
                %in my original code these are actually
                %some 3-by-3 matrices, but the final value of A1
                %A2,A3 and A4 is just a complex number. this part is 
                %not very lengthy. it's around 12-15 expressions which
                %include matrix multiplications, inverse and trace
                %operations on (3by3) matrices.
                
                
                A1 = x*y;
                A2 = x*y;
                A3 = x*y;
                A4 = x*y;
                
                R1(nx,ny) = A1;
                R2(nx,ny) = A4;
                R3(nx,ny) = A3;
                R4(nx,ny) = A4;
                
                nx = nx + 1;
            end
            ny = ny + 1;
        end
        
        D2_I1(nw) = sum(R1(:))*dx*dy;
        D2_I2(nw) = sum(R2(:))*dx*dy;
        D2_I3(nw) = sum(R3(:))*dx*dy;
        D2_I4(nw) = sum(R4(:))*dx*dy;
        WSaved(nw) = w;
    end
    
    D3_I1(nT) = sum(D2_I1)*dw;
    D3_I2(nT) = sum(D2_I2)*dw;
    D3_I3(nT) = sum(D2_I3)*dw;
    D3_I4(nT) = sum(D2_I4)*dw;
    
    nT = nT+1;
end

The memory just keep increasing. I have 128GB with 28 cores. I am using 22 cores for parfor. I was able to run it till T = 5. But then I get "memory error".

Why does parfor memory keep increasing? Should not it be resetted for every next T?

Answer 1

[Raymond Norris]

Hi Luqman,

Two suggestions:

1. Look at ticBytes/tocBytes to see how much data is being passed through to parfor
2. If you're able to upgrade to R2020a, consider using a threads pool, rather than a local pool. See the following for more information on threads vs local

https://www.mathworks.com/help/parallel-computing/choose-between-thread-based-and-process-based-environments.html

Thanks,

Raymond

[Luqman Saleem]

Dear Raymond,

Thank you for reply. I just upgraded to R2020a and now memory problem is resolved.

[Raymond Norris]

Good to hear, Luqman. I'm curious, did you stick with the local pool or switch to the threads pool?

[Luqman Saleem]

I used local pool. Honestly speaking, I am not an expert with MATLAB and I have no idea how to use thread pool.

[Raymond Norris]

There are some limitations with threads pool (see the doc for more information), but you start and use a threads pool the same way as a local pool (see the doc for advantages of threads pool).

Good to hear you're in your good shape.