Executing synchronous data flow graphs on heterogeneous execution architectures using integer linear programming

This paper presents an integer linear programming (ILP) technique to partition and schedule Synchronous Data Flow (SDF) graphs onto heterogeneous execution architectures. Our ILP formulation gives a partition and schedule for SDF graphs, which provide the optimal execution time. We incorporate new methods of exploiting stateless data-parallelism on a heterogeneous architecture. We quantitatively show that our ILP formulation performs better compared to the current state of the art heuristic techniques available, distributing SDF graphs. In fact, our ILP formulation gives an execution boost ranging from 15% to 70% for heterogeneous architectures. Finally, in this paper we also explore a new optimization technique based on actor granularity, which further improves the overall throughput of the SDF graph.