Dataflow-Based Parallelization of Control-Flow Algorithms

Compared to control-flow architectures, dataflow architectures usually offer better performances in high performance computing. Moreover, dataflow architectures consume less electrical power. However, only since recently, the technology enables the development of dataflow architectures that are competitive with control-flow architectures. From a programming perspective, there are a relatively small number of experienced dataflow programmers. As a result, there is a need of translating control-flow algorithms into the dataflow environment. This chapter focuses on extracting methods from various fields, which could be applied on translating control-flow algorithms to dataflow environment, comparing available programming tools for dataflow architectures with respect to the previously established methods, and finally, evaluating speedups and reductions in power consumption on a dataflow implementation of the Lattice-Boltzmann method, implemented using a specific tool and method. Results show a remarkable speedup (one to two orders of magnitude), and, at the same time, a considerable reduction in power consumption.