Parallel programs are complex objects and are hard to build. One way to make building one easier is to insist that the programming language be compositional, that is, the properties of the whole must be (simple) functions of the properties of the pieces. The paper is an argument for this thesis, using the languages Strand and PCN as successful examples.
In fact, only compositionality of behavior is really discussed. I would argue that compositionality of costs is at least as important, perhaps moreso, because without it, it is impossible to break designs into pieces and build them independently.
This paper would have been an interesting contribution to the discussion of parallel programming models, had it appeared in a timely way. In a fast-changing field, the delay in publication has necessarily reduced the value of its contribution.