Kahn process networks, Haskell’s monad-par library, and Intel’s Concurrent Collections language are three diverse examples of deterministic-by-construction models of parallel computation. In a deterministic-by-construction model, all programs written using the model are guaranteed to behave deterministically, so they offer programmers the promise of freedom from subtle, hard-to-reproduce nondeterministic bugs like race conditions.
As I’ll show in my talk, the determinism of all of these models hinges on a notion of monotonic data structures—data structures in which mutators may only add information, never destroy it. In each case, a monotonicity property on some data structure captures the essence of why the model is deterministic. With monotonic data structures as a guiding principle, then, we should be able to develop a deterministic parallel programming model that is general enough to express existing models, as well as guide the design of new ones. Ryan Newton, Amr Sabry, and I are working on developing such a model, and I’ll discuss what we’ve accomplished so far and where we’re headed next. (Warning to implementors: this talk may contain theorems and inference rules! Warning to theorists: this talk may contain running code!)