Parallel computation of recursive tree structure - matlab

Question

I am trying to build a recursive tree structure in matlab. Since it is a pretty big tree I am trying to build, I want to speed up the calculation as much as possible, which is why I want to try making the calculation in parallel.
An example of what I am trying do may look like this:
function result = minimalExample(x)
n = numel(x);
if n == 0
result = 0;
return;
end
if n==1
result = x;
return
end
average = mean(x);
result = minimalExample(x(x<average))+minimalExample(x(x>average));
if any(x==average)
result = result*average;
end
end
I tried to use parfeval to do the calculation like this:
function result = minimalExampleParallel(x,pool)
n = numel(x);
if n == 0
result = 0;
return;
end
if n==1
result = x;
return
end
average = mean(x);
f(1) = parfeval(pool,#minimalExampleParallel,1,x(x<average),pool);
f(2) = parfeval(pool,#minimalExampleParallel,1,x(x>average),pool);
result = 0;
for i = 1:2
[~,value] = fetchNext(f);
result = result + value;
end
if any(x==average)
result = result*average;
end
end
But I get an Workers cannot execute parfeval or parfevalOnAll.-error.
I was hoping there would be a way of adding jobs to a global job queue that the workers may reach too, but I haven't been able to do so.
Is this in any way possible? And if so how? And if not so why?

Answer 1

I know you asked this quite some time ago, but I have an answer. I too was looking for such a solution, but could not find anything. There's nothing really posted online though so I want this to be here in case anyone else wants to know how.
You can do a sort of global job scheduler. Declare a parcluster then create separate jobs on said cluster using createCommunicatingJob. Then, and this is important, set the NumWorkersRange property of each job to something less than the total number of workers, preferably dividing them equally, though you can do it however you want. (Ex: If you have 12 workers and 4 jobs, set each worker range to [1 3]). If you do not do this, then it will queue the other tasks and only perform the first, as it will allocate all workers to the first task.
This works with functions that contain parfeval statements in them and the syntax is similar to parfeval. I verified using MATLAB's paralleldemo_blackjack_parfeval function. I hope this was what you were looking for. I know it's what I was looking for!
Link: http://www.mathworks.com/help/distcomp/createcommunicatingjob.html