Performance and energy optimization of heterogeneous CPU-GPU systems for embedded applications

Abstract

One of the most critical steps of embedded systems design is Hardware-Software partitioning. It is characterized by distributing the components of an application between hardware and software such that the user defined system constraints are satisfied. Heterogeneous computing platforms consisting of CPUs and GPUs have tremendous potential for enhancing the performance of embedded applications. The challenge of application partitioning for CPU-GPU mapping is much greater on such platforms due to their unique and diverse characteristics. In this thesis, an optimization algorithm is devised and presented for partitioning and mapping computational tasks on CPU-GPU platforms while keeping a check on the power consumption. Our methodology also uses parallelism in applications and their tasks by utilizing the architectural capabilities of the GPU. The optimization algorithm was tested with a MJPEG decoder, several benchmarks and synthetic graphs.