Parsley: A Scalable Framework for Dependence-Driven Task Scheduling in Distributed-Memory Multiprocessor Systems

This paper describes the design and implementation of a new parallel programming environment called Parsley, which provides ne-grained scheduling services based on the structure of the application programs. In Parsley, application programs are divided into subtasks which may run serially or in parallel. Parsley provides a programming interface that allows a user to de ne subtasks and to specify the precedence constraints among them. According to this speci cation, the Parsley system schedules subtasks and allocates processors. Thus, the subtasks are executed in a dependence-driven manner. We developed a parallel molecular dynamics simulation program based on the Parsley mechanism and executed it on a scalable multiprocessor system. We achieved good scalability and showed that our system is e cient for large-scale molecular dynamics simulation.