Best approach to this problem - spmd or parfor ?

Am wondering what will be best approach for parallelizing the code for the problem am working on.

I have a time series data (a trajectory) that is ~800,000 x2 matrix and I do some operations on it that involves matrix exponentiation, diagonalization & multiplications. It is a maximum likelihood (ML) calculation & since it produces 'Inf' for chunks of more than 1000 points (1000x2 matrix), I divide the long trajectory into pieces of 1000 points and do the required calculation on them (& the final ML is total additive of that of the pieces). Besides the time series data, I also have some other parameters. So my objFunc is like:

    [MLtotal,MLsplit] = objFunc(params,trajectoryMat)

      trajChunks = reshape(trajectoryMat,...) % split into small chunks

      parfor i=1:length(trajChunks) % easily parallelizable
      MLsplit(i) = calculateML(params,trajChunks)
      end
      MLtotal = sum(MLsplit)

calculateML actually is a fast calculation taking ~0.5 seconds and the serial caculation takes ~ 8 minutes for the whole Matrix (on 8 processor machine).

When I use the parfor with a 100 workers matlabpool, it takes like 8-10 seconds, which is a good scale-up.

Now, I need to do vary the parameters (params) and do the calculation like ~3000 times. (in fact, I want to do a parameter sweep).

    paramsMat = [p1Vect; p2Vect;p3Vect ;p4Vect];
    % p1,p2,p3,p4 are parameters  & p1Vect, p2Vect etc., are 3000 point vectors.
    for ix = 1:length(paramsMat)
           currentparams = paramsMat(ix,:);
           [MLtot(ix),MLsplit(ix,:)] = objfun(currentparams,trajectoryData)
    end

This takes like 24-30,000 seconds using 100 processors (our cluster size is 144 cores). Though it is scaling linearly, is there a better approach to do it?

It is the same program operating on multiple chunks of data in a way. So I was wondering if using spmd is a better choice ? How to use it ? (I haven't used spmd before)

Do you have any other suggestions for speeding it up.

Answer 1

[Edric Ellis]

I would say that it sounds like your PARFOR solution is probably pretty close to the best you can do. It sounds like your problem is well suited to PARFOR - lots of independent chunks of work to do. While you could achieve the same sort of thing with SPMD, you wouldn't get the automatic load-balancing that PARFOR gives you, so it might end up running less efficiently.

One thing you might consider is moving the PARFOR loop to be the outer loop as that will reduce some of the overheads associated with running in parallel.