Modelling deterministic concurrent I/O

Abstract

The problem of expressing I/O and side effects in functional languages is a well-established one. This paper addresses this problem from a general semantic viewpoint by giving a unified framework for describing shared state, I/O and deterministic concurrency. We develop a modified state transformer which lets us mathematically model the API, then investigate and machine verify some broad conditions under which confluence holds. This semantics is used as the basis for a small deterministic Haskell language extension called CURIO, which enforces determinism using runtime checks.Our confluence condition is first shown to hold for a variety of small components, such as individual shared variables, 1-to-1 communication channels, and I-structures. We then show how models of substantial APIs (like a modification of Haskell's file I/O API which permits inter-process communication) may be constructed from these smaller components using "combinators" in such a way that determinism is always preserved. We describe combinators for product, name-indexing and dynamic allocation, the last of which requires some small extensions to cater for process locality.