A Performance Prediction for Iterative Reconstruction Techniques on Tomography

Abstract

Algebraic reconstruction techniques (ART) for image reconstruction were dismissed during the 1970s due to high-demanding computing requirements. Nowadays, in order to meet these requirements, parallelization strategies with domain decomposition have been applied. Furthermore, a performance prediction model would allow added knowledge of the parallel application and predict its behavior under different parameters or hardware platforms. This paper describes an analytical performance prediction model for a parallelization of iterative reconstruction techniques. The techniques' behavior is analyzed step by step to create an analytical formulation of the problem. BPTomo is a parallel distributed application for tomographic reconstruction that uses iterative reconstruction techniques. The model is validated by comparison of the predicted times for representative datasets with BPTomo computation times measured on a PC cluster. The model is shown to be quite accurate with a deviation between experimental and predicted times of lower than 12%.