In principle, reductions can often be staged, since there's no ordering requirements. Imagine a tree of reductions. But you are correct, the reduce phase is what will limit parallelism. If you have a cheap map operation, but a really expensive reduction, you may not see much scalability. (Where "scalable" is a way of saying "performance improves as available hardware increases because more parallelism inherent in the application is exploitable.")