An Efficient Stream Buffer Mechanism for Dataflow Execution on Heterogeneous Platforms with GPUs

The move towards heterogeneous parallel computing is underway as witnessed by the emergence of novel computing platforms combining architecturally diverse components such as CPUs, GPUs and special function units. We approach mapping of streaming applications onto heterogeneous architectures using a Process Network (PN) model of computation. In this paper, we present an approach for exploiting coarse-grain pipeline parallelism exposed by a dataflow graph and describe its mapping onto CPU-GPU architecture. First experimental results conducted on a Tesla C2050 GPU indicate that use of a dataflow model on heterogeneous platforms not only enables exploiting different forms of parallelism (such as task, pipeline and data parallelism), but also has a potential to become an effective solution for reducing I/O overheads.