A Semantic Framework for Scheduling Parallel Programs

Declarative parallel programs oer deterministic results, al- lowing the language implementation to schedule parallel tasks in any order. However, program performance hinges crucially on the way that these tasks are scheduled. In this work, we use formal language semantics to express dierent scheduling policies. These semantics enable us to compare dierent policies and to understand their eects on the use of space. We oer several example programs to demonstrate that scheduling policy can have a dramatic, and even asymptotic, eect on space usage. To predict performance, programmers require a means to understand the eects of scheduling. We define a cost semantics that allows programmers to reason about how space is used by parallel declarative programs. At the same time, these costs provide a specification for how implementations should behave.