Distributed signal processing

This thesis studies system-level issues arising from implementing Digital Signal Processing (DSP) applications on highly heterogeneous and dynamic networks of processors. Compared with traditional DSP computation platforms such as microprocessor chips, distributed networks of processors offer many potential advantages such as resource sharing and fault tolerance. To realize these benefits, however, new issues in system and algorithm design must be resolved. The heterogeneous and time-varying composition of such environments reduces the efficacy of traditional methods for algorithm design and selection. This thesis develops a framework, called algorithm adaptation, which enables the system to make the ultimate selection of the signal processing algorithms applied to a data stream requiring processing. Our formulation of this strategy derives from a system model which accomodates a wide variety of processor types, from Application-Specific Integrated Circuits (ASICs) to general purpose microprocessors. The system is presented with the processing options using a DSP task description language which simultaneously expresses concurrencies and equivalences within DSP algorithms. A graphical visualization of the task descriptions leads to an interesting interpretation of algorithms as similar to communication networks, as well as mathematical formulations for dynamically choosing an execution path according to packet-wise or system-wide optimization criteria. To better quantify the effects of allowing execution paths to be dynamically chosen by the system, the thesis formulates a measure of the system capacity. A software simulation environment demonstrates that the algorithm adaptation framework increases system capacity and reduces processing delay in dynamic networks of heterogeneous processors. Thesis Supervisor: Alan V. Oppenheim Title: Ford Professor of Electrical Engineering