1

I'm using MATLAB's parallel toolbox on a shiny server (32 physical cores, 64 logical, 512GB RAM). All my parallellization needs can be classified as 'embarrassingly parallel', e.g.:

  parfor i_par = 1:n_tot
      [CalcOutputs_array(:,:,i_par), AdvancedCalcOutputs_array(i_par)] = CPU_DemandingFunction(InputsStruct_array(i_par));
  end

In order to minimize overheads, I've already split the input such that each worker receives only what it needs.

The output, however, can become quite large (1GB-10GB), and data collection from the workers is vastly over-weighting the calculation itself.

I'm positive that this is the case since by lowering the amount of data outputted from the function (I have a flag that controls it) - the runtime improvement factor approaches the # of cores i'm using.

Can that be improved? Is there any remedy for data-collection overheads?

Adriaan
  • 17,741
  • 7
  • 42
  • 75
user2617358
  • 51
  • 2
  • Relevant read: [Saving time and memory using `parfor`](http://stackoverflow.com/questions/32146555/saving-time-and-memory-using-parfor-in-matlab/32146700#32146700) – Adriaan Jun 30 '16 at 11:13
  • If you're running workers on your local PC, you can try saving results on disk, and then reading back from disk in the main client. Depending on your use-case this may actually be much slower, but can probably be faster as well. – nirvana-msu Jul 01 '16 at 18:03

0 Answers0