Scalable parallelism using dataflow programming

gram level. Many forms of computation are well suited to dataflow description and implementation. Some common examples include complex media coding, network processing, imaging and digital signal processing, and em bedded control. Digital radio is one application domain that can be successfully modeled using data flows. Dataflow modeling facilitates component isolation and “pluggable” processing chains. Performance-enhancing functionality, such as hybrid automatic repeat request (HARQ), can be implemented as an optional plugin. Given that each new generation of mobile terminals is expected to support a larger number of radio standards, more functionality is being implemented in software or programmable hardware. Building a processing chain from pluggable modules – either standard-specific or universal with parameters – is therefore an attractive proposition. Dataflow modeling is not only a means of creating transceiver implementations. The use of dataflow modeling can also reveal bottlenecks from a lack of parallelism inherent in a radio standard. This is especially useful when developing new standards. To support a wider range of maximum communication speeds, next genera tion radio standards will be scalable in terms of capacity and processing requirements. Ideally, it should be possible to build a high-capacity transceiver by duplicating the hardware of a baseline transceiver. To do this, the potential parallelism should scale in a linear way with processing requirements. Parallel programming is not a new proposition or challenge. At the end of the 1960s, researchers were struggling to make the transition to paralDataflow programming was invented to address the issue of parallel computing. Dataflow models are well suited to describe many forms of computation, particularly in the area of digital signal processing. This article presents case studies about an MPEG-4 video decoder and a multi-standard OFDM receiver. Hardware and software have been synthesized from the dataflow models, and experimental results are presented.