Generally speaking, spmd is used when you need communication to occur between workers in the matlabpool. Another difference is that labindex provides the index of each worker in the pool when used within spmd. In parfor, labindex is always 1 because parfor is intended only for independent tasks.
It is hard to comment on why your specific large data loop was faster with spmd than parfor without seeing a snippet of code. If you could post details about the operations in question then maybe we can identify where the bottleneck is.
