Comparing the StreamIt and SC Languages for Manycore Processors

Embedded many-core systems offering thousands of cores should be available in the near future. Stream programming is a particular instance of data-flow programming where computations are expressed as the data-driven execution of repetitive "filters" on data streams. Stream programming fits these manycore systems\´ requirements in terms of parallelism, functional determinism, and local data reuse. Statically or semi-dynamically scheduled stream languages like e.g. StreamIt and ?C can generate very efficient parallel code, but have strict limitations with respect to the expression of dynamic computational tasks, context-dependent modes of operation, and dynamic memory management. This paper compares two state-of-the-art stream languages, StreamIt and ?C, with the aim of better understanding their strengths and weaknesses, and finding a way to improve them. We also propose an automatic conversion method and tool to transform between these two languages. This tool allows to port and evaluate the classical StreamIt benchmarks on Kalray\´s MPPA, a real-world many-core processor representative of tomorrow\´s embedded many-core chips. We conclude with propositions for the evolution of stream-programming models.