Why We Can't Program Multiprocessors the Way We're Trying to Do It Now

It is now rather easy to build the hardware of a multiprocessor computer, but still quite difficult to exploit it with truly parallel programs. I argue that an important cause of this problem is that the models of computation on which current programming languages are based are inadequate for describing parallelism. The two major language categories are imperative languages, based on machine-oriented models, and declarative languages, based on mathematical abstractions. Imperative languages rely on side-effects to advance computations, and so are inherently sequential. Declarative languages lack general ways of describing data parallelism, and so can only express a fraction of the potential parallelism in a program. The paper closes with a brief description of current work on constraint languages, which might help to reduce the problems of programming multiprocessors