Parallelized MATLAB code takes significantly longer time than serial. - matlab

Question

I am wanting to parallelize some MATLAB code that does some parameter fitting using MATLAB's mle routine. My reason for parallelization is that I want to execute the mle routine for multiple different guesses on the same data set.
Initial simulations show that the parallelized code takes anywhere from 200-400 seconds longer than the serial version (~500 seconds) depending on what data set I fit (to time, I just tic and toc at the beginning and end of my code). Am I using parfor incorrectly?
tic
mitdata; % load data
data=dmso; % take the 'DMSO' column data
num=length(data);
mm=linspace(.1, 1, 2)'; % mean values
vv=linspace(.1, 2, 2)'; % variance values
N=length(mm);
n=length(vv);
% get all combinations of the initial parameter guesses I want to try
pp = cell(N, n);
for i = 1:N
for j = 1:n
pp{(i-1)*N+j} = [mm(i), vv(j)];
end
end
temp = nchoosek([1:N*n, 1:N*n], 2);
temp = sort(temp, 2);
idx = unique(temp, 'rows');
P = zeros(length(idx),4);
for ii = 1:length(idx)
P(ii,:) = [pp{idx(ii,1)},pp{idx(ii,2)}];
end
Pcell = num2cell(P); % convert P to cell so that it's easier to assign parameter values for initial guess
options = statset('MaxIter',10000, 'MaxFunEvals',10000);
pd = zeros(length(P),4); % space for best fit parameters
ld = NaN*ones(length(P),1); % space for likelihood values
parfor i=1:length(P)
[m1,v1,m2,v2] = Pcell{i,:};
x0 = [m1,v1,m2,v2]; % initial guess
[p,conf1]=mle(data,'pdf',#convolv_2invG,'start',x0, 'upperbound', [Inf Inf Inf Inf],'lowerbound',[0 0 0 0],'options',options)
pd(i,:)=p; % get best fit parameters from MLE
l=convolv_2invG(data,p(1),p(2),p(3),p(4)); % use best fit parameters to evaluate pdf
l=sum(log(l));
if l<0
ld(i)=l; % store likelihood values
end
end
toc