constraint based killer sudoku solver performance improvements

[Blockheads Oi Oi]

I have a working program based on [1] that sets up all different 
constraints for each row, column and box and then sets exact sum 
constraints for each cage.  It'll run in around 0.2 secs for a simple 
problem, but a tough one takes 2 hours 45 minutes.  I did some research 
into improving the performance and found [2] but can't work out how to 
implement the constraints given.  Can someone please help, assuming that 
it's even possible.

[1] http://pypi.python.org/pypi/python-constraint/1.1
[2] http://4c.ucc.ie/~hsimonis/sudoku.pdf
-- 
Cheers.

Mark Lawrence.