A dynamically scheduled parallel DSP architecture for stream flow programming

Abstract This paper presents a dynamically scheduled parallel DSP architecture for general purpose DSP computations. The architecture consists of multiple DSP processors and of one or more scheduling units. DSP applications are first captured by stream flow graphs, and then stream flow graphs are statically mapped onto a parallel architecture. The ordering and starting time of DSP tasks are determined by the scheduling unit(s) using a dynamic scheduling algorithm. The main contributions of this paper are summarized as follows: • A scalable parallel DSP architecture : The parallel DSP architecture proposed in this paper is scalable to meet signal processing requirements. For parallel DSP architectures with large configurations, the scheduling unit may become a performance bottleneck. A distributed scheduling mechanism is proposed to address this problem. • A mapping algorithm : An algorithm is proposed to systematically map a stream flow graph onto a parallel DSP architecture. • A dynamic scheduling algorithm : We propose a dynamic scheduling algorithm that will only schedule a node for execution when both input data and output storage space are available. Such scheduling algorithm will allow buffer sizes to be determined at compile time. • A simulation study : Our simulation study reveals the relationships among the grain-size, the processor utilization, and the scheduling capability. We believe these relationships have significant impact on parallel computer architecture design involving dynamic scheduling.