parfor much slower than for loop when using struct variables

I developed some finite element modeling code which is very CPU intensive with solving many equations and I am trying to improve it by using `parfor`. Here is the pseudo code of what I am doing

Two struct variables

Model = (structure variable with about 20 floating point variables and a few matrices of 10x10)
Elements{} = (cell array of structs where there are some nested structs inside, but all holding floating point data variables.  There is one reference to a function, ie: Elements.getState = @getStateN

Code

parfor elmID = 1:numElements
  [Elements{elmID} Results(elmID)] = Elements{elmID}.getState(Elements{elmID}, Model);
end

When I run this code, it is 40 times SLOWER than when I use a for loop.

I assumed here that the Elements{} cell array is being sliced into numElements so the whole Elements cell array wasn't being copied back and forth. Also, the Model structure seemed negligible because I took it out in one case to see.

Is there something here that I am doing wrong or is it just that there is large overhead with copying structs back and forth?

[Edric Ellis]

What size matlabpool are you using? I assume it's a 'local' matlabpool. Also, on the right-hand-side inside your PARFOR loop, don't you mean

[Elements{elmID} Results(elmID)] = getState(Elements{elmID}, Model);

The code as written passes the element into the "getState" method twice.

Finally, do you know roughly how big a MAT file on disk you get when you save Elements, Results, and Model together? That's the amount of data that needs to be transmitted to process the loop.