Large broadcast variable in parfor loop

Hi

Im having a for loop that i want to parallelize. In the loop th essential part is an interpolation. The problem is that the interpolant is huge, (V has 2 to 4 GB). Once I parallelize the loop, this read-only broadcast variable is multiplied in the memory and i run out of memory.

Is there a way to avoid this? i.e. to have the speedup benefits of parallalizing without the need to duplicate the huge interpolant in the memory?

IP=scatteredinterpolant(w,x,y,z,V)
parfor i=1:N
  user_function( parameters, IP(ww(i),xx(i),yy(i),zz(i))  )
end

Thanks Dominik

Answer 1

[Edric Ellis]

Unfortunately, because PARFOR supports operation across multiple machines and uses multiple MATLAB processes, I'm afraid there's no way to avoid having broadcast data replicated across the workers.

If you're using only local workers, you might be able to use this File Exchange submission: http://www.mathworks.com/matlabcentral/fileexchange/28572-sharedmatrix

[dominik]

Thanks Eric, this seems to be what i need (if this allows me to send multiple interpolants to the shared memory). Never having used mex files I have problems getting it running so far, but I guess once working it will do the trick for me. It's a pity this functionality isnt included in matlab as an option for local parallelisation...Thanks!

Answer 2

[dominik]

I got it running.... but it seems this submission does only work for cells and matrixes. When I try to share a griddedInterpolant I get an error.