TripletRun: A Dataflow Runtime Simulator and Its Performance Model

Dataflow, as a representative fine-grained parallel model, has a great potential in heterogeneous computing thanks to its being good at representing dependencies and eliminating global synchronization. In this paper, we present TripletRun, a dataflow runtime simulator. It abstracts hardware features and software behaviors in a very general way. Moreover, it supports some state-of-the-art task scheduling algorithms and provides metrics for building performance model conveniently. TripletRun provides a new perspective for searching better dataflow program execution models and for fine-grained parallel task scheduling research on heterogeneous systems. Moreover, TripletRun also has the potential to be applied in actual systems.