There are many approaches to exploiting multicores, but a particularly promising one is the "data-parallel" paradigm, because it combines massive parallelism (on both shared and distributed memory) with a simple, single-control-flow programming model.

Alas, data-parallel programming is usually restricted to "flat" data parallelism, which is good for implementers but bad for programmers. In particular, all the parallelism must appear at a single point, which excludes useful paradigms such as divide-and-conquer, and inhibits modular programming.

I'll describe the "nested" data parallel programming model, first developed in the 90's. It is great for programmers but much harder to implement; as a result, it's virtually unknown in practice. We are working on fixing this, by building a high-performance implementation in Haskell. I'll tell you how it works and give you some numbers.