Performance measurements of complex signal and image processing algorithms over multiple signal processor systems

Abstract

The evolution of commercial signal processors has enabled migrating these devices toward the sensor of a detection system while eliminating specialized hardware to perform similar operations. As the technology of the signal processor has evolved, systems have been formulated with multiple instances of the signal processor to achieve a linear improvement in throughput performance. Two forces that enable this linear improvement are primarily the ability of an algorithm to be mathematically separated into the several components and secondarily the interconnection structure established between the signal processors. A classic problem achieving this linear improvement is implementing the Fast Fourier Transform (FFT) across several signal processor systems. Other signal-processing and image-processing algorithms that exhibit similar linear improvement of performance over the number of processors also exist. A critical evaluation of generic signal- and image-processing algorithms on multiple signal-processor architectures has been performed, using the singular metric that illustrates true performance—execution time. In this evaluation, several strategies were assessed to reveal strengths and weaknesses sensitive to the target architecture in maximizing performance gain. This paper describes the methodology of this study and presents the results of the types of algorithms that meet this linear improvement with respect to number of signal processors.