MATLAB: Memory Parfor Needs more memory than normal

MATLAB Parallel Serverparallel computingram

Hello, i need your help. I dont understand something with parfor loop.

I load an image which is 8GB.

I use parfor loop with 4 threads.

So the total usage of Ram it had to be 8GB + 4thread*8GB = 40GB.

But when the parfor loops starts, the momory increases more than 64GB (this is my total memory)

everything lags, and after some minutes the memory goes back to 40GB and the program runs normally.

I dont understand. Why does it happen? Is there any possible solution? If it needs 40GB why it gets more memory in the beggining?

For bigger data or if i increase the processors i get a blue screen. Pls help!

Best Answer

When you transfer data to the workers for a parfor loop, there is a transient increase in memory over and above what you'd expect due to the way data is transferred. You can perhaps work around this by loading the image data directly on the workers. One way to do this is by using parallel.pool.Constant. When constructing your Constant, you can pass in a functino handle to load the data, and this gets executed on the worker. Something like this perhaps:

c = parallel.pool.Constant(@() imread('myLargeImage.png'));
parfor idx = 1:N
    im = c.Value; % Get the actual image from within the Constant
    out(idx) = myImageProcessingFunction(im, idx); % or whatever.
end

Related Solutions

MATLAB: Does parallel.pool.const create a copy of the variable in memory for each worker sequentially instead of in parallel

I suspect you're creating the parallel.pool.Constant using data created on the client. It's much more efficient to have the workers create the data, if possible. Consider two cases:

% Case 1: data created on the client
parallel.pool.Constant(ones(1e4));
% Case 2: use the Constant constructor with a function handle to create
% the contents directly on the worker
parallel.pool.Constant(@() ones(1e4));

This results in the following memory usage pattern. In the screen-shot, case 1 is indicated with a red arrow, and case 2 with a green arrow.

As you can see, case 2 happens in parallel, and avoids the data transfer from the client to the workers (it's the data transfer that really causes the lack of parallelism).

If you really cannot create the data on the workers, you can use the parallel.pool.Constant constructor that accepts a Composite, like this:

% Build an empty Composite
c = Composite();
% Transfer the data from client only to worker 1
c{1} = ones(1e4);
c(2:end) = {[]};
spmd
  % Use labBroadcast to copy data to all workers (labBroadcast
  % is more efficient than the client/worker communication)
  c = labBroadcast(1, c);
end
% Build the Constant from the Composite
c = parallel.pool.Constant(c);
% Flush memory on the workers by executing an empty SPMD block
spmd, end

MATLAB: Using all cores with parfor

And because you don't have the Parallel Computing Toolbox, parfor will run serially anyway. There is more overhead with parfor so I would recommend against using it if you don't have Parallel Computing Toolbox.

You can contact the friendly account manager for your organization to get a trial of the Parallel Computing Toolbox if it interests you.

Best Answer

Related Solutions

MATLAB: Does parallel.pool.const create a copy of the variable in memory for each worker sequentially instead of in parallel

MATLAB: Using all cores with parfor

Related Question