http://www.mathworks.cn/cn/help/distcomp/advanced-topics.html
PARFOR loops work by dividing the iterations of the loop among many
workers. Things work best when the inputs to the loop and outputs from
the loop can also be divided in the same way. The simplest example might
be:

N = 10;
in1 = rand(1, N);
in2 = rand(1, N);
out2 = 0;
parfor ii = 1:N
  out1(ii) = in1(ii) * 2 + max(in2);
  out2 = out2 + in1(ii);
end

In this case, both 'in1' and 'out1' are SLICED because PARFOR can see
how to divide them up - not all of 'in1' or 'out1' are required to
perform an iteration of the loop. 'in2' is *not* sliced, because the
whole value is needed for each loop iteration. 'out2' is not sliced
either, but PARFOR can still compute the value because it understands
that it is a REDUCTION variable. More info:

<http://www.mathworks.com/help/toolbox/distcomp/brdqtjj-1.html>

PARFOR warns you if you have a variable inside a PARFOR loop that you're
indexing, but that isn't sliced. This warning appears because you might
have intended the variable to be sliced, but PARFOR couldn't slice it.

PARFOR can slice multi-dimensional arrays providing that the indexing
expression matches various conditions layed out in the page referred to
above. For example:

in3 = rand(N);
parfor ii=1:N
  out3(:,ii) = 1 - in3(:,ii);
end

both 'in3' and 'out3' can be sliced.

Cheers,

Edric.