DFGR an Intermediate Graph Representation for Macro-Dataflow Programs

Abstract

In this paper we propose a new intermediate graph representation for macro-dataflow programs, DFGR, which is capable of offering a high-level view of applications for easy programmability, while allowing the expression of complex applications using dataflow principles. DFGR makes it possible to write applications in a manner that is oblivious of the underlying parallel runtime, and can easily be targeted by both programming systems and domain experts. In addition, DFGR can use further optimizations in the form of graph transformations, enabling the coupling of static and dynamic scheduling and efficient task composition and assignment, for improved scalability and locality. We show preliminary performance results for an implementation of DFGR on a shared memory runtim system, offering speedups of up to 11× on 12 cores, for complex graphs.