Throughput driven transformations of Synchronous Data Flows for mapping to heterogeneous MPSoCs

Abstract

Due to energy efficiency requirements of modern embedded systems, chip vendors are inclined towards multicore architectures with different types of processing engines and non-uniform interconnect fabrics. At the same time multiple applications are intended to run concurrently on the devices with such heterogeneous architectures. This rapid growth in the complexity of the hardware and its use cases imposes new challenges on the software development tools. To overcome this complexity, model of computation based approaches are becoming increasingly promising. Synchronous Data Flow (SDF) is a popular specification formalism for streaming applications with inherently concurrent nature. However, the parallelism expressed in the original representation is often not sufficient to maximally exploit the potential of multicore platforms. In this paper we present a holistic methodology for improving the throughput of streaming applications while mapping them onto heterogeneous architectures. The approach uses transformations that adapt the parallelism in SDF according to available platform resources. We use a genetic algorithm to explore SDF instances with the objective of maximizing throughput on a target platform. Our model supports architecture heterogeneity and multi-application scenarios. The experiments indicate that our approach outperforms other techniques for exploiting parallelism on a single application in most of the test cases and enables concurrent applications optimization.