A parallelization approach for resource-restricted embedded heterogeneous MPSoCs inspired by OpenMP

摘要

Future low-end embedded systems will make an increased use of heterogeneous MPSoCs. To utilize these systems efficiently, methods and tools are required that support the extraction and implementation of parallelism typically found in embedded applications. Ideally, large amounts of existing legacy code should be reused and ported to these new systems. Existing parallelization infrastructures, however, mostly support parallelization according to the requirements of HPEC systems. For resource-restricted embedded systems, different parallelization strategies are necessary to achieve additional non-functional objectives such as the reduction of energy consumption. HPC-focused parallelization also assumes processor, memory and communication structures different from low-end embedded systems and therefore wastes optimization opportunities essential for improving the performance of resource-constrained embedded systems. This paper describes a new approach and infrastructure inspired by the OpenMP API to support the extraction and implementation of pipeline parallelism, which is commonly found in complex embedded applications. In addition, advanced techniques to extract parallelism from legacy applications requiring only minimal code modifications are presented. Further, the resulting toolflow combines advanced parallelization, mapping and communication optimization tools leading to a more efficient approach to exploit parallelism for typical embedded applications on heterogeneous MPSoCs running distributed real-time operating systems.