A new method for minimizing buffer sizes for Cyclo-Static Dataflow graphs

Several optimizations must be considered for the design of streaming applications (e.g. multimedia or network packet processing). These applications can be modelled as a set of processes that communicate using buffers. For this purpose, Cyclo-Static Dataflow graphs, which are an extension of Synchronous Dataflow graphs, allow to consider a large class of industrial applications. This paper presents an original methodology to minimize the global surface of the buffers for a Cyclo-Static Dataflow graph under a given throughput constraint. It is proved that, if the processes are periodic, each buffer introduces a linear constraint described analytically. The optimization problem is then modelled by an Integer Linear Program. A polynomial algorithm based on its relaxation provides a quasi-optimal solution for real life problems. The resolution of the optimization problem for a Reed-Solomon Decoder application is then detailed.