Object-oriented prototypes of parallel architectures for the performance evaluation of algorithms

The author describes a methodology for constructing prototypes of parallel architectures. The prototypes model the functionality of the components of the architecture to any specified level of detail. The prototypes are programmable; thus benchmarks of algorithms can be executed. The prototypes are also flexible, allowing the fine tuning of the architecture to enhance the performance. A specific working case study is covered. A working prototype of a pyramid architecture is constructed to seven levels (546 processors and 53564 interconnections) and is used to evaluate the mapping of a wide range of computer vision algorithms (i.e. low-level image processing through high-level symbolic computations) on such an architecture. The results indicate that a pyramid architecture is not only useful for low-level image processing tasks but that segmentation and even higher-level symbolic computations can be performed efficiently. Results are presented for a pyramid implementation of a split-and-merge segmentation algorithm.<<ETX>>