These days, parallel computing hardware is ubiquitous. However, parallel programs are typically much more difficult to write than sequential ones, so making use of all of the parallelism available in a modern computer remains difficult. Functional programming languages offer a radical and elegant attack on this challenge, by tackling the root cause of the problem, namely, unrestricted side effects.

In this talk I will reflect on the progress and challenges we have had in developing the programming language Accelerate, an embedded language in Haskell for data-parallel array programming, which exposes only a collection-oriented interface over immutable arrays. I will discuss some of the aspects to making such a language design practicable and performant, and give some examples of how this more restrictive programming model can help guide the programmer towards an efficient parallel implementations of their algorithm.