multi-level parallelism in Matlab (2013a)?

Hello,

I am trying to solve a problem of efficient calculation of optical spectra (of some structures). The calculation process itself is represented by tens or hundred independent calculations (according to number of points in the optical spectra, number of wavelengths). The spectral interval is just a vector.

Until now I used parfor loop. I took the interval, split it to several subintervals (according to number of workers in matlabpool) and send it to workers by parfor loop. So, the subintervals were calculated in parallel. Finally I just glued partial results together. I choose "divide et impera" approach to reduce communication between nodes at cluster. If all subintervals have the same length and all nodes are free, the computation time of each subinterval problem is the same(almost).

But now, my problem increases in one dimension. I would like to compute several spectral problem in parallel. For exaple: 4 problems, each for the spectra with 1000points. What I would like to implement is code, which starts with 4 parallels workers (each for one problem). Than each worker solve its partial problem in parallel (lets say on 8 cpus). So finally I will use 1(initial master) + 4(submasters) + 4*8 = 37cpus (strange number, but it is just example).

So, my question is: Is it possible to implement such a parallelism in Matlab? Which construction should I use. Till now I have been working with parfor loops, but there is no way how to put one parfor inside another.

Thanks, Lukas Halagacka

Answer 1

[Edric Ellis]

As you've discovered, you cannot nest parallelism with PARFOR. Depending on the amount of data transfer etc., you may find that "flattening" the problem is one way forward - i.e. convert the nested loops to a single flat PARFOR loop:

     % from something like this
     for r = 1:size(problems, 1)
         for c = 1:size(problems, 2)
             result(r, c) = solve(problems(r, c));
         end
     end

     % to something more like this
     parfor idx = 1:numel(problems)
         result(idx) = solve(problems(idx));
     end

[Lukas Halagacka]

Hi,

thanks for the response. I though about this approach and found it seems to be too complicated in my case. I have been trying rewrite the code to "MPI-like" instead of using parfor, but it is not easy.