Dynamic Determinism Checking for Structured Parallelism

Determinism is a powerful property, making it much easier to design, implement, and debug parallel programs when it can be guaranteed. There has been much work on guaranteeing determinism of programs, both through dynamic and static approaches. The dynamic approaches tend to have high overheads, while the static approaches can be difficult for non-experts to understand and can require high programmer effort in the form of typing annotations and code restructuring to fit the static type system. In this work, we present a compiler and runtime system that ensures determinism of application code, assuming that libraries satisfy their determinism specifications. In our system, the library writer includes determinism specifications in the form of annotations on the methods in the library. When these libraries are used by application programmers, our compiler will insert dynamic checks where necessary to ensure determinism at runtime. We demonstrate that our inserted dynamic checks lead to an average (geometric mean) slowdown of only 1.26× on 16 cores for our benchmark, thereby establishing the practicality of our approach.