parfor - Attempt to serialize data which is too large

I'm working on a large data set that is is stored as a cell array X, where each cell contains an array of about 2MB, totaling 1GB. I'm running it in parallel on 50 cores that are each on separate nodes. I get the error "Attempt to serialize data which is too large" when I try to run the whole array . When I run a small array X I don't get the error. I check on the nodes and there's plenty of memory left, it seems the bottleneck is data transmission

I'm using the following code

Y1= cell(1,N1);
Y2 = cell(1,N1);

parfor i = 1:N1                 
        XX = X{i};                  
        Np = size(XX,1);            
        Y1_temp = zeros(400, Np-1);
        Y2_temp = zeros(400, Np-1);

          for j= 1:(Np-1)           

              Y1_temp(:,j) = someFunction;
              Y2_temp(:,j) = someOtherFunction;
          end

          Y1{i} = Y1_temp;
          Y2{i} = Y2_temp;
  end

I don't understand why this is, has anyone ever got this error? I couln't find anything online. I also tried writing it as an spmd block where I operate on the local worker data but I get the same error. Any help will be very appreciated.

Regards Germán

Answer 1

[Edric Ellis]

The 2GB PARFOR transmission limitation was removed from Parallel Computing Toolbox in R2013a.

Answer 2

[Konrad Malkowski]

There is a 2GB transmission limit between client and workers. Could you provide me with additional values for Np, N1, Nt, etc so that I have a better picture of which variables are causing the problem?

[Germán]

sure, the values are usually Np=100000, N1=Nt=512.

My understanding is that only the sliced portions of X, Y1, Y2 would be transmitted in each iteration, is that correct? If so the the size of a single cell in those variables is never above 2 MB.

Thanks for your help!

Answer 3

[James]

I have the same problem. I think it is a common problem, but I couldn't find a solution for this. If it is the 2GB transmission limit, then how can we get around this then?

[Germán]

I got around it using normal arrays instead of cell arrays since there was a bug with cell arrays, but the issue might have been solved in the latest releases.

[Edric Ellis]

The 2GB PARFOR transmission limitation was removed from Parallel Computing Toolbox in R2013a.