Parallelizing Independent Tasks Help

Hi,

I have a model that I'm trying to parallelize. In this section, I run the same function four times on four independent sets, each time the function outputs three tables of reaction probabilities. I struggled using parfor to parallelize because it behaves strangely with indexing order and I don't understand it. Can anyone chime in with advice about how they would parallelize this task?

[HF,HA,HS] = reactions(H,1,Hprobs);
[OF,OA,OS] = reactions(O,16,Oprobs);
[U25F,U25A,U25S] = reactions(U25,235,U25probs);
[U28F,U28A,U28S] = reactions(U28,238,U28probs);

Thank you!

[Walter Roberson]

We probably need to see the code for reactions() to get an idea of why there is a problem.

[Nimrod Sadeh]

function [F,A,S] = reactions(in,isotope,p)
% separate by activity of choice

if isempty(in); F = []; A = []; S = []; return; end
N = size(in,1);

switch isotope
    case 1 % hydrogen no fission
        absorbed = rand(N,1) <= p(in(:,1),2);
        F = [];

          try
              A = in(absorbed,:);
          catch ME
              if strcmp(ME.identifier,'MATLAB:badsubscript')
                  A = [];
              else
                  throw(ME);
              end
          end

          try
              S = in(~absorbed,:);
          catch ME
              if strcmp(ME.identifier,'MATLAB:badsubscript')
                  S = [];
              else
                  throw(ME);
              end
          end

      case 16 % oxygen no fission
          [F,A,S] = reactions(in,1,p);
      case 235 % U235 - fission likely!
          fissioned = rand(N,1) <= p(in(:,1),1);

          try
              F = in(fissioned,:);
          catch ME
              if strcmp(ME.identifier,'MATLAB:badsubscript')
                  F = [];
              else
                  throw(ME);
              end
          end

          try
              p = p(:,2:3)./sum(p(:,2:3),2);
              N = size(in(~fissioned),1);
              in = in(~fissioned,:);
              absorbed = rand(N,1) <= p(in(:,1),1);
          catch ME
              if strcmp(ME.identifier,'MATLAB:badsubscript')
                  A = []; S = []; return;
              else
                  throw(ME);
              end
          end

          try
              A = in(absorbed,:);
          catch ME
              if strcmp(ME.identifier,'MATLAB:badsubscript')
                  A = [];
              else
                  throw(ME);
              end
          end

          try
              S = in(~absorbed,:);
          catch ME
              if strcmp(ME.identifier,'MATLAB:badsubscript')
                  S = [];
              else
                  throw(ME)
              end
          end

      case 238
          [F,A,S] = reactions(in,235,p);
      otherwise
          msgID = 'reactions:WrongType';
          msg = 'enter the isotope number for one of the r. components';
          ex = MException(msgID,msg);
          throw(ex);
  end

% r = {F,A,S};
end

[Walter Roberson]

Could you describe further what you mean by "it behaves strangely with indexing order" ?

[Nimrod Sadeh]

Sure. I've tried putting the inputs and outputs in two separate cell arrays, and run essentially

parfor i=1:4
output(i,1:3) = reaction(input(i,1:4));
end

but MATLAB marks that as an error (valid indices are restricted in parfor loops).

[Walter Roberson]

parfor i=1:4
  temp = reaction(input(i,1:4));
  output(i,1:3) = temp;
end

[Nimrod Sadeh]

That gives me a

The variable output in a parfor cannot be classified

error.

[Greg]

You've both missed the 's' in the function name reactions, but I'm assuming that's a typo in-comment only.

I think you're having curly-brace vs. parenthesis problems, and 1:3 in input, not 1:4:

input = { H,  1,  Hprobs; ...
          O, 16,  Oprobs; ...
        U25,235,U25probs; ...
        U28,238,U28probs};
parfor i=1:4
  [F{i},A{i},S{i}] = reactions(input{i,1:3});
end

If this works, I'll move it to an Answer for acceptance.

[Nimrod Sadeh]

You're right on all counts, Greg. The code works, too. Thanks for your answer - please post it so I can accept. Thank you Walter, too.

I ran the speed test on the data, and it seems the old method is about 4 times faster than the parfor loop, so I'll probably stick to that. But thanks for the help.

[Greg]

If you're looking for general performance improvement (rather than specifically multi-threading), run the profiler. It will identify individual lines of code that are taking especially long to execute. You can then post a new (related) question, identifying those lines and we can try to help optimize.

Answer 1

[Greg]

Per the comments, I think you're having curly-brace vs. parenthesis problems, and 1:3 in input, not 1:4:

input = { H,  1,  Hprobs; ...
          O, 16,  Oprobs; ...
        U25,235,U25probs; ...
        U28,238,U28probs};
parfor i=1:4
  [F{i},A{i},S{i}] = reactions(input{i,1:3});
end