A software environment of architecture prototypes for evaluating parallel vision systems and algorithms

Presents a software environment consisting of a collection of parallel architecture prototypes used to evaluate the performance of parallel vision systems and algorithms. A consistent evaluation of parallel algorithms on various architectures is difficult due to the variety and complexity of parallel architectures and the fact that parallel algorithm-architecture performance is governed by both task characteristics and architecture characteristics. The authors use the prototyping approach for performance evaluation because theoretical analysis as well as simulation of massively parallel computer systems cannot accurately estimate the performance of these systems. The software environment presented offers a uniform basis for the comparison of a variety of parallel system configurations, such as meshes, rings, hypercubes, and pyramids. Performance measurements of algorithms on various configurations can then be derived and used to choose the most efficient algorithm for a particular vision task to run on a particular parallel system configuration. The authors describe the software environment and its operations. Its usefulness was demonstrated by developing a collection of vision tasks, executing them on the software architecture prototypes, and evaluating them by gathering actual performance measurements