Should parallelization be switched on manually

Hi there! I would like to know whether parallelization in Matlab switches on automatically or should I do that manually. I have a code to solve an optimization problem (determining constants in kinetic equations, ODE), and I start with many initial guesses(doing Latin hypercube sampling). Last time I run my program, it took approx. 6 hours. Here is the code I am running:

load matlab S data input
dimS = size(S);
rankS = rank(S);
    td=data(:,1);
    xd=data(:,2);
    p = 1000;   % Number of points (samples)
    N = 8;   % Number of dimensions 
        lb = [0.00001 0.000002 0.00045 0.01 0.0001 0.00001 0.0002 0.000001]; % lower bounds 
        ub = [0.5 0.5 0.5 1 0.5 0.5 0.5 0.5]; % upper bounds 
        X = lhsdesign(p,N,'criterion','correlation');
        D = bsxfun(@plus,lb,bsxfun(@times,X,(ub-lb)));
    for i=1:size(D,1)        
            k0(i,:)=D(i,:);            
            options =optimset('Display','off','LargeScale','off','MaxFunEvals',2000);
            ke=lsqnonlin(@myobjective,k0(i,:),[1e-5,1e-5,1e-5,1e-5,1e-5,1e-5,1e-5,1e-5],[],options, td, xd, input, S);
            k_all(i,:)=ke(1,:);
    end

Answer 1

[Matt J]

Its usual internal multithreading should be enabled by default, but you can check with

>>maxNumCompThreads

Answer 2

[Titus Edelhofer]

Hi,

on first sight it looks as if the loop on D

for i=1:size(D,1)        
    k0(i,:)=D(i,:);            
    options =optimset('Display','off','LargeScale','off','MaxFunEvals',2000);
    ke=lsqnonlin(@myobjective,k0(i,:),[1e-5,1e-5,1e-5,1e-5,1e-5,1e-5,1e-5,1e-5],[],options, td, xd, input, S);
    k_all(i,:)=ke(1,:);
end

could be parallelized using parfor and the Parallel Computing Toolbox ...

Titus

[Sean de Wolski]

This is probably a better option than the 'UseParallel' flag, parallelize out as far as possible.

Answer 3

[Sean de Wolski]

If you have the Parallel Computing Toolbox, you can turn on Parallel Computing using the 'UseParallel' flag in optimset (or optimoptions).

The speedup will vary but if there is a high number of dimensions it can help significantly for gradient calculations.

[Matt J]

I don't see a "UseParallel" option in the lsqnonlin documentation...

[Steven Lord]

Not all of the options are available for all of the solvers, and even when an option is available for one of the solvers it may not be available for all the algorithms that solver can use. So for this particular situation Sean's answer doesn't apply, but in general it's something to consider when using Optimization Toolbox (or PARTICLESWARM in Global Optimization Toolbox, according to the OPTIMOPTIONS page.)