Ok. I think the following paper answers my question: http://community.haskell.org/~simonmar/papers/threadscope.pdf
In summary, the problem with
a `par` b `par` a+b
and
a `par` a+b
is the lack of ordering of evaluation. In both versions, the main thread gets to work on a (or sometimes b) immediately, causing the sparks to “fizzle” away immediately since there is no more need to start a thread to evaluate what the main thread has already started evaluating.
The original version
a `par` b `pseq` a+b
ensures the main thread works on b before a+b (or else would have started evaluating a instead), thus giving a chance for the spark a to materialize into a thread for parallel evaluation.